MeasuredOne study (Aggarwal et al., KDD 2024), run on a 2023 research engine and a 200-query Perplexity sample. A 2026 critical survey finds such gains hold only once a page is already among the retrieved sources.
Cite-worthy statistics are specific, checkable numbers placed in content, such as a dated measurement with its source, that an AI answer engine can lift to support a claim. The GEO paper's "Statistics Addition" method was one of its top three tactics, but the evidence comes from one benchmark study that measured pages already selected as sources.
Cite-worthy statistics are numbers a reader or a machine can verify: a figure, its unit, the population or sample it describes, the period it covers and who produced it. "Many searches now show AI summaries" is a qualitative claim; "Pew Research Center found an AI summary on 18% of the Google searches in its March 2025 panel" is a statistic an engine can quote and attribute.
The term comes from practice built on the GEO paper (Aggarwal et al., arXiv:2311.09735, KDD 2024), whose "Statistics Addition" method rewrote a source page to include quantitative statistics instead of qualitative discussion wherever possible. The lever covers adding, sourcing and dating numbers inside the page text. Publishing data of your own is a separate lever (original research and data studies). The lever also says nothing about whether a page gets retrieved in the first place: the paper's tests started from pages already handed to the engine as sources.
How it works
In the paper's setup, a generative engine (gpt-3.5-turbo) answered each query from the top five Google results, and an LLM rewrote one source per query with each method. Visibility was scored with Position-Adjusted Word Count (how much of the answer a source accounts for, weighted toward earlier citations) and Subjective Impression (an LLM-judged score covering relevance, influence and similar qualities).
On GEO-bench, Statistics Addition scored 25.9 on word count against a 19.5 baseline. The authors report that their top three methods gained 30 to 40% on that metric and 15 to 30% on the impression metric, while a table caption gives 41% and 28% for the best methods; the two framings come from the paper itself. On Perplexity, tested on 200 queries with sources supplied as uploaded files, Statistics Addition reached the highest impression score (33.9 against 24.7) with a smaller word-count gain (26.2 against 24.1). The paper's best two-method combination was Fluency Optimization plus Statistics Addition, which beat any single method by more than 5.5% on a 200-example subset.
Why it matters
Answer engines ground claims in passages they can attribute, and a precise, sourced number gives them a self-contained fact to use. Microsoft's Bing team makes a similar point: "Precise, structured language makes it easier for AI to classify your content as relevant." What a team does: replace vague claims with figures that carry a source and a date, keep each number next to the claim it supports, and refresh figures when newer data appears. What a team should not expect: a guaranteed citation. The effect was measured on a 2023 research engine, and Martinez's 2026 survey notes such gains are "conditional on a source already being present in a fixed context".
Example
A hypothetical payroll software company's page says "most small businesses struggle with payroll errors." Rewritten, it says "In our March 2026 survey of 1,140 US firms with fewer than 50 employees, 31% reported at least one payroll correction per quarter." The second version gives an engine a number, a sample, a date and an owner it can cite.
Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024), arXiv:2311.09735v3 full text, published 2024-06-28, read 2026-11-07: https://arxiv.org/html/2311.09735v3
Aggarwal et al., GEO: Generative Engine Optimization, arXiv abstract and submission history (v1 2023-11-16), published 2023-11-16, read 2026-11-07: https://arxiv.org/abs/2311.09735
Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), arXiv:2607.14035, published 2026-07-15, read 2026-11-07: https://arxiv.org/abs/2607.14035
MeasuredOne study (Aggarwal et al., KDD 2024). Results differ by metric and engine: on Perplexity the method raised the word-count score but lowered the impression score.
Citing sources inside your content means naming, and usually linking, the reliable sources behind your claims within the page itself. The GEO paper's "Cite Sources" method was one of its three strongest tactics on its benchmark and helped lower-ranked pages most, but the evidence is one study whose results varied by metric and engine.
Citing sources inside your content is the practice of attributing a page's claims to identifiable sources: a named study, an official document, a dataset or a standard, usually with a link and a date. It differs from cite-worthy statistics (the numbers themselves) and from quotations (a person's own words). In the GEO paper (Aggarwal et al., KDD 2024), the "Cite Sources" method had an LLM add relevant citations from credible sources to one page per query.
The lever concerns outbound attribution in the visible text. It is distinct from backlinks pointing at your page, from citation markup, and from the citations an AI answer shows its users. A page can cite well and still never be retrieved; the method acts on how a source fares once it is among the documents an engine reads.
How it works
On GEO-bench the authors list Cite Sources among their top three methods, with the 30 to 40% relative gain on Position-Adjusted Word Count they report for that group; its table value is 24.9 against a 19.5 baseline. Three findings qualify that.
First, the effect depended on where a page started: pages ranked fifth in the underlying Google results gained 115.1% in visibility with Cite Sources, while pages ranked first lost 30.3%. Second, on Perplexity (200 queries, sources supplied as uploaded files), the appendix table shows the word-count score rising to 26.6 from 24.0 while Subjective Impression fell to 19.0 from 24.7, a drop the text does not discuss. Third, the authors report that Cite Sources added most when combined with other methods (an average of 31.4%) despite being weaker alone.
Why it matters
Engines that ground answers favor claims they can trace, and a page that shows its sources makes its own claims easier to check and reuse. The rank-five result explains the attention this lever gets: in that experiment, attribution helped pages a classic search engine placed lower. A team can act on it cheaply: attribute every non-obvious claim to a named, dated source, link to the primary document instead of a summary of it, and remove citations that no longer support the claim. Treat any AI visibility gain as unproven for your engine and topic until you measure it, because the study's own Perplexity numbers point in two directions.
Example
A hypothetical guide to US small-business taxes writes "Estimated tax payments are generally due four times a year (per the IRS's estimated-tax guidance)" and links the IRS page, in place of "experts recommend paying quarterly." An engine reading the first version finds a claim and its origin in one sentence.
Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024), arXiv:2311.09735v3 full text, published 2024-06-28, read 2026-11-07: https://arxiv.org/html/2311.09735v3
Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), arXiv:2607.14035, published 2026-07-15, read 2026-11-07: https://arxiv.org/abs/2607.14035
MeasuredOne study (Aggarwal et al., KDD 2024), on a 2023 research engine and a 200-query Perplexity sample; the quotations were inserted by the LLM that rewrote each page.
Quotations and expert voices are attributed statements from named, credible people or organizations included in a page. In the GEO paper, "Quotation Addition" was the best single method on its benchmark and on Perplexity's word-count metric, one study suggesting that attributed, distinctive statements are easy for answer engines to reuse.
Quotations and expert voices are verbatim statements, attributed to a named person or organization with relevant standing, embedded in a page's text: a researcher explaining a result, a regulator's published wording, a practitioner describing what they saw. The GEO paper's "Quotation Addition" method had an LLM add relevant quotations from credible sources to one page per query.
The lever covers real, attributable quotations. Invented quotes, anonymous testimonials and paraphrases dressed up as quotes fall outside it, and they create accuracy problems when an engine repeats them. In the experiment the quotations were produced by the rewriting model, so the benchmark tested how an engine responds to the presence of quotes; it did not test whether they were genuine.
How it works
On GEO-bench, Quotation Addition had the highest scores of any single method on both metrics in Table 1: 27.8 Position-Adjusted Word Count against a 19.5 baseline, and 24.7 Subjective Impression against 19.3. On Perplexity, the authors report it "performs best in Position-Adjusted Word Count with a 22% improvement over the baseline." It also lifted lower-ranked pages: sources in fifth position gained 99.7%, while first-position sources lost 22.9%. By domain, it performed best on the paper's People & Society, Explanation and History query categories.
Why it matters
A named quotation is a compact, attributable unit: an engine can reproduce it with its speaker and keep the answer traceable. For topics that lean on judgment (health guidance, policy, how-to advice), expert statements also show that a page goes beyond common knowledge, which Google's 2026 generative AI guide values as "unique expert or experienced takes". What a team does: interview real experts, publish their statements with full names and roles, quote primary documents exactly, and date each quote. The evidence remains one benchmark; no published test shows quotes raising citation rates in current production engines.
Example
A hypothetical page on heat pumps in cold climates adds a quote from a named, licensed HVAC contractor in Minnesota describing how her customers' systems performed at minus 15 degrees Fahrenheit, with her license type and the interview date. A generic line such as "heat pumps work well in winter" gives an engine nothing to attribute.
MeasuredOne study (Aggarwal et al., KDD 2024), where stylistic methods gained 15 to 30%. Google says no special writing style is needed for its AI features.
Fluency and readability optimization is rewriting a page so its prose is smooth, well organized and simple to follow, without changing the facts. In the GEO paper, "Fluency Optimization" and "Easy-to-Understand" raised visibility by 15 to 30% on its benchmark, one study suggesting presentation affects how much of a source an engine reuses.
Fluency and readability optimization covers two of the GEO paper's nine methods: "Fluency Optimization", which improves the fluency of a page's text, and "Easy-to-Understand", which simplifies its language. Both change presentation only; unlike statistics, citations or quotations, they add no new information.
In practice the lever means clear sentences, a logical order, terms defined on first use, short paragraphs under informative headings, and no filler. It overlaps with plain-language editing and accessible writing. It is distinct from tone changes meant to sound authoritative, a separate method with weaker results in the paper, and from rewriting text to match particular query phrasings, which Google's 2026 guidance calls unnecessary because its systems understand synonyms and general meaning.
How it works
The paper reports that these stylistic changes "resulted in a significant visibility boost of 15-30%", which the authors read as engines valuing information presentation as well as content. In Table 1, Fluency Optimization scored 25.1 on Position-Adjusted Word Count against a 19.5 baseline, close to Cite Sources (24.9), while Easy-to-Understand scored 22.2. The best two-method combination in the paper was Fluency Optimization plus Statistics Addition, which beat any single method by more than 5.5% on a 200-example subset. Gains were uneven: in the rank analysis, Fluency Optimization moved visibility between a 4.4% loss and a 5.2% gain across the five rank positions, and its strongest query categories were Business, Science and Health.
Why it matters
Clear prose is the cheapest lever in this section and the one least likely to backfire: it helps human readers, screen readers and machine extraction at once. Microsoft's Bing guidance asks for sentences "that make sense even when pulled out of context", and Google asks that content be "well written and easy to follow". What a team does: edit for one idea per paragraph, define terms on first use, cut filler, and read each section as if it were the only part an engine will see. Do not expect a measurable citation gain from style alone; the 15 to 30% figure comes from one 2023 benchmark.
Example
A hypothetical insurance explainer replaces "Coverage parameters are subject to variability contingent upon policy configuration" with "What your policy covers depends on the options you chose when you bought it." The facts are unchanged, and the second sentence can stand alone in an answer.
Also called: keyword stuffing (negative control), keyword spam, keyword repetition, term stuffing
DocumentedMeasuredDocumented as spam in Google's policies, which since May 2026 state they also cover generative AI responses; measured as a negative result in one study (Aggarwal et al., KDD 2024).
Keyword stuffing is repeating query terms, or lists of keywords, in a page to force relevance, a classic SEO manipulation. In the GEO paper it gave little to no gain and performed about 10% worse than baseline on Perplexity's word-count metric, and Google's spam policies, which apply to its generative AI responses, prohibit it.
Keyword stuffing means filling a page with keywords or numbers to manipulate rankings, often in unnatural lists, repeated phrases or out-of-context blocks. Google's spam policies define it in those terms and give examples such as blocks of city and region names a page is trying to rank for. In the GEO paper, the "Keyword Stuffing" method added more keywords from the query, "as expected in classical SEO optimization", and served as the study's classic-SEO comparison.
The entry covers deliberate repetition. Using the words your audience uses, naturally and in context, is ordinary writing. Google's 2026 generative AI guide adds that its systems understand synonyms and general meaning, so a page does not need to capture every phrasing of a query.
How it works
In Table 1 on GEO-bench, Keyword Stuffing scored 17.8 on Position-Adjusted Word Count, below the 19.5 baseline, and the authors conclude such methods "offer little to no improvement". On Perplexity (200 queries supplied as uploaded files), they report it "performs 10% worse than the baseline" on word count (21.9 against 24.1). The same Perplexity table shows a higher Subjective Impression score than baseline (28.1 against 24.7), which the text does not discuss, so the negative result rests on the word-count metric. The paper's explanation is that the model in a generative engine "is not limited to keyword matching", so repetition adds nothing for it to use.
Why it matters
This is the one GEO lever where the research and the official policy agree: stuffing did not help in the benchmark, and in Google it is spam. Google's policies cover attempts "to manipulate generative AI responses in Google Search", and sites that violate them may rank lower or drop out of results. What a team does: audit pages for repeated exact-match phrases, city lists and keyword-laden alt text, rewrite them as natural prose, and cover a topic's real questions instead of its phrasing variants. Google's 2026 guide also says that producing separate pages for every query variation, fan-out queries included, primarily to manipulate AI responses violates its scaled content abuse policy.
Example
A hypothetical plumbing page repeats "emergency plumber Denver" eleven times and ends with a list of 40 suburb names. A rewrite that explains response times, pricing and service areas in plain sentences gives an engine facts to cite; the stuffed version gives it noise and risks a spam action in Google.
MeasuredOne study (Aggarwal et al., KDD 2024) with mixed results: the authors found no significant gain from a persuasive tone on their benchmark, while the same method scored higher on Perplexity's impression metric.
Authoritative tone and technical language are style changes that make a page sound more persuasive and expert, or add domain terms. In the GEO paper, the "Authoritative" and "Technical Terms" methods produced small, uneven effects: the authors found no significant gain from persuasive tone on their benchmark, and results varied by domain and engine.
This entry groups two of the GEO paper's nine methods. "Authoritative" modified a page's style to be "more persuasive and authoritative"; "Technical Terms" added technical terms wherever possible. A related method, "Unique Words", added distinctive vocabulary and scored 20.7 against a 19.5 baseline on the word-count metric. All three change wording without adding facts, sources or quotations.
