Also called: GEO: Generative Engine Optimization, Aggarwal et al. 2023, Aggarwal et al. 2024, arXiv:2311.09735, KDD 2024 GEO paper, the Princeton GEO paper
MeasuredPeer-reviewed (KDD 2024); one study, run on a benchmark harness, not a measurement of current production engines.
The GEO paper (Aggarwal et al.) is the 2023 study, published at KDD 2024, that coined generative engine optimization, defined ways to measure a source's visibility inside AI-written answers, released the GEO-bench benchmark, and found that adding citations, quotations and statistics raised visibility by up to 40 percent in its test setting.
Full reference: Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, "GEO: Generative Engine Optimization", arXiv:2311.09735 (v1 2023-11-16, v3 2024-06-28); KDD '24 proceedings, Barcelona, pages 5 to 16, DOI 10.1145/3637528.3671900. Code is on GitHub under GEO-optim and the benchmark is on Hugging Face.
What it establishes: the vocabulary of generative engines and GEO; two visibility metrics for synthesized answers; a 10,000-query benchmark; and, within that benchmark, measured gains from specific rewrites that varied by domain, with a smaller replication on Perplexity.ai.
What it does not establish: how today's production engines behave. The experiments used one model, GPT-3.5-turbo, over a fixed set of five retrieved sources, so the rewritten source was already in the answer's context. The results say nothing about whether a page gets retrieved in the first place, and no effect size should be assumed to carry over to ChatGPT, Gemini or AI Overviews as they ship today.
How it works
Each GEO-bench query was paired with the cleaned text of its top five Google results. The engine (GPT-3.5-turbo, temperature 0.7, five samples per query) wrote a cited answer; the researchers then rewrote one source with one of nine methods and measured the change in that source's visibility. Visibility had two measures: Position-Adjusted Word Count, the words attributed to a source weighted toward earlier citations, and Subjective Impression, seven model-judged qualities such as relevance, influence and likelihood of a click. Quotation Addition, Statistics Addition and Cite Sources gained most, roughly 30 to 40 percent on the first metric and 15 to 30 percent on the second; Keyword Stuffing scored below the unmodified baseline. Gains concentrated on low-ranked sources: Cite Sources lifted the fifth-ranked source by 115.1 percent while cutting the first-ranked source by 30.3 percent. On a 200-query Perplexity.ai subset, Quotation Addition improved the position-weighted metric by 22 percent.
Why it matters
Nearly every GEO tactic list traces back to this paper, often with its conditions stripped away. Reading it directly lets a team separate what was measured (rewrites of a source already in a fixed context, in 2023 and 2024, on one model) from what is commonly claimed (that the same rewrites win citations in any engine today). Its durable lessons are modest and useful: specific, verifiable, well-sourced writing helped in its tests, keyword stuffing did not, and the effect differed by topic. Cite it as the field's founding study, never as a current manual.
Example
A vendor deck claims that statistics boost AI visibility by 40 percent. Checked against the paper, the figure comes from a 2023 benchmark that used GPT-3.5-turbo and pre-selected sources, and the paper itself reports that gains varied by domain.
Also called: GEO-Bench, GEO benchmark, geo-bench dataset
DocumentedThe benchmark as published with the GEO paper and its public dataset.
GEO-bench is the benchmark released with the GEO paper: 10,000 user queries drawn from nine sources and tagged across domains and intents, each paired with the text of up to five top Google results, built to measure how visible a source is inside answers written by a generative engine. It is a common reference point for later GEO experiments.
The paper splits GEO-bench into 8,000 training, 1,000 validation and 1,000 test queries (the Hugging Face viewer shows 9,000 train and 1,000 test rows). Queries come from MS MARCO, ORCAS-1, Natural Questions, AllSouls, LIMA, Davinci-Debate, Perplexity.ai Discover, ELI5 and GPT-4-generated questions. About 80 percent are informational, with transactional and navigational queries making up the rest. GPT-4 tagged each query on seven dimensions: difficulty, nature of query, genre, topic, sensitivity, user intent and answer type; the dataset card warns that the tags may be noisy.