The lever is about voice and vocabulary. Demonstrated authority (named authors, credentials, first-hand evidence) is a different matter, covered by author authority, E-E-A-T and expert quotations. Sounding confident and being a credible source are separate properties, and the study tested only the first.
How it works
On GEO-bench, Authoritative scored 21.8 on Position-Adjusted Word Count and Technical Terms 23.1, against a 19.5 baseline, gains well below the top methods. The authors note that instruction-following models might be expected to favor persuasive text, yet "we find no significant improvement", and they conclude that engines already resist such changes to a degree. The appendix's Perplexity table tells a different story for tone: Authoritative reached 30.6 on Subjective Impression against 24.7. By domain, Authoritative performed best on the Debate, History and Science categories. In the rank analysis, its effect ranged from a 6.0% loss for first-ranked pages to a 12.6% gain for fourth-ranked ones.
Why it matters
Tone is where much GEO advice overpromises. The one controlled study found persuasive style weaker and less consistent than adding facts, sources and quotes, and Google's 2026 guidance says you do not need to write in a specific way for its AI features. What a team does: use the precise technical terms its audience uses, define them, and let authority come from evidence: named experts, data, sources and first-hand detail. Treat tone rewrites as an editorial choice made for readers, and measure before claiming any AI visibility effect.
Example
A hypothetical cybersecurity vendor rewrites its ransomware explainer from neutral prose into confident claims ("the definitive solution trusted by industry leaders"). In the GEO paper's terms that is the Authoritative method, which its authors found gave no significant improvement on their benchmark; adding a dated incident statistic and a named analyst's quote matches the methods that scored higher.
MeasuredOne study (Aggarwal et al., KDD 2024); the categories come from tags on its own benchmark queries. A 2026 critical survey of the field also finds that generic heuristics transfer poorly.
Domain-specific optimization is the principle that the content tactics that raise visibility in AI answers differ by topic and query type, so no single recipe works everywhere. The GEO paper found each method had different best-performing categories, such as statistics for law and government queries and quotations for questions about people and society.
Domain-specific optimization comes from a central finding of the GEO paper (Aggarwal et al., KDD 2024): "the efficacy of these strategies varies across domains", so website owners should make "domain-specific targeted adjustments". The authors tagged GEO-bench queries by category and reported which categories each method served best.
The term covers choosing tactics by subject area and query intent: factual lookups, opinion and debate questions, explanations, legal and government topics, health, business. It maps only loosely onto industry verticals, since one industry's pages answer many query types. The category results also have limits: they show where a method did relatively best on one benchmark, which falls short of proof that the method works for that category on today's engines.
How it works
Table 3 of the paper lists the top three categories per method. Authoritative: Debate, History, Science. Fluency Optimization: Business, Science, Health. Cite Sources: Statement, Facts, Law & Government. Quotation Addition: People & Society, Explanation, History. Statistics Addition: Law & Government, Debate, Opinion. The categories come from tags on GEO-bench's 10,000 queries, drawn from nine sources (including MS MARCO and Natural Questions) plus synthetic queries. Martinez's 2026 critical survey reaches a compatible conclusion from a wider body of work: topical relevance is among the most reproducible levers, and "generic heuristics transfer poorly".
Why it matters
Teams often copy one checklist across every page. The study's category results argue for matching the tactic to the question: data for policy and legal topics, attributed sources for factual statements, expert quotes for questions about people and history, clear explanation for business and health. What a team does: group its target questions by type, test one tactic per group against a control set of pages, and keep only what moves its own citation numbers. Because the categories come from one benchmark built in 2023, treat them as hypotheses to test before rolling anything out.
Example
A hypothetical legal-information site testing GEO tactics starts with dated statistics on its law and government pages, where the paper found Statistics Addition strongest, and with cited primary sources on pages answering factual questions, where Cite Sources did best. It measures citation share for each group against untouched pages before extending either tactic.
Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024), arXiv:2311.09735v3 full text, published 2024-06-28, read 2026-11-07: https://arxiv.org/html/2311.09735v3
Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), arXiv:2607.14035, published 2026-07-15, read 2026-11-07: https://arxiv.org/abs/2607.14035
Also called: answer capsule, answer block, answer-first paragraph, snippable answer, direct answer paragraph, BLUF (bottom line up front)
DocumentedClaimedRecommended in Microsoft's Bing guidance (Bing scope); Google says writing or chunking content specially for AI is unnecessary. No published controlled test measures the effect on AI citations.
Answer capsules, also called answer blocks, are short, self-contained passages, often one to three sentences, placed right under a question or heading and written so an AI engine can lift them whole as an answer. Microsoft's Bing team recommends such snippable answers; Google says no special formatting is required, and no study has measured the effect.
An answer capsule (or answer block) answers the question its section raises in the first sentences, completely enough to make sense with nothing around it: it names the subject, gives the answer and includes the key qualifier. The rest of the section then adds detail, evidence and exceptions. Length conventions circulate among practitioners, but no engine publishes a target length; Microsoft's guidance speaks of one- to two-sentence responses.
The pattern descends from featured-snippet practice and resembles the journalistic inverted pyramid and the military "bottom line up front". It is a writing pattern, so it needs no markup and has nothing to do with FAQ schema. It also differs from splitting content into many tiny fragments, which Google's 2026 guide lists among practices site owners can ignore.
How it works
Answer engines retrieve and quote passages, so a passage that carries its own subject and answer survives extraction intact. Microsoft's October 2025 guidance says assistants "break content into smaller, usable pieces", recommends "one- to two-sentence responses that directly address a question" and phrasing that makes sense out of context, and notes that assistants "can often lift these pairs word for word" from question-and-answer formats. Google's featured snippets, which also appear inside People Also Ask, apply a similar logic in classic results. Google's 2026 generative AI guide sets the limit: its systems understand multiple topics on one page and show the relevant piece, so breaking content into tiny pieces is not required.
Why it matters
A capsule costs little to write and serves three readers at once: the skimming human, the classic snippet and an AI engine pulling a passage. Its value is clarity; the risk is turning every paragraph into a disconnected fragment, which Google calls unnecessary. What a team does: open each key section with a direct, complete answer that names its subject, place the evidence and caveats right after it, and check that the capsule stays accurate when quoted alone. Do not promise citations from it: Microsoft says nothing guarantees selection, Google says indexing and serving are never guaranteed, and no public test measures the effect.
Example
Under the heading "How long does a US passport renewal take?", a hypothetical travel site opens with: "Routine renewals by mail usually take several weeks; expedited service costs extra. Check the State Department's current processing times before applying." The capsule answers, qualifies and points to the primary source in two sentences.
Also called: extractability, machine-readable content, liftable content, parseability, textual availability
DocumentedDocumented for Google (important content in textual form) and for Bing (content hidden in tabs, images or PDFs may be skipped); each statement applies to its own engine.
Content extractability is how easily a machine can pull complete, accurate statements out of a page. It depends on key information being present as text in the HTML, instead of locked inside images, PDFs, hidden tabs or script-only rendering. Google and Microsoft both document the underlying practices for their AI features.
Content extractability describes whether a page's meaning survives the trip from the page into a machine's text representation. Extractable content is present in the served HTML as text, sits in coherent paragraphs, lists and tables, and keeps each claim next to its subject and qualifiers. Poor extractability looks like a pricing comparison published only as a screenshot, a specification sheet available only as a scanned PDF, answers hidden behind tabs that load on click, or text injected by JavaScript that some crawlers never run.
The term is a practitioner umbrella; engines document the practices under other names. It overlaps with semantic HTML and JavaScript rendering, the usual technical causes, and with answer capsules, which address extractability at the level of writing.
How it works
Google's AI features page lists "making sure that important content is available in textual form" among its best practices for AI Overviews and AI Mode. Microsoft's October 2025 guidance for Bing's AI answers warns that "AI systems may not render hidden content", says PDFs often lack the structural signals HTML provides, and advises alt text or HTML for critical details in images. Crawler behavior adds a third constraint: Vercel and MERJ's December 2024 analysis found that the major AI crawlers they observed did not execute JavaScript, so text that exists only after client-side rendering may be missing for them.
Why it matters
An engine can only cite what it can read, and extractability failures are silent: the page looks complete to a person while a crawler sees an empty container, an image, or a fragment cut off from its subject. What a team does: view the raw HTML, or fetch the page with JavaScript disabled, and confirm the key facts appear as text; move critical tables out of images; give important PDFs an HTML equivalent; keep essential answers visible without a click. These fixes help classic search, accessibility and AI answers at once, which makes them low-risk even where the AI-specific payoff is unmeasured.
Example
A hypothetical SaaS company publishes its plan comparison as a PNG graphic. An assistant asked "Does this tool include single sign-on on the basic plan?" finds no text to quote. Rebuilt as an HTML table with a one-line summary above it, the same comparison becomes quotable.
ObservedMicrosoft's Bing guidance recommends descriptive over vague headings. No engine or study has tested headings that state a claim, so the AI-specific benefit is an inference from how passages are retrieved.
Declarative headings are headings that state a section's point, such as "Most AI crawlers skip JavaScript", instead of teasing it. They give readers and machines an accurate label for each passage. Microsoft recommends descriptive headings for AI answers; the extra benefit of headings that state a claim is an untested inference.
A declarative heading summarizes what its section says. "Most AI crawlers skip JavaScript" is declarative; "The hidden problem" and "Learn more" are vague. Descriptive headings ("JavaScript rendering and AI crawlers") sit in between: they name the topic without stating the conclusion.
The lever concerns heading text at every level, from the page title to H3. It complements question-based headings, which mirror the user's question, while a declarative heading carries the answer; a page can mix both, with a question as the H2 and a declarative opening sentence below it. Accuracy is part of the definition: a heading that overstates its section misleads people and any system that relies on it to label the passage.
How it works
Retrieval systems split pages into passages and judge each one's relevance, and headings are among the clearest signals of what a passage covers. Microsoft's October 2025 guidance says headings "act like chapter titles that define clear content slices", suggests replacing a vague label such as "Learn More" with a descriptive one, and describes the title, meta description and H1 as signals AI systems use to interpret a page's purpose and scope. Google's 2026 guide asks for "headings that provide a clear structure to navigate content", for readers first. The step from descriptive to declarative headings has not been tested by any engine or study.
Why it matters
When an engine pulls one passage, the heading is often the only context that travels with it. A heading that states the point makes the passage's claim obvious to retrieval and to the reader who skims an AI answer's sources. What a team does: rewrite vague or clever headings so each names its topic, state the conclusion where the section has a firm one, keep heading levels logically nested, and avoid generic labels like "Overview" repeated across pages. This is low-cost editorial work with clear benefits for readers; its AI-specific effect remains unmeasured.
Example
A hypothetical customer-success platform changes the H2 "The surprising truth about onboarding" to "Onboarding calls cut first-month churn in our customer data". A retrieval system and a skimming reader both know what the section claims before reading the paragraph.
DocumentedMeasuredRecommended in Microsoft's Bing guidance (Bing scope). The measured part is adjacent: one Pew study found question-word searches triggered AI summaries most often; the effect of the headings themselves is untested.
Question-based headings are H2 or H3 headings phrased as the questions people actually ask, such as "How long does a robots.txt change take to apply?", each followed by a direct answer. Microsoft recommends question-and-answer formats for AI answers, and Pew found Google showed AI summaries on 60% of searches starting with a question word.
A question-based heading turns a section's topic into the question a user would type or say, and the text below answers that question first. The format descends from FAQ pages and People Also Ask optimization and spread in AI search work because assistant prompts are often full questions.
The lever is a heading pattern and needs no FAQ schema; Google stopped showing FAQ rich results in May 2026. It also does not call for one page per question: Google's 2026 guidance warns that creating content for every variation of how people search, fan-out queries included, primarily to manipulate rankings or AI responses violates its scaled content abuse policy. One strong page can answer a cluster of related questions under their own headings.
How it works
Two kinds of evidence support the pattern. Microsoft's October 2025 guidance for Bing says "direct questions with clear answers mirror the way people search" and that assistants "can often lift these pairs word for word". Pew Research Center's analysis of 900 US adults' Google searches in March 2025 found an AI summary on 60% of queries beginning with a question word such as who, what, when or why, against 18% of all searches; 53% of searches of 10 words or more produced one. That shows where AI answers appear, and it says nothing about whether question headings win citations. Query fan-out, which Google describes as concurrent related queries generated by the model, adds a rationale: each sub-question is another chance for a matching section.
Why it matters
Question space is where AI answers appear most, so headings phrased the way people ask make it easier for an engine to match a passage to a prompt or one of its sub-queries. What a team does: collect real questions from search data, support tickets and sales calls; use them verbatim as headings where they read naturally; answer each in the first sentences below; and group related questions on one strong page instead of spinning out thin pages. Google's guidance cuts both ways here: it values clear structure and penalizes mass-produced variants.
Example
A hypothetical tax-software help center restructures a long article into sections headed "Can I deduct a home office if I rent?", "What records do I need to keep?" and "What if I work from home only part of the week?", each opening with a two-sentence answer. The same questions appear in its support tickets and in assistant prompts.
Also called: semantic markup, HTML semantics, structural HTML, accessible markup
DocumentedObservedDocumented as web practice and in Google's, web.dev's and Microsoft's guidance; Google says perfect semantics are not required. Any effect on AI citations is unmeasured.
Semantic HTML is markup that uses elements for their meaning, such as headings, lists, tables, buttons and captions, instead of generic styled containers. It lets parsers, screen readers and AI agents see a page's structure without rendering it visually. Google calls it a good idea but not a requirement; its effect on AI citations is unmeasured.
Semantic HTML means choosing the element that matches the content's role: a real heading hierarchy (h1 to h3), ul and ol for lists, table with header cells for tabular data, figure and figcaption for captioned images, button and a for controls and links, label tied to form inputs, and landmarks such as main, nav and article. The alternative, a page built from div and span elements styled to look structured, conveys its structure only visually.
The scope is document structure. Structured data (JSON-LD) is a separate layer that describes entities, while semantic HTML describes how the visible content is organized. The two work together, and neither replaces readable text.
How it works
Machines read a page's structure in three main ways: the HTML DOM, the browser's accessibility tree, and screenshots. Google's web.dev guide to agent-friendly websites (last updated 2026-04-01) calls the accessibility tree "the page's semantic summary", which browser agents use to understand roles and controls, and advises real button and a elements over modified div and span elements. Microsoft's Bing guidance says lists, numbered steps and comparison tables "break complex details into clean, reusable segments". Google's 2026 generative AI guide keeps expectations modest: semantic HTML is "generally a good idea" and helps screen readers, while perfect markup is not required because the web in general is not valid HTML and Google can still understand it.
Why it matters
Semantic markup is cheap insurance for every machine that reads a page. A table marked up as a table keeps its rows and columns when extracted, a list stays a list, and an FAQ with real headings keeps each question attached to its answer. Browser agents that act on pages (booking, comparing, filling forms) rely on the accessibility tree, so semantic controls make their tasks possible. What a team does: audit templates for heading order and real list and table elements, label form fields, and inspect the accessibility tree in browser developer tools. Expect accessibility and agent usability gains; a citation lift is unproven.
Example
A hypothetical clothing retailer's size chart is built from styled div elements; extracted as text, it becomes a run of numbers with no column labels. Rebuilt as a table with header cells for chest, waist and length, it keeps its meaning for screen readers, crawlers and AI agents alike.
Also called: server-side rendering, SSR, client-side rendering, CSR, static site generation, SSG, pre-rendering, JS rendering
DocumentedMeasuredGoogle documents that Googlebot renders JavaScript and that not all bots can. The finding that major AI crawlers do not render rests on one vendor network analysis (Vercel and MERJ, December 2024); the Perplexity and Cloudflare crawler references checked for this entry say nothing about rendering.
JavaScript rendering is the question of whether a page's content exists in the HTML the server sends or appears only after JavaScript runs in a browser. Googlebot renders JavaScript, but a December 2024 analysis of Vercel's network found that the major AI crawlers it observed did not, so client-only content may be invisible to them.
Server-side rendering (SSR) and static site generation produce the full content in the HTML response. Client-side rendering (CSR) sends a near-empty shell and builds the content in the browser with JavaScript. Pre-rendering generates HTML snapshots ahead of time, and hydration attaches interactivity to server-rendered HTML.
For AI search the distinction matters because crawlers differ. Google documents that Googlebot queues pages for rendering and runs JavaScript with an evergreen version of Chromium; other crawlers and on-demand fetchers may read only the raw HTML. The entry concerns the visibility of meaningful content; JavaScript used for non-essential features such as chat widgets or view counters can stay client-side. Dynamic rendering, serving bots a separate pre-rendered version, is described by Google as a workaround and "not a recommended solution".
How it works
Google's JavaScript SEO documentation describes three phases (crawling, rendering, indexing) and notes that "server-side or pre-rendering is still a great idea" because it is faster for users and crawlers and "not all bots can run JavaScript". Vercel and MERJ's analysis, published 2024-12-17, reported that "none of the major AI crawlers currently render JavaScript", naming OpenAI's OAI-SearchBot, ChatGPT-User and GPTBot, Anthropic's ClaudeBot, Meta-ExternalAgent, ByteDance's Bytespider and PerplexityBot. ChatGPT's and Claude's crawlers fetched JavaScript files without executing them, while Gemini (through Googlebot's infrastructure) and AppleBot did render. Crawler capabilities change and that analysis is almost two years old, so confirm current behavior in your own server logs.
Why it matters
If the facts an engine should cite (prices, specifications, answers) appear only after client-side JavaScript runs, a non-rendering crawler or fetcher sees an empty page. The failure is easy to miss because Google, which renders, may index the same page normally. What a team does: server-render or statically generate primary content, metadata and navigation; check key pages with JavaScript disabled or by viewing the raw HTML; keep client-side rendering for features that carry no meaning; and review AI crawler requests in server logs. Vercel's own advice is to server-render critical content because ChatGPT's and Claude's crawlers do not execute JavaScript.
Example
A hypothetical SaaS pricing page loads its plan table through a client-side component after page load, and the raw HTML contains only an empty div. Googlebot can render and index the table, but a crawler that reads only the HTML response, as the December 2024 analysis found for GPTBot and ClaudeBot, sees no prices to cite.
Also called: structured data, schema markup, schema.org markup, JSON-LD, rich results markup, schema for AI
DocumentedClaimedDocumented positions from Google (no special schema needed; markup must match visible content) and Microsoft (schema helps AI systems understand pages). Claims that specific schema types raise AI citations are vendor claims unless tied to a named study.
Structured data is machine-readable markup, usually JSON-LD using the schema.org vocabulary, that labels what a page and its entities are. Google states AI Overviews and AI Mode need no special schema, while existing markup must match the visible text. Microsoft says schema helps AI systems interpret pages. Neither presents it as a citation shortcut.
Structured data is a standardized format for describing a page and classifying its content: an article with its author and dates, a product with price and availability, an organization with its address and profiles. Google recommends JSON-LD and treats its own documentation, over schema.org's, as definitive for how Google Search uses the vocabulary.
For AI search, the question is what markup does and does not do. Google's AI features page says there is "no special schema.org structured data that you need to add", and its 2026 generative AI guide lists "overfocusing on structured data" among myths to ignore, while recommending markup for rich result eligibility. Microsoft's Bing guidance says schema "helps search engines and AI systems understand your content". Entity markup such as Organization and sameAs is covered under entity SEO.
How it works
Google uses structured data to understand page content and to power rich results, and its guidelines require markup to reflect the visible page: "Don't mark up content that is not visible to readers of the page." Its AI features page adds "making sure your structured data matches the visible text on the page" as a best practice. Rich result types come and go: Google stopped showing FAQ rich results from May 7, 2026, and removed the FAQ documentation in June 2026, after dropping How-to rich results in September 2023. Microsoft recommends accurate JSON-LD that matches the content type. Other answer engines do not publish whether or how they read schema at answer time.
Why it matters
Structured data is worth maintaining for what it verifiably does: rich result eligibility in Google, clearer entity information, and accurate machine-readable facts such as prices, dates and authors. As an AI citation ticket it is a poor bet: Google says none is required, and markup that contradicts the visible page breaks its guidelines. What a team does: keep Organization, Article, Product or LocalBusiness markup accurate and consistent with the page, validate it with Google's Rich Results Test, stop expecting FAQ markup to produce a visible result in Google, and treat vendor claims that a schema type "boosts AI citations" as untested unless a named study backs them.
Example
A hypothetical software vendor adds FAQPage markup to every product page hoping to appear in AI Overviews. Google says no special schema is needed for AI features and no longer shows FAQ rich results at all, while an outdated price in the vendor's Product markup, contradicting the visible price, breaks Google's rule that markup must match the page.
Also called: llms.txt file, /llms.txt, LLMs.txt, LLMS.txt, llms-txt
DocumentedClaimedDocumented as a proposal with self-reported adoption. Google states Google Search ignores the file; the other engine documentation reviewed here does not mention it, so treating it as a citation lever is a claim.
llms.txt is a proposed Markdown file at a site's root that gives AI agents a short description of the site and links to clean Markdown versions of key pages. Proposed by Jeremy Howard in September 2024, it is read by some agents and tools on demand. Google states Google Search ignores it, and no engine documents using it for citations.
The /llms.txt file, proposed by Jeremy Howard on 2024-09-03 and now presented as v2 of the proposal (page last modified 2026-08-10), is a Markdown document whose only required element is an H1 naming the site or project. Optional parts follow: a blockquote summary, free text, and H2 sections listing links with short notes, including an "Optional" section an agent can skip when it needs less context. The proposal also asks sites to serve Markdown versions of useful pages (the page URL with .md appended) and describes discovery through rel="alternate" type="text/markdown" and rel="describedby" links or HTTP headers.
The proposal targets inference: an agent fetching guidance while helping a user. It contrasts the file with robots.txt, which governs bot access, and sitemap.xml, which lists indexable pages. This entry covers its use as a visibility lever; the file format has its own entry.
How it works
An agent or tool that knows the convention requests /llms.txt, reads the summary, and follows links to Markdown pages instead of parsing full HTML, which saves context window space. The proposal page reports that "thousands of sites" publish one, that documentation platforms such as Mintlify generate it automatically, that OpenAI, Anthropic and Google's Gemini publish files for their developer docs, and that Chrome's Lighthouse checks for one in its agentic browsing audits; those adoption figures are the proponent's own. On the consumption side, Google added a note to its generative AI guide on 2026-06-15: Google Search does not use llms.txt, and maintaining one "will neither harm nor help" a site's visibility in Search.
Why it matters
llms.txt is cheap to publish and useful where agents actually read it: developer documentation, API references, and sites that coding assistants or research agents consult. Google has ruled it out as a route into its AI features, and none of the other engine documentation reviewed for this entry mentions it. What a team does: publish one if agents are a real audience (developer tools, technical products), keep it short and current, provide the .md page versions it points to, and check server logs to see which agents request it. Measure that traffic before budgeting more work on it.
Example
A hypothetical payments API publishes /llms.txt with an H1, a two-sentence summary and sections linking to Markdown versions of its quickstart, authentication guide and error-code reference. A coding assistant asked to integrate the API can fetch those clean pages directly; Google Search, by its own statement, ignores the file.
Also called: content freshness, freshness, last updated date, publication date, byline date, content recency, content maintenance
ObservedClaimedGoogle documents how to expose dates, not how much freshness weighs. The citation pattern comes from one vendor study (Ahrefs, about 17 million citations), in which Google AI Overviews stayed close to organic results.
Content freshness and dating means keeping pages current and showing clear publication and update dates, so engines and readers can judge whether information still holds. A 2025 Ahrefs vendor study found AI assistants cited somewhat newer pages than Google's organic results, ChatGPT most of all; no engine publishes how it weighs freshness.
Freshness has two parts. Maintenance is the substance: updating facts, figures, prices and recommendations when they change, and removing what is obsolete. Dating is the signal: a visible "Published" or "Last updated" date on the page, matched by datePublished and dateModified in structured data, so machines and people know when the content was written or significantly revised.
The lever matters most for time-sensitive topics (prices, regulations, software features, statistics) and much less for stable reference material. It excludes changing dates without changing content, which misleads readers and which Google's John Mueller has warned against, as Ahrefs' study notes.
How it works
Google's byline date documentation says its systems estimate when a page was published or significantly updated from several factors, and recommends a prominent, labeled visible date plus datePublished and dateModified in structured data using a CreativeWork type such as Article or BlogPosting. For AI answers, Ahrefs analyzed about 17 million cited URLs across ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews (published 2025-07-28). By its first-seen measure, cited URLs averaged 25.7% fresher than organic results, ChatGPT favored newer content most, and AI Overviews stayed close to the organic average; the average cited page was still 2.9 years old. Engines also differ in how often they fetch: Vercel's crawl analysis saw AI tools apparently relying on cached or training data instead of fetching fresh pages.
Why it matters
For questions whose answers change (current pricing, this year's rules, the latest version), an engine that sees a recent date and current figures has a reason to prefer your page, and an outdated page risks being cited for wrong information. What a team does: set a review schedule for time-sensitive pages, update the substance before the date, show "Last updated" visibly, keep structured data dates in sync, and retire or redirect pages that can no longer be kept accurate. Do not expect date changes alone to move citations; the vendor evidence shows a modest tilt, and almost none for Google AI Overviews.
Example
A hypothetical HR software company's 2025 state minimum wage guide still ranks in early 2026. The team updates every rate, retitles the page for 2026, adds "Last updated: January 12, 2026" above the first paragraph with the matching dateModified value, and links an archived copy of the 2025 version.
Also called: original research, first-party data, proprietary data, data studies, primary research, non-commodity content
DocumentedObservedGoogle documents that unique, non-commodity content is the strongest long-run factor for its generative AI features (Google scope). The citation advantage of being the primary source is an inference with no controlled study.
Original research and data studies are first-party findings a publisher produces itself, such as surveys, benchmarks, experiments or usage data, which make its site the primary source for a fact. Google's 2026 guidance names unique, non-commodity content as the most influential factor for its AI features; no study has measured a citation advantage.
Original research covers information that did not exist before the publisher produced it: a survey of customers or an industry, a product benchmark, an analysis of anonymized usage data, a controlled experiment, a first-hand test or review. Its defining property is that anyone else who uses the finding has to refer back to it.
Google's 2026 generative AI guide frames the same idea as non-commodity content: material offering "unique expert or experienced takes" beyond common knowledge, as opposed to summaries that restate what exists elsewhere. The lever is distinct from cite-worthy statistics, which can quote other people's numbers, and from information gain as a retrieval concept. Being the origin of a number gives a page the strongest claim to attribution; no engine or writer is obliged to honor it.
How it works
Google's guide says content people find unique, compelling and useful "will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions" it gives, and advises against recycling what others have said or what a generative AI model could easily produce. The reasoning for citations is mechanical but untested: when an engine grounds an answer in a figure, the page that first published it is the most direct source to cite, and each attributed mention elsewhere adds a path back to the origin. The GEO paper's statistics result points the same way, though it tested numbers added to existing pages, not new data.
Why it matters
Commodity content competes with every other summary of the same facts and with what the model already knows. A first-party dataset gives an engine something it can only get from you, and gives journalists, analysts and other publishers a reason to mention you, which feeds the third-party mentions AI answers draw on. What a team does: pick a question its market asks that nobody answers with data, collect data with a stated method and sample, publish the method, date and key figures as text, and repeat the study on a schedule. Expect occasional paraphrase without credit: attribution is likely, never guaranteed.
Example
A hypothetical scheduling-software company analyzes 2.3 million anonymized appointments booked through its product in 2025 and publishes no-show rates by industry and day of week, with its method and sample described. When an assistant answers "What is the average no-show rate for dental appointments?", the company's page is the origin of the figure, though whether the answer credits it depends on the engine.
Also called: internal links, internal linking, site architecture, crawlable links, orphan pages, link discoverability
DocumentedDocumented by Google for its Search and AI features (Google scope). Vercel's crawl data adds field observation on AI crawlers' broken-URL requests.