The benchmark scores visibility with the paper's two metrics, Position-Adjusted Word Count and Subjective Impression. It measures how an engine uses and cites sources already supplied to it. It does not measure crawling, indexing or retrieval, so it cannot show whether a live engine would find a page.
Why it matters
When someone says a GEO tactic was tested, the first question is where: on GEO-bench, on a live engine, or nowhere. A GEO-bench result is real evidence about how a language model uses sources it was handed, which matters once a page is retrieved. It is weak evidence about the step that decides most visibility in production, retrieval itself. Teams can use the public dataset and code to rerun experiments on newer models, but should still test on their own buyer questions in the engines their buyers actually use.
Example
A startup reports a new rewriting method that beats the paper's best result on the GEO-bench test split. That shows the method helps a model cite a supplied source, and leaves open whether it changes which pages ChatGPT search retrieves.
Aggarwal et al., GEO: Generative Engine Optimization, full text v3 (benchmark section), published 2024-06-28, read 2026-11-07: https://arxiv.org/html/2311.09735v3
Also called: How Generative Engine Optimization (GEO) Rewrites the Rules of Search, How GEO Rewrites the Rules of Search, GEO over SEO, a16z GEO, Act II of search
ObservedInfluential analysis from an interested source; it measures nothing itself.
The a16z GEO essay is "How Generative Engine Optimization (GEO) Rewrites the Rules of Search", published by venture firm Andreessen Horowitz on May 28, 2025. It helped carry GEO from research into marketing, framing AI answers as "Act II of search" and arguing that reference rates now matter alongside click-through rates.
Written by Zach Cohen and Seema Amble, the essay argues that search is moving from ranked results pages to language-model interfaces, so visibility depends on being cited inside generated answers. Its central metric is the reference rate: how often a brand or its content is cited or used as a source in model answers. It names a new category of monitoring platforms (Profound, Goodie, Daydream) and incumbent responses (Ahrefs Brand Radar, Semrush's AI toolkit), and predicts GEO tools could grow into a performance-marketing channel.
What it establishes: a framing much of the industry adopted, and an early map of the tool category. What it does not establish: any measurement. Its usage figures are borrowed. "Queries are longer (23 words, on average, vs. 4)" links to Semrush data in which 23 words was ChatGPT prompts without search and 4.2 words was ChatGPT queries with search on, so both figures describe ChatGPT and neither describes Google. The "6 minutes" session figure carries no attribution. The page carries a general disclaimer that portfolio companies may be mentioned.
Why it matters
The essay is widely quoted as the source for the idea that AI answers are a new marketing channel, so its claims travel into board decks and vendor pitches. Reading it as what it is, an investment thesis from a venture firm, helps teams keep its framing (measure how often engines reference you) without importing its borrowed figures as facts. Its reference-rate idea is a sound starting question for an AI visibility program; the numbers in a business case should come from the team's own measurement or from primary studies.
Example
A board deck states that AI queries run 23 words against 4 on Google, citing a16z. Traced to Semrush's underlying data, both numbers describe ChatGPT, so the slide needs a different comparison.
Also called: Pew Research AI Overviews study, Google users are less likely to click on links when an AI summary appears in the results, Pew 2025 AI summary clicks study
MeasuredOne study: independent, with published methodology; US only, Google only, March 2025.
The Pew Research AI summaries study is a July 2025 analysis of the real browsing of 900 US adults, which found that Google users clicked a traditional result on 8 percent of visits when an AI summary appeared versus 15 percent without one, and clicked a link inside the summary on just 1 percent.
Reference: Athena Chapekis and Anna Lieb, "Google users are less likely to click on links when an AI summary appears in the results", Pew Research Center, 2025-07-22. Panel: 900 US adults in KnowledgePanel Digital who shared browsing data for March 2025, covering 68,879 unique Google searches, of which 12,593 produced an AI summary. Results pages were re-collected in April 2025 through a scraping service.