Internal linking and discoverability means connecting a site's pages with crawlable links and descriptive anchor text so crawlers can find every page worth citing. Google lists findability through internal links among the SEO fundamentals for AI Overviews and AI Mode, and says every important page should be linked from at least one other page.
Discoverability is the precondition for every other lever: a page crawlers never find cannot be indexed, retrieved or cited. Internal links are the main way crawlers move through a site, alongside XML sitemaps. Google's link documentation specifies what counts as crawlable, generally an a element with an href attribute, and notes that it cannot reliably extract links that work only through script events.
The lever includes anchor text (the visible words of a link, which tell people and Google what the target is about), navigation and hub pages, contextual links within body text, and the elimination of orphan pages that no other page links to. External backlinks are a separate topic, as is organizing topics into clusters, which builds on internal links.
How it works
Google's AI features page lists "making your content easily findable through internal links on your website" among the fundamentals that apply to AI Overviews and AI Mode, which draw on pages in Google's index. Google's link best practices add that "every page you care about should have a link from at least one other page on your site" and that descriptive anchor text helps both people and Google. For AI crawlers, Vercel and MERJ's December 2024 analysis found that 34.82% of ChatGPT's fetches and 34.16% of Claude's hit 404 pages, and recommended maintaining redirects, current sitemaps and consistent URL patterns.
Why it matters
Pages reachable only through search boxes, filters or script-driven navigation may never reach an engine's index, and every AI answer built on that index inherits the blind spot. Internal links also signal which pages a site considers important. What a team does: make sure each page worth citing is linked from at least one relevant, crawlable page with descriptive anchor text; link new articles from hubs and related posts; fix broken internal links and redirect chains; and keep the XML sitemap current. This is standard SEO hygiene, and Google states it applies unchanged to its AI features.
Example
A hypothetical analytics company publishes a strong benchmark report but links to it only from a JavaScript-driven "Resources" filter, and months later the page has no organic or AI visibility. Links from the homepage, the related product page and three relevant blog posts, with anchor text such as "2026 dashboard load-time benchmark", give crawlers paths to it.
Also called: technical SEO basics, crawlability, indexability, snippet eligibility, technical floor, WAF allowlisting, bot access
DocumentedDocumented by Google (Search and AI features), Perplexity (crawler access through firewalls) and Cloudflare (default AI crawler blocking for new domains); each statement applies to its own scope.
Page experience and crawlability basics are the technical floor for appearing in AI answers: crawlers allowed by robots.txt, CDNs and firewalls, pages indexed and eligible for snippets, and a usable page. Google says its AI features add no technical requirements beyond these, apart from a Search Console control that includes sites by default.
This entry gathers the conditions a page must meet before any content lever can matter. Access: robots.txt allows the relevant crawlers, and nothing upstream (a CDN rule, a web application firewall, a bot-fighting feature, a login wall) blocks them. Indexing: the page returns a 200 status, carries no noindex, and is indexed. Snippet eligibility: preview controls such as nosnippet or a restrictive max-snippet are not set by accident, since Google requires snippet eligibility. Experience: the page loads quickly, works on every device and makes its main content easy to distinguish.
Each engine sets its own floor. Google documents its requirements; Perplexity documents how to let its crawlers through firewalls; other engines document their user agents with varying detail.
How it works
Google's AI features page says a page must be "indexed and eligible to be shown in Google Search with a snippet", adds that there are no additional technical requirements, and lists "ensuring that crawling is allowed in robots.txt, and by any CDN or hosting infrastructure" and a great page experience among its best practices. Its 2026 guide adds the Search generative AI control in Search Console, which includes sites by default and lets owners exclude them from AI Overviews, AI Mode and Discover's generative features; Google says it reached all websites worldwide by August 31, 2026. Infrastructure can block silently: since July 1, 2025, Cloudflare blocks AI crawlers by default on new domains and asks each one at sign-up whether to allow them, and Perplexity documents how to allowlist PerplexityBot and Perplexity-User by user agent plus published IP ranges.
Why it matters
Many AI visibility failures that a team can fix quickly are access failures. A firewall rule, a CDN bot setting, a stray noindex or a nosnippet tag can remove a site from AI answers while its pages look fine in a browser. What a team does: list the crawlers it wants (Googlebot, Bingbot, OAI-SearchBot, PerplexityBot, Claude-SearchBot and others), confirm robots.txt and CDN or firewall settings allow them, check server logs for successful fetches, verify indexing in Google Search Console and Bing Webmaster Tools, review the Search generative AI control, and fix speed and mobile issues. None of this guarantees inclusion, as Google states plainly.
Example
A hypothetical B2B site switches on its CDN's option to block AI bots during a scraping scare, and weeks later its pages stop appearing as sources in Perplexity answers. Server logs show PerplexityBot receiving 403 responses; an allow rule matching the bot's user agent and published IP ranges, placed above the blocking rules as Perplexity's documentation describes, restores access.
Also called: multimodal content, image SEO, video SEO, alt text, alternative text, visual content optimization
DocumentedDocumented as a recommendation in Google's AI features guidance and in Microsoft's Bing guidance; neither engine states an effect size.
Multimodal content means supporting a page's text with relevant images and video, each made machine-readable through alt text, captions, nearby text and metadata. Google recommends high-quality images and video for its AI features, which can show visual results, and Microsoft advises alt text or HTML for critical image details. No engine publishes an effect size.
Multimodal content covers a page's non-text assets and the text that makes them understandable to machines: descriptive alt attributes, captions, filenames, surrounding text, and metadata such as VideoObject structured data, video sitemaps or Open Graph tags. The assets must be indexable; for video, Google requires an indexed watch page with the video embedded and not hidden behind other elements.
The lever has two halves. Visual assets can themselves appear in AI answers and visual results, and their text equivalents ensure that whatever an image conveys is also available as text. Critical facts should not live only inside an image, which is an extractability problem: alt text describes an image and is no substitute for putting key information in the page text.
How it works
Google's AI features page recommends "supporting your textual content with high-quality images and videos, when applicable", and its 2026 generative AI guide says these features "can bring in relevant images and video", so a site can appear beyond plain links; it adds that following Google's image and video SEO documentation already covers this. Google's image guidance says it uses alt text together with computer vision and page content to understand an image, and warns against stuffing alt attributes with keywords. Microsoft's Bing guidance says to "always provide alt text or present critical details in HTML" so assistants understand image content reliably.
Why it matters
Google says its generative AI features can show images and video, which makes visual assets a second route into an answer, and an asset without text cues is hard for any machine to interpret. A key figure shown only in a chart is also invisible to a text-only crawler. What a team does: add original, relevant images and video where they help the reader; write alt text that states what the image shows; caption charts with their main finding; publish video on indexable watch pages with a transcript or summary; and keep critical numbers in the page text. Expect gains in usability and image or video search; the AI citation effect is unmeasured.
Example
A hypothetical outdoor-gear retailer publishes a tent comparison with original photos. Its old alt text read "tent1.jpg"; the new alt text reads "Two-person backpacking tent pitched on rocky ground with the rainfly attached", and each caption repeats the packed weight listed in the comparison table, which also exists as HTML text.
Also called: hreflang, multilingual SEO, international SEO, localized versions, language variants, x-default
ObservedClaimedGoogle documents hreflang for its Search; no engine documents how language affects AI citations. The finding that ChatGPT runs many fan-out searches in English comes from one vendor study (Peec AI, February 2026).
Multilingual and hreflang coverage means publishing language and regional versions of key pages and declaring them with hreflang, so search systems can send each user to the right version. Google documents hreflang for Search, which its AI features draw on; how language shapes AI citations is undocumented, and vendor data suggests ChatGPT often searches in English.
Multilingual coverage is the presence of real translated or localized versions of important pages; hreflang is the annotation, placed in HTML link elements, HTTP headers or the sitemap, that tells Google which URLs are language or regional variants of one another. Each version must list itself and every other version, and an x-default value can name a fallback page for unmatched languages. Google states that it does not use hreflang or the HTML lang attribute to detect a page's language; it determines language algorithmically.
The lever covers localized versions of the content an engine might cite (product facts, pricing, policies, support answers), connected by hreflang and kept consistent across languages.
How it works
In classic Google Search, hreflang helps Google "point users to the most appropriate version of your page by language or region". Google's 2026 guide says its generative AI features rely on its core Search ranking systems to retrieve pages, so the same annotations plausibly carry over, though Google does not say so for AI features specifically. Other engines may behave differently: Peec AI, a vendor of AI search analytics, reported in February 2026 that in its data 43% of ChatGPT's fan-out searches for non-English prompts ran in English, and that English searches supplemented nearly 78% of such sessions. If that pattern holds, a brand with content only in a local language may be missing from part of the research behind a local-language answer.
Why it matters
People ask assistants in their own language, and an answer can only cite pages the engine found and judged relevant in that language or in the languages it chose to search. What a team does: pick the markets where it wants to be recommended, localize the pages that carry cite-worthy facts, connect versions with complete hreflang sets and an x-default, keep facts identical across languages, and test prompts in each language on the assistants that matter. If ChatGPT does research partly in English, a solid English version of key facts helps even when the target market speaks another language.
Example
A hypothetical German project-management vendor has detailed German pages but a thin English site. Testing German-language prompts in ChatGPT, it finds competitors with thorough English comparison pages cited more often. It expands its English pages, links the German and English versions with hreflang and an x-default, and tracks citations in both languages.
Also called: Google Business Profile, GBP, local SEO for AI, NAP consistency, Merchant Center, business listings
DocumentedDocumented by Google for its AI features and local results (Google scope). NAP consistency is long-standing practitioner guidance; no engine documents it as an AI factor.
Local and business-profile signals are the business facts engines use in local and commercial answers: Google Business Profile details such as category, address, hours and reviews, Merchant Center product data, and consistent name, address and phone listings. Google lists up-to-date Business Profile and Merchant Center information among its best practices for AI features.
This lever covers structured business information that lives outside a website's pages: the Google Business Profile (name, category, address or service area, hours including special hours, phone, website, attributes, photos and reviews), Merchant Center product feeds, and listings on maps, directories and review sites. NAP consistency, meaning the same name, address and phone number everywhere, is the classic local floor.
Google's 2026 generative AI guide says AI responses can include product listings and information about local businesses, and that Merchant Center feeds and Business Profiles help products and services appear in AI responses and other results. Google's local ranking help names three factors for local results: relevance, distance and prominence. The entry concerns the accuracy and completeness of these facts; ads are a separate topic.
How it works
Google's AI features page asks site owners to check "that your Merchant Center and Business Profile information is up-to-date". Its Business Profile help says businesses with complete and accurate information are more likely to show up in local results, that verification helps, that replying to reviews shows customers their feedback is valued, and that "there's no way to request or pay for a better local ranking on Google". The 2026 guide also points merchants to Business Agent, a conversational experience on Google Search where customers can chat with a brand. How other assistants source local business facts is largely undocumented.
Why it matters
For "near me", "open now" and product availability questions, an AI answer is only as accurate as the business data behind it. Stale hours, a wrong category or an outdated price can be repeated to many people at once, and a missing profile can leave a business out of the answer entirely. What a team does: claim and verify the Business Profile, complete every relevant field, keep regular and holiday hours current, respond to reviews, maintain Merchant Center feeds with correct prices and availability, and keep name, address and phone consistent across directories and the website. Then test local prompts in the assistants customers use.
Example
A hypothetical bakery in Portland changes its Sunday hours but updates only its website. An assistant asked "Is the bakery on Hawthorne open Sunday morning?" gives the old hours still listed on its Business Profile. Updating the profile, including special holiday hours, corrects the source the answer most likely drew on.
Also called: entity-based SEO, entity optimization, entity-first SEO, semantic SEO
DocumentedObservedGoogle documents entity understanding (the Knowledge Graph) and the markup that helps it disambiguate organizations; entity SEO as a named practice, and its effect on AI answers, rest on field observation.
Entity SEO is the practice of making a brand, person, product or concept recognizable to search systems as one distinct thing with stable facts, instead of a string of words. It relies on consistent naming, disambiguating descriptions, structured data and corroborating third-party sources, so that engines and the AI answers built on them resolve the right entity.
In search, an entity is a thing whose identity persists across names, spellings and languages: a company, a person, a place, a product, a concept. Google framed the shift in 2012, when it introduced the Knowledge Graph as a way to understand "things, not strings." Entity SEO applies that idea to a publisher's own presence. It covers choosing one canonical name and using it everywhere, writing a short description that separates the entity from namesakes, publishing structured data (Organization, Person, Product) that states the entity's facts and links its official profiles, and making sure independent sources describe the same facts the same way.
The boundary: entity SEO concerns the identity of a thing and the facts attached to it. It overlaps with topical authority (how completely a site covers a subject) and with keyword targeting (matching query wording), and it is distinct from both. It is an industry label, so its exact scope varies by practitioner.
How it works
Search systems map queries and documents to entities, then use what they hold about each entity to interpret and assemble results. Google's Knowledge Graph is the best-documented example: a database of facts about people, places and things compiled from many sources, licensed data and content owners. Google's Organization structured data guide says the markup helps Google tell one organization apart from others; it recommends sameAs links and identifiers such as iso6523Code, leiCode and vatID, while properties such as logo shape how the organization appears in results and its knowledge panel. When a site's markup, its official profiles and third-party references agree, a system has less ambiguity to resolve. When they conflict, two namesakes can be merged, or one entity can be split into fragments.
Why it matters
An AI answer that names a company has to decide first which company it means. Many answer engines ground their responses in search results, so the entity understanding of the underlying search system travels into the answer: a brand whose name, category and founders are stated inconsistently can be confused with a namesake, described with stale facts, or left out. No engine publishes how much entity clarity weighs in citation selection, so the evidence is observational. The work itself is low-risk: one canonical name, Organization markup that matches visible text, consistent official profiles, and errors corrected at their source.
Example
A payroll startup shares its one-word name with a regional bank and a mobile game. It adopts one descriptor everywhere ("the payroll software for restaurants"), publishes Organization markup with its legal name, logo and sameAs links to its LinkedIn and Crunchbase pages, and asks two directories to fix an outdated founding year. Engines now receive consistent signals about which entity the name refers to; when each system picks them up is outside the startup's control.
Also called: brand as an entity, organization entity, company entity, machine-readable brand identity
ObservedGoogle documents the inputs (Organization markup, knowledge panel claims); how answer engines resolve a brand entity is not documented by any engine and rests on field observation.
Brand entity means the machine-readable identity of a company or product: its canonical name, logo, description, official website and profiles, plus core facts such as founders, founding date and category. Search systems and AI answers rely on that identity to decide which organization a question is about and which facts to state about it.
Entity SEO is the practice; the brand entity is the object it works on. It is the set of facts a system holds about a brand and the identifiers that tie those facts together: the official URL, a logo, a legal name and any alternate names, profiles on social and business networks, registry identifiers, and records in knowledge bases such as Wikidata. A well-formed brand entity has facts that are complete, current and consistent across sources, so a system can resolve every mention of the brand to the same record.
It is separate from brand reputation (what people think of the brand) and from brand awareness (how readily it comes to mind), though AI answers draw on all three. It is also wider than a knowledge panel: the panel is one display of the entity on Google, while the entity exists wherever a system stores facts about the brand.
How it works
Google documents several inputs. Its Organization structured data guide says the markup helps Google understand an organization's administrative details and disambiguate it from other organizations, and that some properties influence the logo and knowledge panel shown in results; no property is required, and every applicable one is recommended. Its Knowledge Panel help lets an official representative claim a panel by signing in through an associated account (YouTube, Search Console and others) and then suggest changes. Facts in the Knowledge Graph come from many sources, licensed data and content owners. Other engines do not document their entity pipelines, so the same inputs are a reasonable bet there, without a documented mechanism.
Why it matters
When someone asks an assistant what a company does, who founded it or how it compares with rivals, the answer depends on whether the system resolved the right brand and holds correct facts. Namesakes, rebrands, acquisitions and inconsistent descriptions are the usual causes of conflated or outdated answers. A team can audit how major engines describe the brand, list every place its core facts appear (site, markup, profiles, directories, Wikidata), fix discrepancies at the source, and claim the Google knowledge panel where one exists. Expect lag, since each system refreshes on its own schedule.
Example
A company renamed from "Pinecrest Labs" to "Pinecrest" updates its website, its Organization markup (with alternateName set to the old name), its LinkedIn page and its Crunchbase record in the same week, then corrects the label on its Wikidata item. Assistants that described the old name and product line begin, each on its own schedule, to use the new name and category.
DocumentedGoogle's own documentation, authoritative for Google; how AI Overviews and AI Mode consult the graph is not documented.
Google Knowledge Graph, usually called the Knowledge Graph, is Google's database of facts about people, places, organizations and things, and of the relations between them. Introduced in 2012, it powers knowledge panels and helps Google interpret what a query or page is about. Its facts come from public sources, licensed data and content owners who claim their panels.
Google describes the Knowledge Graph as its database of billions of facts about people, places and things. At launch on 2012-05-16, Google said it held more than 500 million objects and more than 3.5 billion facts, rooted in public sources including Freebase, Wikipedia and the CIA World Factbook. Each entity has a machine identifier (the public Knowledge Graph Search API returns IDs of the form kg:/m/...) and typed properties expressed with schema.org vocabulary.
The term refers to Google's system. Other companies run their own knowledge graphs, and "knowledge graph" is also a generic computer-science term for any graph of entities and relations. The knowledge panel shown beside search results is one display built from the Knowledge Graph.
How it works
According to Google's Knowledge Panel help, facts come from a variety of sources that compile factual information, from licensed data for items such as sports scores and weather, and directly from content owners, including people who claim a panel and suggest changes. Google says it relies mainly on automated systems and analyzes what its public reporting tools capture; anyone can send feedback through the Feedback link on a panel. Developers can query the graph read-only through the Knowledge Graph Search API, which returns JSON-LD; Google warns that the API is not suitable as a production-critical service and is migrating it to Cloud Enterprise Knowledge Graph.
Why it matters
For Google surfaces, the Knowledge Graph is where an entity's facts are reconciled, and Google's AI features run inside the same Search product. Google does not document how AI Overviews or AI Mode consult the graph, so the link is plausible and unconfirmed. The actions are concrete: check whether the brand has a knowledge panel, claim it where possible, correct wrong facts through the Feedback link, and keep the sources the graph draws on (official site and markup, Wikipedia, Wikidata, major profiles) consistent. Inconsistent naming across those sources is a common route to being merged with a namesake.
Example
A founder searches her company's name and finds a knowledge panel listing the wrong founding year. She claims the panel by signing in through the company's verified Search Console property, suggests the correction, and fixes the same date on the company's About page and its Wikidata item, so that every source the graph might read agrees.
Also called: schema.org sameAs, sameAs property, sameAs links, sameAs markup
DocumentedObservedDocumented as schema.org vocabulary and in Google's Organization and Article guidance; any effect on AI citations is unmeasured.
sameAs is a schema.org property that links an entity described on a page to other web pages that unambiguously identify the same entity, such as its Wikidata item, Wikipedia article or official social profiles. It helps search systems reconcile scattered mentions into one identity for an organization, a person or a product.
Schema.org defines sameAs as the "URL of a reference Web page that unambiguously indicates the item's identity." It is a property of Thing, so it is available on every schema.org type, and its expected value is a URL. Publishers use it inside Organization, Person, Product or LocalBusiness markup, usually in JSON-LD, listing one or more URLs.
Google's Organization structured data guide recommends sameAs for pages on other websites with additional information about the organization, such as a social media or review profile, and its Article guidance accepts sameAs as a way to identify an author. The property asserts identity and does not replace visible content. Pointing sameAs at pages about a different entity, or at pages that do not describe the entity at all, adds noise to the identity it is meant to clarify.
How it works
A system reading the markup gets a list of URLs that the publisher claims describe the same entity. If those pages (a Wikidata item, a LinkedIn company page, a Crunchbase profile) carry matching names, logos and facts, the connection is easier to make with confidence. Because sameAs is the publisher's own assertion, it is most useful when the linked pages independently confirm it.
Why it matters
Being described correctly in an AI answer starts with being resolved to the right entity, and sameAs is one of the few reconciliation signals a publisher fully controls. No engine has published evidence that sameAs changes citation rates, so treat it as low-cost hygiene. Teams typically add Organization markup on the homepage with sameAs links to the Wikidata item, LinkedIn, Crunchbase and official social accounts, add Person markup with sameAs for named authors, and update the list whenever a profile is created, renamed or retired.
Example
An analytics company's homepage carries Organization markup in JSON-LD whose sameAs array lists four URLs: its Wikidata item, its LinkedIn company page, its Crunchbase profile and its GitHub organization. Each of those pages names the same company, links back to the same domain and shows the same logo, so the site and the four profiles point to one identity.
Also called: Wikipedia article, Wikipedia coverage, Wikipedia page for a company, Wikipedia notability
MeasuredObservedCitation frequency: one Pew Research Center study of Google AI summaries (US, March 2025). Wikipedia as a lever for AI visibility: observed, not measured.
Wikipedia presence means a brand, person or product being covered accurately on Wikipedia, in its own article or within existing ones. It matters for AI search because Wikipedia is among the most cited sources in AI summaries and is widely reused as reference data, but an article has to be earned through independent coverage; the subject cannot simply write one.
Wikipedia presence covers two situations: a dedicated article about the subject, and accurate mentions of the subject in other articles (an industry, a technology, a founder's biography). Both are governed by Wikipedia's community policies. Under the general notability guideline, a topic qualifies when it has "significant coverage in reliable sources that are independent of the subject," and the separate guideline for organizations and companies applies stricter coverage requirements; advertising, press releases, autobiographies and the subject's own website do not count as independent.
The conflict-of-interest guideline strongly discourages people from editing articles about themselves, their employer or their clients. Paid editors must disclose their employer and client under the Wikimedia Foundation's Terms of Use and are expected to propose changes on the article's talk page for volunteers to review. Presence is therefore an outcome of being written about elsewhere, outside the subject's direct control.
How it works
In Pew Research Center's analysis of the browsing of 900 US adults in March 2025 (68,879 Google searches, 12,593 of which produced an AI summary), Wikipedia, YouTube and Reddit were the most frequently cited sources in both AI summaries and standard results. Together they made up 15% of the sources listed in AI summaries and 17% in standard results, and Pew found Wikipedia links somewhat more common in AI summaries than on standard result pages. That is one study, of Google only, at one point in time.
Why it matters
Wikipedia is cited by answer engines and reused by knowledge bases, so accurate coverage there can shape how a brand is described across many systems at once. The path is slow and indirect: earn independent coverage in reliable publications first, then, if the notability bar is met, submit a draft through Articles for Creation or a talk-page edit request with full disclosure. Self-written promotional articles run against both the notability and conflict-of-interest guidelines, and the guideline itself warns that such editing can embarrass the people and companies it promotes. Treat it as long-horizon entity work with no guaranteed outcome.
Example
A climate-tech startup is covered in depth by two national newspapers and a trade journal after a funding round. Its communications lead does not write the article herself: she discloses her affiliation on her user page and submits a sourced draft through Articles for Creation, leaving volunteer editors to decide whether the company meets the notability bar.
Also called: Wikidata item, QID, Wikidata entry, Wikidata record
DocumentedObservedWikidata's structure, license and notability rules are documented by Wikidata; which AI systems consume it, and how often they refresh, is not documented engine by engine.
Wikidata is the free, multilingual knowledge base run by the Wikimedia community that stores facts as structured statements about items, each with a permanent identifier such as Q42. Its data is dedicated to the public domain (CC0), so search engines, assistants and data projects can reuse it, which makes an accurate item a widely shared reference for an entity.
Wikidata describes itself as a free, collaborative, multilingual, secondary knowledge base collecting structured data. Its building blocks are items, identified by a Q followed by a number (Douglas Adams is Q42); properties, identified by a P and a number (educated at is P69); and statements that pair a property with a value, ideally with a reference to a source. Each item has a label, a description and optional aliases per language. The structured data is released under the CC0 public domain dedication, so anyone may copy and reuse it.
An item must meet at least one notability criterion: a valid link to a Wikimedia page such as a Wikipedia article; being a clearly identifiable entity that can be described using serious and publicly available references; or filling a structural need for other items. Wikidata holds facts while Wikipedia holds prose, and a company can qualify for an item without qualifying for an article.
How it works
Editors and approved bots add statements, and anyone can reuse the result through the website, the query service or full data dumps. Google's documentation for its Knowledge Graph Search API points developers to Wikidata dumps when they need relationships between entities. A correction made on Wikidata reaches a downstream system only when that system next ingests the data, on a schedule the editor does not control.
Why it matters
Because Wikidata is open, structured and reused, an accurate item gives every system that reads it the same official name, website, founding date, industry and identifiers, along with links to the brand's other profiles. That makes it one of the cheapest ways to reduce entity ambiguity, provided the entity meets the notability rules and each statement carries an independent reference. Teams should check for an existing item before creating one, keep descriptions neutral, cite a source for each statement, and link the item from the site's sameAs markup. No engine documents how much weight it gives Wikidata.
Example
A B2B software company covered by two trade publications creates a Wikidata item with its official website, inception date, headquarters location and industry, each statement referenced to one of those articles. It then adds the item's URL to the sameAs list in its homepage markup, so its own site and the open record point to each other.
Also called: E-A-T, Double-E-A-T, EEAT, Experience, Expertise, Authoritativeness, Trustworthiness
DocumentedClaimedDocumented by Google as a quality-rater concept and a description of what its systems aim to reward; claims of a direct E-E-A-T boost inside AI answers are vendor or community assertions.
E-E-A-T stands for experience, expertise, authoritativeness and trustworthiness, the qualities Google's search quality raters use to judge whether content is credible, with trust named as the most important. Google presents E-E-A-T as a guide to what its systems should reward, with extra weight on health, finance and safety topics, and states that it is not a specific ranking factor.
Google introduced E-A-T in its search quality rater guidelines and added a second E, for experience, on 2022-12-15, asking whether content shows first-hand use of a product, a visit to a place or a lived experience. Google's guidance on helpful content defines the four aspects and says that trust is the most important of them. It also states that "E-E-A-T itself isn't a specific ranking factor." Raters are people who give Google feedback on whether its results seem good; Google says rater data is not used directly in its ranking algorithms and that raters have no control over how pages rank.
E-E-A-T carries extra weight on "Your Money or Your Life" (YMYL) topics, those that could significantly affect people's health, financial stability or safety. The term names Google's quality concept. Other engines do not use it, though they face the same need to ground answers on credible sources.
How it works
Raters are trained to assess E-E-A-T using the criteria in the rater guidelines, and their feedback helps Google evaluate whether its systems are returning good results. For publishers, the visible expression is the set of questions in Google's helpful-content guidance under "Who, How, and Why": accurate authorship with bylines that lead to the author's background, disclosure of how content was produced (including the use of automation or AI where readers would reasonably expect it), and a clear purpose of helping people. The same guidance warns against deceptive authorship, such as invented names or false credentials.
Why it matters
Answer engines have a practical reason to prefer sources they can check: a wrong claim in a generated answer reflects on the engine that produced it. E-E-A-T-shaped evidence on a page (a named, qualified author, citations to primary sources, an editorial policy, accurate contact details) makes content easier to verify and safer to quote. No engine has published evidence of an E-E-A-T boost inside AI answers, so treat such claims as unproven. The defensible move is to make credibility visible on the pages where accuracy matters most, starting with YMYL content.
Example
A health publisher's article on medication interactions names its author, a licensed pharmacist, links to her credentials, cites drug labels and peer-reviewed studies, and shows a last-reviewed date with a link to the editorial policy. A competing page with the same advice, no author and no sources gives a reader, a rater or an answer engine far less to check.