Key findings: 18 percent of searches showed an AI summary. Users ended their browsing session on 26 percent of pages with a summary versus 16 percent without. Wikipedia, YouTube and Reddit made up 15 percent of sources cited in summaries (17 percent in standard results); government sites were 6 percent of summary sources versus 2 percent of standard ones. Most summaries (88 percent) cited three or more sources, and the median summary ran 67 words. Summaries appeared on 60 percent of queries starting with a question word and 53 percent of queries of ten or more words, against 8 percent of one- or two-word queries.
Limits: one month, US only, Google only, a small opt-in panel, and an association rather than proof of cause.
Why it matters
Pew's study is one of the few independent measurements based on observed browsing rather than modeled click-through rates, and it shows Google's AI summaries going with fewer clicks to the open web. For a publisher or brand, a summary can answer the query without sending the visit, and being cited inside the summary rarely produces traffic on its own. Teams should treat citation in AI summaries as a visibility and trust outcome, measure it separately from traffic, and pay attention to the longer, question-shaped queries that most often trigger summaries.
Example
A health publisher sees impressions hold steady but clicks fall on question-style queries. Pew's finding that summaries appear on 60 percent of question-word queries is consistent with those pages, more than its brand-name pages, losing traffic.
Also called: AI Overviews Reduce Clicks by 34.5%, Update: AI Overviews Reduce Clicks by 58%, Ahrefs AI Overviews CTR study, Ahrefs 34.5% study, Ahrefs 58% study
MeasuredOne vendor's study, re-run once on the same method; methodology disclosed; correlational by its own account.
The Ahrefs AI Overviews click studies are two vendor analyses of 300,000 keywords. The April 2025 study linked AI Overviews to a 34.5 percent lower click-through rate for the top-ranking page; the February 2026 re-run, over two years of data, put the gap at 58 percent. Both are correlational, per Ahrefs' own framing.
Both studies are by Ryan Law with Xibeijia Guan. Method: 150,000 keywords that showed an AI Overview and 150,000 informational keywords that did not, drawn from Ahrefs' Keywords Explorer, with average monthly desktop click-through rate from aggregated Google Search Console data. Informational keywords were the comparison group because, per Louise Linehan's separate October 2024 Ahrefs analysis, 99.2 percent of keywords that trigger AI Overviews are informational.
"AI Overviews Reduce Clicks by 34.5%" (2025-04-17) compared March 2024 with March 2025: position-1 CTR for AI Overview keywords fell from 0.073 to 0.026, against a forecast of 0.040 had they followed the comparison group, which fell from 0.056 to 0.031. "Update: AI Overviews Reduce Clicks by 58%" (2026-02-04) compared December 2023 with December 2025: position-1 CTR for AI Overview keywords fell from 0.073 to 0.016, against a forecast of 0.037.
Limits: one vendor's data, desktop only, a forecast-based counterfactual, keywords grouped (in 2026) by whether they showed an Overview at the end of the period, and Search Console's inability to separate AI Overview clicks.
Why it matters
The two studies supply widely repeated numbers on what AI Overviews do to organic clicks, and Ahrefs reads the move from 34.5 to 58 percent as the effect worsening as Overviews spread. The practical point for teams is that ranking first for an informational query that triggers an Overview brings far fewer visits than it did in 2023. Traffic forecasts built on historical click-through curves should be revised, and citation inside the Overview itself should be tracked separately from rank.
Example
A software company's top-ranked how-to guide earned about 7 percent of clicks in late 2023. Using Ahrefs' 2026 figures as a guide, it plans for under 2 percent while an Overview appears, and adds citations inside the Overview to its reporting.
Also called: SparkToro/Datos zero-click studies, SparkToro zero-click study, 2024 Zero-Click Search Study, zero-click search study, Rand Fishkin zero-click study
MeasuredPanel studies by one author, with different panels across editions; caveats published with each.