DocumentedObservedGoogle documents byline and author-markup guidance; no engine documents author-level weighting when selecting citations for AI answers.
Author authority and bylines together form the visible, checkable evidence that a named person with relevant expertise wrote or reviewed a piece of content: a byline, a bio, credentials and a track record elsewhere. Bylines and author markup make that evidence readable by people and machines, though no engine documents weighting individual authors when it chooses citations.
The practice has three parts. On the page, a byline names the author and links to a profile describing background and areas of expertise. In markup, Article structured data identifies the author by name and by a URL that uniquely identifies the person, such as a profile page or a social profile, using the Person type for people and the Organization type for organizations, with each author in a separate author field. Off the page, the same person has a consistent footprint: publications, talks and professional profiles, ideally tied together with sameAs.
Google's helpful-content guidance encourages accurate authorship information wherever readers would expect it and warns against deceptive authorship, such as made-up names, AI-generated headshots or false credentials. The term concerns individual people. The reputation of the publication itself is related and separate.
How it works
Google's Article markup guidance asks that author.name contain only the person's name, with job titles in jobTitle and publisher names in publisher, and that author.url point to a page that uniquely identifies the author; sameAs is an accepted alternative, and an internal profile page can carry ProfilePage markup. These conventions are meant to let a system connect a byline to one person across pages and sites.
Why it matters
AI answers attribute claims to sources, and a source with a named, verifiable expert is easier to rely on than an anonymous page. Even without documented author weighting, bylines help on two fronts that are documented: Google's quality guidance favors clear authorship, and readers who click through judge credibility by who wrote the piece. Teams can replace generic bylines such as "Staff" on expert content, build real author pages, mark them up consistently, and keep each author's name identical across the site and external profiles.
Example
A cybersecurity vendor republishes its incident-response guide under the name of its head of threat research instead of "Marketing Team". The byline links to her profile page, which lists her certifications and conference talks and carries ProfilePage and Person markup with sameAs links to her LinkedIn and GitHub accounts.
Also called: digital PR for GEO, earned media for AI search, AI search PR, GEO PR
MeasuredClaimedOne study (Chen et al., 2025) measured a strong preference for earned media in three AI search services; the effect of specific PR campaigns on AI visibility rests on vendor case studies.
Digital PR for AI answers is the work of earning coverage of a brand in independent publications and reference sites that AI search engines already cite, so the brand appears in answers through third-party sources as well as through its own site. One 2025 study found that AI search services lean heavily on such earned media.
The practice adapts classic digital PR (original data, expert commentary, newsworthy announcements offered to journalists) to a new goal: being among the sources an answer engine retrieves and cites for the questions buyers ask. Targets are chosen by observing which domains engines cite for a category, then earning accurate mentions there through stories, data studies, expert quotes, reviews or inclusion in comparisons. Andreessen Horowitz's 2025 essay on GEO framed the related metric as how often a brand is cited or used as a source in model answers.
Digital PR differs from link building, which pursues ranking signals for a page, and from brand mention seeding, which places mentions without any editorial gatekeeper. Whether coverage happens stays the publication's editorial decision.
How it works
Chen, Wang, Chen and Koudas (arXiv:2509.08919, submitted 2025-09-10) ran controlled experiments across verticals, languages and query paraphrases, comparing ChatGPT, Perplexity and Gemini with Google search. They report a "systematic and overwhelming bias towards Earned media" in the AI services, against brand-owned and social content, while Google's sources were more balanced, and they recommend building earned media to establish authority with AI systems. Martinez's 2026 critical survey of 45 GEO studies adds a caution: no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability.
Why it matters
If answer engines draw mostly on third-party sources for a category, a brand that exists only on its own domain is easy to leave out. Earned coverage puts the brand inside the sources engines already trust and gives them independent corroboration of its claims. The evidence is one study for the bias and vendor case studies for campaign results, so measure: track which domains engines cite for target prompts, pursue coverage in those outlets with real news or data, and compare citation share before and after on a fixed prompt panel.
Example
A payroll software company publishes an annual survey of 2,000 small-business owners on late payments. Two trade publications and a national business site that engines often cite for payroll questions report the findings and name the company. On the company's tracking panel, prompts such as "how common are late payroll runs for small businesses" start returning answers that cite those articles.
Chen, Wang, Chen and Koudas, Generative Engine Optimization: How to Dominate AI Search (arXiv:2509.08919), published 2025-09-10, read 2026-11-07: https://arxiv.org/abs/2509.08919
Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026) (arXiv:2607.14035), published 2026-07-15, read 2026-11-07: https://arxiv.org/abs/2607.14035
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Cohen and Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: mention seeding, brand seeding, seeded brand mentions
ClaimedA named practice in vendor glossaries with no published evidence of efficacy; its risks are documented in platform and search-engine spam policies. Recorded because the vocabulary exists; this glossary does not recommend it.
Brand mention seeding is the deliberate placement of brand mentions across forums, Q&A sites, comment threads and similar spaces in the hope of influencing what AI models learn and retrieve. Its efficacy is unproven, and when done covertly or at volume it breaches platform spam rules and Google's spam policies, which makes it a reputational risk.
The term appears in vendor glossaries for a tactic that sits between community marketing and astroturfing. In its benign form, a company's staff answer questions in communities where they genuinely participate and disclose their affiliation. In its problematic form, mentions are posted from undisclosed, multiple or automated accounts, or produced at volume, to make a brand look more discussed than it is.
This glossary records the term because the vocabulary is in use; it does not recommend the tactic. Seeding differs from digital PR, where an editor decides whether to cover the brand, and from customer advocacy, where real users speak in their own words.
How it works
Platforms set the limits. Reddit's spam policy defines spam as repeated or unsolicited actions, automated or manual, that negatively affect users, communities or Reddit, names the use of tools such as bots and generative AI to spread spam, and asks users whose posts mainly link to a business they benefit from to watch their posting frequency or use Reddit's ad platform; community moderators add their own rules. On the search side, Google's spam policies now explicitly cover attempts to manipulate generative AI responses in Google Search, and define user-generated spam as spammy content added to a site through a channel intended for user content.
Why it matters
The appeal is plain: engines cite discussion platforms often, and a mention there costs little. The risks are concrete. Moderators remove seeded posts, platforms ban accounts, spam systems on platforms and search engines are built to catch coordinated or automated posting, and exposure as astroturfing damages trust in the brand on the very platforms that shape its reputation. No study shows that seeding raises AI visibility durably. The defensible alternative is disclosed participation by real employees, and a product customers choose to talk about.
Example
An agency offers a client 200 Reddit comments a month recommending its project-management tool from aged accounts. Within weeks, several subreddits remove the comments and ban the accounts under their rules against undisclosed self-promotion, and a thread calling out the pattern starts appearing in results for the brand's name.
Also called: user-generated content platforms, UGC citations, forum citations, Reddit citations, community platforms in AI answers
MeasuredObservedCitation frequency: one Pew Research Center study (Google AI summaries, US, March 2025). The Google and OpenAI data partnerships with Reddit are documented by the companies; UGC presence as a lever is observed.
UGC platforms as cited sources (Reddit, Quora, forums) describes the role of discussion sites, where users write the content, in AI answers. Engines cite them often for opinions, comparisons and first-hand experience, and both Google and OpenAI have data partnerships with Reddit. How a brand is discussed on these platforms can shape how AI answers describe it.
User-generated content (UGC) platforms host posts, answers and threads written by their users rather than by an editorial team. For AI search they matter as sources: engines cite them for questions about lived experience, product opinions and troubleshooting, where a discussion thread may be the most direct evidence available.
The entry covers these platforms as sources that engines cite. Review sites have their own entry, and video platforms such as YouTube are treated separately. Presence here means being discussed accurately by real users; corporate posting without community standing is a different and riskier activity, described under brand mention seeding.
How it works
Two documented agreements give engines structured access to Reddit. Google announced on 2024-02-22 that it would receive real-time, structured access to Reddit content through Reddit's Data API, with the stated aim of showing more Reddit information in its products. Reddit and OpenAI announced on 2024-05-16 that OpenAI would use the same Data API to bring Reddit content into ChatGPT, especially on recent topics. In Pew Research Center's March 2025 panel, Wikipedia, YouTube and Reddit were the most frequently cited sources in Google AI summaries, together 15% of the sources listed; Pew did not report Reddit's share on its own.
Why it matters
For "is it worth it", "X vs Y" and "what do people think of" questions, a discussion thread can be the source an assistant summarizes, so customer sentiment on these platforms flows into answers about the brand. A team can monitor the threads engines cite for its category, fix the product problems those threads describe, answer questions openly from disclosed staff accounts where community rules allow, and make it easy for satisfied customers to share their experience. Buying or faking presence runs into platform spam rules and tends to backfire.
Example
A buyer asks an assistant whether a mid-range robot vacuum copes with pet hair. The answer draws on a two-year-old Reddit thread in which owners complain about tangled brushes. The manufacturer, which has since shipped a redesigned brush, has its support lead reply in the thread from a disclosed account, and the newer discussion becomes available to later answers.
ObservedNo general answer engine documents how it weighs review sites; Amazon and OpenAI document the use of review data in their own shopping surfaces, which is their scope only.
Reviews platforms (G2, Trustpilot, Capterra) are third-party sites where customers rate and review products and companies. AI answers to "best X" and comparison questions often draw on them, and shopping assistants use review content directly, so a brand's ratings and review text there feed into how it is described.
These platforms collect structured reviews (star ratings, pros and cons, use cases, company size) and organize them into category pages, comparisons and rankings. G2 and Capterra focus on business software; Trustpilot covers companies across sectors; marketplaces such as Amazon host reviews for their own catalogs.
The entry covers review sites as sources for AI answers. Discussion forums belong to UGC platforms, and a brand's own testimonials are outside the term. Google's review snippet guidelines draw a similar line for rich results: when an organization controls the reviews about itself, its pages using LocalBusiness or Organization markup are ineligible for star review features, and ratings aggregated from other websites are not allowed.
How it works
Amazon states that its shopping assistant, launched as Rufus and renamed Alexa for Shopping on 2026-05-13, was trained on Amazon's product catalog, customer reviews, community Q&As and information from across the web. OpenAI's commerce documentation lists reviews among the recommended product-feed attributes that improve ranking, relevance and user trust in ChatGPT shopping experiences. General answer engines publish nothing comparable; the use of review pages is visible in the citations they show for comparison prompts.
Why it matters
For shortlist questions ("best help desk software for a 20-person team"), an assistant needs comparable evidence about many products at once, and review platforms supply it. A thin or outdated profile leaves the assistant little to say about the brand, or only old complaints. Teams can claim and complete their profiles, keep category and feature data current, invite every customer to review at the same point in the journey, and respond to negative reviews. Trustpilot's business guidelines, for example, prohibit fake reviews, incentivized reviews and inviting only customers known to be happy.
Example
An HR software vendor sees that assistants answering "best onboarding software for small companies" cite a G2 category page where its product has 14 reviews, most from 2023. It starts asking every customer for a review 90 days after go-live and updates its feature listing; within two quarters the page shows a larger, recent set of reviews for assistants to draw on.
Also called: YouTube citations, video citations in AI answers, video in AI search
MeasuredObservedCitation frequency: one Pew Research Center study of Google AI summaries; how engines read video (transcripts, captions, metadata or the video itself) is not documented and likely varies.
YouTube as a cited source refers to videos linked or summarized in AI answers. In Pew's 2025 study of Google AI summaries, YouTube was one of the three most-cited sources, with Wikipedia and Reddit. Captions, transcripts, titles and descriptions give text-based retrieval something to read, though no engine documents how it processes video.
AI answers cite videos for how-to, review and demonstration questions, often as a link with a thumbnail or a timestamp. The entry is named for YouTube because the available data singles it out; the same logic applies to other video hosts.
Pew Research Center found that Wikipedia, YouTube and Reddit were the most frequently cited sources in both AI summaries and standard Google results in March 2025, together 15% of the sources in AI summaries. It also found YouTube links somewhat more common in standard results than in AI summaries. Video is prominent in AI answers, then, and in that study a little less prominent than in classic results. Engines do not say whether they read a video's transcript, its captions, its metadata or the video itself.
How it works
YouTube can create automatic captions with speech recognition in dozens of languages. Its help center warns that their quality varies, that they can misrepresent speech because of mispronunciations, accents, dialects or background noise, and that creators should review and edit them. Uploaded or corrected captions, accurate titles, descriptions with chapters, and a written companion page on the creator's own site all give text-based systems a faithful version of what the video says.
Why it matters
When a category's questions are best answered by a demonstration (setup, repair, side-by-side comparison), engines may cite a video, and a brand without one cedes that slot to creators and competitors. Because much retrieval runs on text, the words around and inside the video count: a product name mangled by auto-captions makes the video harder to match. Teams can publish focused videos for key how-to questions, correct their captions, write descriptive titles and chapter markers, and pair each video with a transcript page on their own domain.
Example
A camera maker's five-minute video on cleaning a sensor ships with auto-captions that turn the model name into a common word. The team uploads corrected captions, adds chapters ("remove lens", "blower", "swab") and posts the transcript on its support site, so both the video and the support page can match questions about cleaning that model's sensor.
Also called: topic authority, topical coverage, topical depth, subject-matter authority
ObservedAn SEO-industry principle. Google documents a news-specific "topic authority" system and query fan-out, which makes coverage mechanistically relevant, but no study isolates topical authority as a cause of AI citation.
Topical authority is the degree to which a site covers a subject completely, accurately and coherently enough to be treated as a go-to source on it. In AI search it matters because one question fans out into many sub-queries, and a site with strong pages across the whole topic gives more of those sub-queries a passage worth citing.
The concept comes from SEO practice: instead of ranking one page for one keyword, a site earns standing on a subject by covering its subtopics in depth, from basics to edge cases, with consistent quality and a clear internal structure. It is judged informally, by coverage and by how often others cite the site on the topic.
Google uses the phrase for a specific system. Its "topic authority" system, described on 2023-05-23, helps surface expert and local publications for news queries in specialized areas such as health, politics and finance, using signals such as a publication's notability for a topic or location, original reporting that other publishers cite, and its reputation. That system is scoped to news and should not be read as a general score for every site.
How it works
Google documents that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources. Each sub-search is a separate chance to retrieve a page, so a site that answers the main question and its neighbors (definitions, comparisons, costs, failure cases) has more entry points into the final answer. Martinez's 2026 survey of 45 GEO studies names topical relevance as one of the two most reproducible levers, which supports matching content closely to the questions people ask.
Why it matters
Fan-out turns one buyer question into a dozen narrower ones, and the engine picks a passage for each. A site with one strong page and nothing around it competes for one slot; a site that covers the topic end to end competes for several. Coverage also earns citations from readers and other publishers, who link to complete resources. Teams can map the sub-questions in their category from prompt research, support tickets and sales calls, find the gaps, and fill them with specific, well-linked pages. Coverage alone guarantees no citation.
Example
A heat-pump manufacturer has a strong product page and nothing on sizing, cold-climate performance, rebates or noise. It publishes a page on each, written with its installers and linked from a central guide. A question like "will a heat pump work at minus 20" now has a matching passage on its site, where before only competitors and forums had one.
Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026) (arXiv:2607.14035), published 2026-07-15, read 2026-11-07: https://arxiv.org/abs/2607.14035
Also called: topic clusters, content hubs, pillar pages, hub and spoke, pillar-cluster model
ObservedA site-architecture practice from SEO; Google documents why internal links matter for discovery and relevance, and does not document clusters as a ranking or citation factor.
Content clusters / hub structure is the practice of grouping pages on one subject around a hub, or pillar, page that covers the topic broadly and links to deeper pages on each subtopic, which link back. It is the site-architecture form of topical authority, helping crawlers find every page and helping engines match sub-questions to specific passages.
A cluster has three parts: a hub page that introduces the topic and orients the reader, spoke pages that each answer one subtopic in depth, and internal links in both directions with descriptive anchor text. Clusters can nest, with a spoke becoming the hub of a narrower cluster.
The term describes structure. Topical authority is the outcome the structure supports, and internal linking is the mechanism that makes it navigable. A cluster that repackages the same thin content across many pages adds nothing for readers and drifts toward the scaled-content patterns that search engines act against.
How it works
Google's link best practices say Google uses links as a signal when determining the relevance of pages and to find new pages to crawl, that every page you care about should have a link from at least one other page on your site, and that internal anchor text helps both people and Google understand the site. A cluster makes those links systematic. In AI search, the payoff comes through fan-out: when an engine issues sub-searches across subtopics, a cluster offers a dedicated, findable page for each.
Why it matters
Pages that are orphaned or buried deep are harder to crawl, refresh and retrieve, so they rarely surface for the sub-questions they answer. A cluster gives each subtopic a clear page and tells crawlers how the pages relate. Teams can audit a topic for orphaned or overlapping pages, merge duplicates, give the hub a short answer plus a map of subtopics, and link spokes to each other where readers would naturally go next. This is ordinary information architecture, and the evidence that it raises AI citations specifically is observational.
Example
A payments company's hub page, "Accepting card payments", summarizes the topic in 150 words and links to spokes on interchange fees, chargebacks, PCI compliance, payment terminals and online checkout. Each spoke links back to the hub and to two related spokes, and the chargebacks page later becomes the hub of its own cluster on dispute types.
Also called: AGENTS.md file, agents.md, README for agents, AGENT.md
DocumentedClaimedThe format and its Linux Foundation stewardship are documented; the adoption figure of 60,000+ projects comes from its proponents; no answer engine documents reading AGENTS.md.
AGENTS.md is an open Markdown file format that gives AI coding agents project-specific instructions, such as how to build, test and follow conventions, placed at a repository's root or in subfolders. Released by OpenAI in August 2025 and now stewarded by the Linux Foundation's Agentic AI Foundation, it serves agents working on code; no answer engine documents reading it.
The agents.md site describes the format as a "README for agents": a dedicated, predictable place for the build steps, test commands and code conventions that would clutter a README written for people. It is standard Markdown with no required fields. In a monorepo, each package can carry its own file, and agents read the nearest one in the directory tree, so the closest AGENTS.md to the edited file wins; explicit instructions from the user in chat override it.
The Linux Foundation announced the Agentic AI Foundation on 2025-12-09 with three founding projects: Anthropic's Model Context Protocol, Block's goose and OpenAI's AGENTS.md. AGENTS.md is a file for agents doing work inside a project. It differs from llms.txt, which describes a website's content for language models, and from robots.txt, which states crawl permissions.
How it works
A coding agent opening a repository looks for AGENTS.md files, reads the one closest to the files it is changing, and follows its instructions alongside the user's request. Tools named as compatible by the agents.md site and the Linux Foundation include OpenAI Codex, Google's Jules and Gemini CLI, Cursor, GitHub Copilot's coding agent, VS Code, Devin, Factory and Amp, among many others; some, such as Aider and Gemini CLI, need a configuration line to read it. The project's site claims use by more than 60,000 open-source projects.
Why it matters
AGENTS.md belongs to the agent-facing layer of software that AI systems increasingly work through. Its direct effect on being cited in AI answers is nil as far as any documentation shows, since no answer engine says it reads the file. It matters for teams whose product is software or documentation used by developers: coding agents that follow accurate project instructions produce working integrations, and that shapes whether developers recommend the tool. It also shows the field's move toward files written for machines. Adding one to a marketing site will not improve visibility.
Example
An open-source SDK adds an AGENTS.md at its repository root stating the package manager, the test command, the style rules and "never edit generated files in /gen". A second AGENTS.md in the docs folder explains how to build the documentation site. A developer who asks a coding agent to add a feature gets a change that follows the project's conventions and runs its tests.
Also called: Natural Language Web, NLWeb protocol, Microsoft NLWeb
DocumentedEmergingLaunch, protocol and reference code are documented by Microsoft; no adoption figures are published, and no major answer engine documents querying NLWeb endpoints.
NLWeb is an open project from Microsoft, launched on 2025-05-19, that lets a website answer natural-language questions about its own content through a simple protocol, returning results as Schema.org JSON. Every NLWeb instance also acts as a Model Context Protocol server, so one endpoint can serve people in a chat box and AI agents alike.
NLWeb is led by R.V. Guha, who created RSS, RDF and Schema.org and joined Microsoft as a corporate vice president and technical fellow. It has two parts: a protocol for asking a site questions in natural language, and an MIT-licensed reference implementation that indexes a site's existing semi-structured data (Schema.org markup, RSS feeds and similar lists of products, recipes, events or reviews) into a vector store and answers with a language model. Microsoft named early collaborators including Eventbrite, Shopify, Tripadvisor and O'Reilly Media, along with the data vendors Milvus, Qdrant and Snowflake.
NLWeb is about a site exposing its own content to queries. It differs from llms.txt, a static file describing content, and from MCP itself, the general protocol NLWeb uses to talk to agents. The project's README sums up the relation as NLWeb being to MCP and A2A "what HTML is to HTTP."
How it works
The reference server exposes two endpoints that take the same arguments: /ask for web clients and /mcp for MCP clients. A request carries the user's query, optionally a site filter, earlier queries (prev) for conversational context, and a mode: list returns the most relevant items, summarize adds a summary, and generate writes an answer in the style of retrieval-augmented generation. Results come back as JSON with each item's URL, name, score, description and the underlying Schema.org object; the server keeps no state between requests. Cloudflare announced a one-click NLWeb deployment on its AutoRAG service on 2025-08-28, which crawls and indexes a site and serves both endpoints.
Why it matters
NLWeb bets that sites will answer agents directly as well as being crawled and summarized by them. For a publisher, it makes Schema.org markup do double duty: the structured data that describes pages for search becomes the index an agent can query. Evidence of any visibility benefit is thin, since no major answer engine documents calling NLWeb endpoints and no adoption figures are published. A team with rich structured content can experiment through Cloudflare or the reference code, while keeping crawlable pages and markup as the primary path to AI answers.
Example
A recipe publisher with Schema.org Recipe markup on 40,000 pages deploys NLWeb. A visitor types "vegetarian dinner under 30 minutes without mushrooms" into the site's chat box and gets matching recipes; an AI agent connected over MCP sends the same question to the /mcp endpoint and receives the recipes as structured objects.
DocumentedGoogle's published spam policy, authoritative for Google Search and its AI features; other engines have not published an equivalent policy.
Scaled content abuse is Google's spam-policy term for generating many pages primarily to manipulate search rankings instead of helping users, whether the pages are written by AI, by people or by both. Google's spam policies now also cover attempts to manipulate generative AI responses in Search, so mass-produced pages are a risk in AI search too.
Google announced the policy on 2024-03-05, together with its March 2024 core update, expanding the older policy on automatically generated content to cover content produced at scale however it is made. The current policy page lists examples: using generative AI or similar tools to generate many pages without adding value for users; scraping feeds or search results and transforming them (synonymizing, translating, obfuscating) into many pages; stitching content from different pages without adding value; creating multiple sites to hide the scaled nature of the content; and creating many pages that make little sense to a reader but contain search keywords.
The policy judges purpose and value. Google's 2023 guidance on AI-generated content says that appropriate use of AI or automation is not against its guidelines, while using it primarily to manipulate rankings violates its spam policies.
How it works
Google says it detects violations through automated systems and, where needed, human review that can lead to a manual action. Its spam policies define spam as techniques used to deceive users or manipulate Search systems, and their introduction now names attempts to manipulate generative AI responses in Google Search. To appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Search with a snippet, so pages demoted or removed under the policy lose that path as well. Google's Search Status Dashboard lists recurring spam updates, including ones starting 2026-03-24, 2026-06-24, 2026-08-18 and 2026-09-24.
Why it matters
AI search created a new temptation: hundreds of templated pages written to be quoted by answer engines. Google treats that pattern as spam when the pages exist to manipulate rankings or AI responses, and its AI features draw only on pages eligible in Search. Other engines have published no equivalent policy, but near-duplicate pages give a retrieval system little reason to prefer any one of them. Before publishing at scale, a team should ask what each page adds that existing pages do not, and keep only the ones with a real answer.
Example
A marketing site generates 600 pages titled "What is [term] in AI search?" from one prompt template, each 300 words long and nearly identical apart from the term. That matches the policy's first example, many pages generated with AI tools without adding value for users, and exposes the whole set to demotion or removal from Search and, with it, from Google's AI features.
Also called: product feeds, shopping feeds, Merchant Center feed, product catalog data, agentic commerce feeds, ChatGPT product feed
DocumentedEmergingGoogle and OpenAI document feed-based eligibility for their own surfaces; agentic checkout protocols are new and changing, and other assistants' ingestion of product data is undocumented.
Product feeds and AI shopping visibility concern the structured catalog files (titles, prices, availability, images, identifiers) that merchants submit to platforms such as Google Merchant Center or OpenAI's commerce program. AI shopping surfaces use them to find, describe and recommend products and, through new agentic checkout protocols, to let assistants complete purchases. Stale feeds lead to wrong prices or exclusion.
In commerce, AI visibility depends more on structured product data than on prose. That data reaches AI shopping surfaces through three routes: feeds submitted to a platform (Google Merchant Center, or OpenAI's product feed under the Agentic Commerce Protocol), on-page Product structured data, and marketplace catalogs. Google's guidance on AI features lists checking that Merchant Center and Business Profile information is up to date among its best practices.
The term covers discovery and recommendation driven by structured product data, plus the checkout protocols built on top of it. Paid shopping placements belong to advertising in AI answers. A complete, current feed is necessary for eligibility; it does not decide which product an assistant recommends.
How it works
OpenAI describes the Agentic Commerce Protocol (ACP) as an open standard that lets ChatGPT ingest structured catalog data, understand merchant inventory and surface relevant products in context. Its feed specification accepts CSV, TSV or JSONL with nine required fields for discovery (item_id, title, description, url, brand, seller_name, image_url, availability, price), turns search eligibility on by default and checkout eligibility off, and asks merchants to update the feed when sales start or end and when stock changes. OpenAI's documentation still describes Instant Checkout in ChatGPT for approved partners, with the merchant as the merchant of record, although OpenAI said in March 2026 that checkout was moving to merchants' own ChatGPT apps and that it would focus on product discovery.
Google announced the Universal Commerce Protocol (UCP) on 2026-01-11, developed with Shopify, Etsy, Wayfair, Target and Walmart, to let shoppers buy directly from eligible businesses in AI Mode and the Gemini app. Merchants can bring their existing Merchant Center shopping feeds and remain merchant of record; access to the protocol runs through a waitlist. Amazon's shopping assistant, renamed from Rufus to Alexa for Shopping on 2026-05-13, draws on Amazon's own catalog and reviews.
Why it matters
Shopping assistants recommend from what they can parse and trust. A product missing from a feed, or listed with a stale price or availability, may be skipped or described wrongly, and the error repeats in every answer that uses the data. Teams can treat the feed as a primary marketing asset: submit complete feeds to Merchant Center and to OpenAI's program where eligible, keep price and availability in sync with the site, make on-page Product markup match visible content, and enrich the attributes buyers filter on (materials, sizes, compatibility). Checkout integrations are optional and still changing.
Example
A running-shoe retailer syncs its Merchant Center feed every hour and submits a JSONL feed to OpenAI's commerce program with sizes, widths, heel drop and cushioning as attributes. When a shopper asks an assistant for "wide-fit trail shoes under $150 in size 11", the retailer's in-stock models carry exactly the attributes the question filters on.
Also called: AI search ads, ads in AI Overviews, ads in AI Mode, ChatGPT ads, Copilot ads, sponsored placements in AI answers, conversational ads
DocumentedClaimedPlacement, eligibility and policies are documented on each operator's pages as of 2026-11-07 and change month to month; statements that ads do not influence answers, and pilot figures, are operator claims not independently verified.
Advertising in AI answers means paid placements shown inside, beside or below AI-generated responses, such as ads in Google's AI Overviews and AI Mode, sponsored units in ChatGPT and ads in Microsoft Copilot. They are bought through each platform's ad system and labeled as sponsored, separately from the organic citations an answer earns.