The SparkToro zero-click studies are clickstream analyses by Rand Fishkin measuring how many Google searches end without a click. The 2024 edition, on Datos panel data, found 360 open-web clicks per 1,000 US searches and 374 in the EU; the 2026 edition, on Similarweb data, found 68.01 percent of US searches ended with no click.
The 2024 study ("2024 Zero-Click Search Study", 2024-07-02) used the Datos clickstream panel from September 2022 to May 2024, with mobile data for January to May 2024 only. It found that 58.5 percent of US and 59.7 percent of EU Google searches ended in zero clicks, and that almost 30 percent of US clicks went to Google-owned properties. The page's URL slug swaps the US and EU figures; its title and body give 360 for the US and 374 for the EU. Earlier editions in 2019 (Jumpshot panel) and 2020 (Similarweb) set the trend line.
The 2026 study ("In 2026, Less than One Third of Google Searches Still Send a Click", 2026-06-09) used Similarweb's US desktop and mobile panel for January to April 2026. It reports 68.01 percent zero-click searches, gives 60.45 percent for 2024 on its own basis, and attributes most of the rise to AI Overviews without isolating that effect. Fishkin calls comparisons across editions "apples and oranges" because panels, users and devices differ.
Why it matters
These studies set the baseline the AI era builds on: well before AI Overviews, most Google searches already ended without an open-web click. That matters for attribution. A traffic decline cannot be blamed on AI answers alone without checking the longer zero-click trend, and stable rankings with falling visits reflect a shift that predates generative AI and that AI answers appear to have accelerated. Plans should value presence on the results page and inside answers, since many searches never produce a visit.
Example
A retailer's analytics show organic sessions down 18 percent year over year with steady rankings. Read against SparkToro's zero-click trend and Ahrefs' AI Overview studies, the team attributes part of the drop to queries being resolved on Google's own page.
Also called: AI features and your website, Optimizing your website for generative AI features on Google Search, Google AI optimization guide, Google AI Overviews documentation
DocumentedAuthoritative about Google Search only, and silent on how sources are weighted for citation.
Google Search Central AI features documentation is Google's official guidance for site owners on AI Overviews and AI Mode: how pages become eligible, how query fan-out works, which controls limit inclusion, and how to measure it. Its 2026 guide states that optimizing for these features is still SEO and that Google Search ignores llms.txt.
"AI features and your website" says AI Overviews and AI Mode may use query fan-out, issuing related searches across subtopics; that there are "no additional requirements" beyond being indexed and eligible for a snippet; that preview controls (nosnippet, data-nosnippet, max-snippet, noindex) limit what is shown; and that AI feature traffic counts under the Web search type in Search Console's Performance report.
"Optimizing your website for generative AI features on Google Search" (2026-05-15) adds that the features are "rooted in our core Search ranking and quality systems", names AEO and GEO, says content need not be chunked, discourages chasing inauthentic mentions, and covers agents. A 2026-06-15 update states that Google Search ignores llms.txt. The crawler reference explains that Google-Extended governs Gemini training and grounding outside Search and does not affect Search inclusion or ranking.
In 2026 Search Console added a Generative AI performance report (announced 2026-06-03, impressions only) and a site-wide Search generative AI control, both available worldwide by 2026-08-31.
How it works
The Search generative AI control sits in Search Console under Settings. Including a site is the default. Opting out removes the site's links and content from AI Overviews, AI Mode and generative AI features in Discover, including their use to ground answers, and the site then gets no traffic or impressions from those features. Google says the setting is not used as a ranking or inclusion signal elsewhere in Search, and that it does not limit AI training, which remains governed by Google-Extended.
Why it matters
These pages are the first-party statement of how Google's AI surfaces treat websites, so they settle several debates for Google specifically: special AI files and markup are unnecessary, classic indexing and snippet eligibility are the gate, and a site can now leave AI features without leaving Search. Teams should audit indexability and preview controls first, read the Generative AI performance report next to the Web numbers, and decide deliberately about both the new control and Google-Extended. None of this documents how other engines behave.
Example
A publisher wants out of AI Overviews but not out of Search. Instead of restricting snippets page by page, it switches off the Search generative AI control in Search Console and keeps Google-Extended disallowed to limit Gemini training.
Google Search Central, Latest Google Search documentation updates (2026-06-15 llms.txt note; 2026-10-21 control page update), published n/d, read 2026-11-07: https://developers.google.com/search/updates
Also called: Overview of OpenAI Crawlers, OpenAI bots documentation, GPTBot documentation
Documented
OpenAI crawler documentation is the official page "Overview of OpenAI Crawlers", which describes OpenAI's four web agents, GPTBot, OAI-SearchBot, ChatGPT-User and OAI-AdsBot: what each one does, how robots.txt applies to each, where their IP ranges are published, and how long robots.txt changes take to apply.
The page assigns one job to each agent. OAI-SearchBot surfaces sites in ChatGPT's search features; a site that disallows it will not appear in ChatGPT search answers, though it may still show as a navigational link. GPTBot crawls content that may be used to train OpenAI's foundation models; disallowing it signals that the content should not be used for training. ChatGPT-User visits pages when a user asks ChatGPT or a custom GPT something, and also serves GPT Actions; because those visits are user-initiated, the page says "robots.txt rules may not apply". OAI-AdsBot checks the safety of landing pages submitted as ChatGPT ads and visits only those pages; its data is not used to train foundation models.
Each robots.txt setting is independent of the others, so a site can allow search while refusing training; if both are allowed, OpenAI may use one crawl for both purposes. IP ranges are published as one JSON file per agent (for example openai.com/searchbot.json), and search-related robots.txt changes take about 24 hours to apply. The page carries no date.
Why it matters
This page decides whether a site can be found in ChatGPT search at all. Blocking GPTBot to refuse training is a common, defensible choice; blocking OAI-SearchBot by accident, often through a broad bot rule or a firewall, quietly removes a site from ChatGPT's search answers. Teams should name each agent explicitly in robots.txt, check firewall rules against the published IP lists, look for each user agent in server logs, and wait a day before testing a change.
Example
A SaaS company's robots.txt disallows GPTBot and allows OAI-SearchBot. Its pages keep appearing as cited sources in ChatGPT search, while the company has signaled that its content should not be used for training.
Also called: Perplexity Crawlers, PerplexityBot documentation, Perplexity bots documentation
Documented
Perplexity crawler documentation is the official "Perplexity Crawlers" page, which describes PerplexityBot, the crawler that makes sites appear and get linked in Perplexity's answers, and Perplexity-User, the fetcher that visits pages when a user asks a question, along with robots.txt behavior, published IP lists and firewall allowlisting steps.
PerplexityBot exists to make websites discoverable and linked in Perplexity search results, and the page states that it is not used to crawl content for AI foundation models. Perplexity-User fetches a page when a user's question calls for it, so Perplexity can answer and link to it; it is not used for crawling or for training, and because a user requested the fetch, it "generally ignores robots.txt rules". Each setting works independently, and robots.txt changes may take up to 24 hours to apply.
The page also gives unusually concrete firewall guidance: match each bot by both its user-agent string and its published IP ranges (perplexity.ai/perplexitybot.json and perplexity.ai/perplexity-user.json), with step-by-step rules for Cloudflare and AWS WAF, a warning that Cloudflare's Bot Fight Mode can still challenge the bots, and advice to refresh the IP lists automatically. The page carries no date.
Why it matters
Perplexity shows its sources prominently, so being crawlable by PerplexityBot is a precondition for appearing in its answers. Default bot protection at the CDN or firewall can block it without anyone deciding to. Teams should allow PerplexityBot explicitly, implement the user-agent plus IP checks the page describes, and remember that Perplexity-User may still fetch a page a user points it to despite a robots.txt disallow, so content meant to stay private needs real access control.