The term covers any paid unit placed in or around a generated answer: text and shopping ads within Google's AI features, an ad unit below a ChatGPT response, ads in Copilot answers, and sponsored prompts in shopping assistants. Its boundary is payment. Organic citations, the sources an engine chooses to reference, are earned through indexability and content and, according to the operators, are not for sale.
Google, OpenAI, Microsoft and Perplexity each say their ads are labeled and kept apart from the answer. OpenAI states that advertisers "have no ability to shape, rank, or alter ChatGPT's responses"; Microsoft says organic Copilot responses are never influenced by sponsored content; Perplexity said in 2024 that the content of its answers would not be influenced by advertisers. These are operator policies. Availability changes month to month, so each platform's live help pages are the reference.
How it works
Google (as of 2026-11-07). Search and Shopping ads can appear above or below AI Overviews in all 200+ countries and territories, and all languages, where AI Overviews are available. Ads within AI Overviews and AI Mode run in English, labeled "Sponsored", in the United States, Australia, Canada, India, Indonesia (AI Overviews only), Kenya, Malaysia, New Zealand, Nigeria, Pakistan (text ads only), the Philippines and Singapore. Google began testing AI Mode ads in May 2025, expanded them to the ten non-US markets on that list on 2026-10-28, launched multiturn ads in follow-up questions in the US, and plans Brazil, Mexico and the United Kingdom in Q4 2026. Existing Search, Shopping and Performance Max campaigns are eligible automatically, and AI Overviews and AI Mode cannot be targeted or excluded separately from Search.
OpenAI. Ad testing began in the US on 2026-02-09 for adults on the Free and Go plans; Plus, Pro, Business, Enterprise and Edu are ad-free, and Free users can choose an ads-free mode with lower limits. On 2026-10-21 OpenAI reported availability in 89 countries and more than 100,000 advertisers (its own figures). One ad unit can appear below a response, and a multi-image format is in testing during image generation. Ads are chosen from the current conversation's context and, with personalization on, past chats, memory and ad interactions; ranking weighs relevance and advertiser bids. Users can hide or report ads, see why one appeared, clear ad data and turn personalization off; personalized ads are not initially offered in the EEA or Switzerland. No ads go to accounts identified as under 18, into Temporary Chats or near sensitive topics such as health, mental health or politics, and political ads are not allowed. Advertisers receive aggregated performance data only.
Microsoft. A Microsoft Advertising post of 2025-03-05 said ads ran in Copilot in English, French and German markets, with Spanish and Japanese to follow, and introduced Showroom ads, an interactive format piloted with select advertisers.
Amazon. Sponsored Products and Sponsored Brands prompts, AI-generated prompts shown in shopping results and on product pages that can open a conversation in Amazon's assistant, reached general availability for US advertisers on 2026-03-25; eligible campaigns are enrolled automatically and clicks are billed on cost per click.
Perplexity. It began testing sponsored follow-up questions and paid media beside answers in the US in November 2024; trade press in February 2026, citing the Financial Times, reported that it is winding the program down by the end of 2026.
Why it matters
Paid and earned visibility now share the same answer surfaces but follow different rules: ads are bought by auction through each platform, citations are earned through content and indexability. Reporting has to separate the two, or a brand will mistake ad impressions for organic presence. Targeting differs as well: Google matches on queries and the context of AI responses, OpenAI on conversation context and, with personalization, chat history. Answer-independence statements are operator policies that the sources reviewed here do not verify, so monitor answers over time, and check each platform's live pages before planning spend.
Example
A US outdoor retailer running Shopping and Performance Max campaigns becomes eligible, with no new campaign, for ads inside English AI Overviews and AI Mode answers, and separately runs a ChatGPT campaign through OpenAI's Ads Manager to appear below relevant Free and Go conversations. Its reporting tracks both paid placements apart from organic citations of its gear guides; eligibility does not guarantee that an ad serves.
Also called: position in context, lost in the middle, positional bias, context placement
MeasuredPosition effects measured in controlled experiments (Liu et al., 2023) and named one of the two most reproducible GEO levers by one 2026 survey of 45 studies; behavior varies by model and engine.
Context position is where a retrieved passage sits within the text a language model reads before it answers. Research shows models use information at the start or end of a long context more reliably than information in the middle. Publishers cannot choose that slot, but self-contained passages remain usable wherever retrieval places them.
When an answer engine retrieves sources, it assembles them, with the user's question and its instructions, into the model's context window. The engine's retrieval and ranking set the order, and the order affects what the model uses. Liu and colleagues ("Lost in the Middle", arXiv:2307.03172, published in Transactions of the Association for Computational Linguistics) tested multi-document question answering and key-value retrieval and found performance was often highest when the relevant information sat at the beginning or end of the context and dropped significantly when it sat in the middle, including for models built for long contexts.
Martinez's critical survey of 45 GEO studies (arXiv:2607.14035, 2026-07-15) concludes that topical relevance and context position are the most reproducible levers. Context position is decided by the engine; the related publisher-side concepts are chunking and answer capsules.
How it works
Retrieval decides whether a passage enters the context at all; ranking and assembly decide where it lands. A passage ranked first may sit at the top of the context, while the fifth of ten may land in the middle, where the Liu et al. results suggest it is less likely to be used. Engines can also reorder, truncate or summarize sources, and none documents its assembly order.
Why it matters
Being retrieved is necessary and not sufficient: a passage placed mid-context may be read and still not used. The publisher controls the passage and not its slot, so the practical response is to make each important passage self-contained, with the subject, the claim and its key figure in the same few sentences, so it stays usable wherever it lands. Martinez also finds that citation-oriented rewrites can impair retrieval, so making a passage more quotable must not weaken its match to the query. Per-engine behavior varies, and none of this guarantees use.
Example
In the Liu et al. experiments, the same relevant document was moved between the first, middle and last positions among the documents given to the model, and accuracy fell when it sat in the middle. For a publisher, the lesson is modest: the first sentence of a page does not control where an engine places the passage, but a passage that restates its subject and answer can be understood without its neighbors, wherever it is placed.
Liu et al., Lost in the Middle: How Language Models Use Long Contexts (arXiv:2307.03172), published 2023-07-06, read 2026-11-07: https://arxiv.org/abs/2307.03172
Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026) (arXiv:2607.14035), published 2026-07-15, read 2026-11-07: https://arxiv.org/abs/2607.14035
Also called: AI seeding, LLM content seeding, seeding for AI citations
ClaimedEmergingPractitioner coinage popularized by Backlinko (owned by Semrush); the supporting figures are vendor-reported and no independent causal test has been published.
LLM seeding is the practice of placing a brand's information in the places and formats AI systems are likely to retrieve, summarize and cite, such as review sites, forums, industry publications and comparison articles, instead of relying only on the brand's own website. It is an off-site visibility tactic whose effects remain poorly measured.
LLM seeding treats the open web as the input set of AI answers and tries to make a brand present wherever those answers draw from. Backlinko defines it as publishing content in the formats and places LLMs are most likely to scrape, summarize and cite, and lists targets such as articles on Medium, Substack and LinkedIn, guest posts and expert quotes in trade publications, Reddit, Quora and niche forums, review sites like G2 and Capterra, and video platforms.
The term overlaps with digital PR for AI answers but is broader in format (a brand's own posts on third-party platforms count) and narrower in intent (the goal is retrieval and citation by AI systems rather than coverage for its own sake). It must be kept apart from brand mention seeding, the coordinated planting of mentions in forums and comments, which this dictionary treats as a risky, astroturf-adjacent practice. Legitimate LLM seeding publishes attributable, useful content where it is welcome.
Why it matters
AI answers about a category often cite review platforms, forums and roundups more than vendor sites, so a brand absent from those sources can be absent from the answer. Seeding asks a practical question: when engines answer our buyers' questions, which third-party pages do they read, and are we represented there accurately? Teams start from citation data for their key prompts, then earn reviews, contribute expert answers under real names, and pitch inclusion in comparisons engines already cite. Evidence is thin: most figures come from vendors, and one vendor study found about 11% of AI citations in software reviews came from roundups written by the brand ranked first.
Example
A project-management startup finds that ChatGPT and Perplexity answers to 'best tool for agencies' cite a G2 category page, two Reddit threads and a trade magazine roundup. It asks customers for G2 reviews, has its head of product answer agency questions on Reddit under her own name, and pitches the magazine's editor with usage data.
MeasuredObservedAI citation share from one vendor study (Peec: 232,000 citations, software niche, 12 weeks); the Google demotion was observed by practitioners in early 2026 and has not been confirmed by Google.
Self-promotional listicles are 'best X' roundup articles in which the publishing company ranks its own product first, often written to be cited when AI assistants answer comparison questions. They still earn AI citations, but practitioners linked them to Google visibility losses in early 2026, which makes them a contested tactic.
A self-promotional listicle is a comparison or 'best of' article on a vendor's own site that lists competitors alongside the vendor and places the vendor at the top. The format predates AI search, but it spread as a GEO tactic because assistants answering 'what is the best tool for...' questions often cite roundup articles.
Peec analyzed 232,000 citations from 13,000 listicles across six AI platforms over 12 weeks in the software-review niche and found that about 11% of citations came from self-promotional listicles: 3.6% on ChatGPT and about 10% on Google AI Mode and Perplexity, with no sign of demotion during the period. On the Google side, Lily Ray reported in February 2026 that several SaaS and B2B sites with hundreds of such articles lost between 29% and 49% of visibility from mid-January, while stressing that her sample was small, selected and not proof of cause. Google has not confirmed a targeted change.
Why it matters
A page that loses Google visibility can also lose its place in Google's AI features, even while other assistants keep citing it. The format also carries a credibility cost: a list that always crowns its author reads as advertising to buyers. Teams that publish comparisons can lower the risk by disclosing their affiliation plainly, using transparent criteria and evidence, giving competitors fair treatment, and earning inclusion in independent roundups instead of relying on their own. Evidence on both sides is limited to one niche and anecdotal site samples.
Example
A CRM vendor publishes 'The 12 best CRMs for 2026' with itself at number one. Perplexity cites the page for a buyer's comparison question, while the vendor's blog folder loses Google visibility the same quarter, a pattern the team cannot pin on one cause without testing.
Also called: AI Max, AI Max for Search campaigns, Google AI Max
DocumentedFeatures and AI-surface eligibility per Google Ads documentation; performance uplifts are Google's own internal figures.
AI Max for Search is a Google Ads setting, announced on 2025-05-06, that adds AI-driven keyword expansion, generated ad text and automatic landing-page selection to a Search campaign. It matters for AI search because it qualifies ads for contextual matching inside AI Overviews and AI Mode, which exact and phrase match keywords mostly cannot reach.
AI Max for Search is an add-on to standard Google Search campaigns. Google's help pages group it into three features: search term matching, which extends existing keywords through broad match and keywordless technology; text customization, which generates headlines and descriptions from the advertiser's ads, landing pages and assets (previously called automatically created assets); and final URL expansion, which can send a click to a different page on the site when it fits the query better. Brand controls, location-of-interest targeting and expanded search-term reporting come with it.
Google's documentation on ads in AI experiences states that AI Max for Search, broad match, Shopping and Performance Max are eligible for both contextual matching and direct query matching in AI Overviews and AI Mode, while exact and phrase match keywords are eligible only when a query shows direct commercial intent. Google reported that advertisers activating AI Max typically see 14% more conversions or conversion value at a similar cost, based on its internal 2025 data.
Why it matters
Paid visibility inside AI answers runs through Google's existing campaign types, and AI Max is the switch that lets a Search campaign compete for the long, exploratory questions typical of AI Mode. It also hands Google more control over which queries trigger ads, what the ads say and which page they open, so brand and compliance teams need to review generated text and set URL exclusions. Marketers should treat it as the paid counterpart to organic AI visibility, report it separately from earned citations, and judge it on their own conversion data rather than Google's average uplift.
Example
A B2B software company turns on AI Max for its expense-management campaign. Its ad starts appearing beside AI Mode answers to questions like 'how do I stop employees losing receipts', with a generated headline and a landing page Google picked, which the team then narrows with URL exclusions.
Also called: Direct Offers, AI Mode offers, Google Direct Offers pilot
DocumentedEmergingPilot per Google's announcements of 2026-01-11 and 2026-05-20; participation and results are not published in aggregate.
Direct Offers is a Google Ads pilot, announced on 2026-01-11, that lets advertisers attach exclusive deals such as discounts to their products so Google's AI can present them to shoppers in AI Mode. Since May 2026 offers can surface directly within AI Mode responses, and offer types are widening beyond discounts.
Direct Offers lets a retailer define promotions in its Google Ads campaign settings and leaves the decision of when to show them to Google's AI, which judges when a shopper looks ready to buy. Google introduced it on 2026-01-11 as a pilot focused on discounts, naming Petco, e.l.f. Cosmetics, Samsonite, Rugs USA and Shopify merchants as early participants, with bundles and free shipping planned.
On 2026-05-20 Google said Chewy, Gap and L'Oreal had surfaced deals through the pilot, that offers will now appear within AI Mode responses as shoppers compare options, that brands can upload giveaways and local coupons, and that Gemini can assemble product bundles for a specific search. Merchants integrated through the Universal Commerce Protocol can offer native checkout so a shopper claims the promotion faster. As an ads product, Direct Offers sits apart from organic product listings in AI Mode.
Why it matters
Direct Offers inserts a price incentive at the point of decision inside an AI shopping conversation, which changes how retailers compete in AI Mode: a well-described product and an attached discount now reach the same shopper. Teams need to coordinate promotions, feeds and campaign settings so offers are eligible, measure whether offers shift conversion or merely discount sales that would have happened anyway, and keep paid offers separate from organic visibility in reports. Access is limited to a pilot, so most advertisers cannot use it yet.
Example
A pet retailer in the pilot sets a 20% first-order discount on a premium dog food. When a shopper in AI Mode compares grain-free options and narrows the choice to two brands, the response shows the retailer's offer next to its listing.