Example
A news site behind Cloudflare finds no PerplexityBot requests in its logs. Bot Fight Mode was challenging the crawler; the team adds a skip rule matching Perplexity's user agent and published IPs, and the bot starts fetching pages.
Also called: Does Anthropic crawl data from the web, and how can site owners block the crawler?, ClaudeBot documentation, Claude crawler documentation, Anthropic bots documentation
Documented
Anthropic crawler documentation is the Claude Help Center article on Anthropic's three web agents: ClaudeBot, which collects training data; Claude-User, which fetches pages for user questions; and Claude-SearchBot, which indexes content for search quality. It states what blocking each one changes, confirms the bots honor Crawl-delay, and notes a training-preference signal recognized since October 2026.
The article moved from support.anthropic.com to support.claude.com, and the old address redirects. It describes ClaudeBot as collecting web content that may be used for model training; blocking it signals that a site's future material should be excluded from training data. Claude-User fetches sites when a Claude user's question calls for it; blocking it stops those retrievals and may reduce visibility in user-directed search. Claude-SearchBot navigates the web to improve search result quality; blocking it stops indexing for search and may reduce visibility and accuracy in search results.
Anthropic says its bots honor robots.txt directives, respect the non-standard Crawl-delay extension where appropriate, and do not try to bypass CAPTCHAs. It publishes crawler IP addresses at claude.com/crawling/bots.json but recommends robots.txt over IP blocking. As of 2026-10-30 it also recognizes a machine-readable "no AI training" signal that follows a draft built on the IETF aipref working group's work; a robots.txt rule naming ClaudeBot takes precedence over that general signal. Unlike OpenAI's and Perplexity's pages, this one does not carve its user-triggered fetcher out of robots.txt. The page shows no publication date.
Why it matters
The three-way split lets a site separate training from search more finely than most vendors allow: refuse ClaudeBot and still be fetched and indexed for Claude's search answers. Blocking all three, a common side effect of blanket AI-bot rules, removes a site from Claude's answers as well as from its training data. Teams should decide agent by agent, write an explicit robots.txt group for each, note that a ClaudeBot rule overrides the new general training signal, and confirm in logs that requests come from the published IPs.
Example
A publisher adds User-agent: ClaudeBot with Disallow: / and leaves Claude-SearchBot and Claude-User allowed. Its articles can still be retrieved and cited when Claude answers with web search, while training crawls are refused.
Also called: /llms.txt, llms.txt specification, llmstxt.org, The /llms.txt file
DocumentedEmergingThe proposal and its adoption claims are documented by its author; use by answer engines at answer time is not established.
The llms.txt proposal is Jeremy Howard's specification, first published at llmstxt.org on September 3, 2024 and revised as v2 in August 2026, for a Markdown file at a site's root that gives language models a curated summary and links to clean Markdown versions of key pages, because HTML built for people is hard for models to use.
Reference: Jeremy Howard, "The /llms.txt file, v2", llmstxt.org, published 2024-09-03 and modified 2026-08-10, with a separate page listing changes since v1. The format: an H1 with the site or project name (the only required part), a blockquote summary, optional detail, H2 sections listing Markdown links with short notes, and an "Optional" section agents can skip. Pages may offer Markdown twins at the same URL with .md appended, discoverable through rel="alternate" links or HTTP Link headers.
What it establishes: a clear statement of the problem ("web pages are built for people", and context windows are still too small for whole sites) and, per the proposal's own page, adoption: thousands of sites publish the file, tools such as Mintlify, GitBook, Wix, Yoast SEO and AIOSEO generate it, OpenAI, Anthropic and Google's Gemini publish one for their developer docs, and Chrome's Lighthouse audits for it. These claims are the proposal's own.
What it does not establish: that answer engines read the file when composing answers. The page speaks of agents using it on demand and names no engine, and Google states that Google Search ignores it.
Why it matters
llms.txt is cheap to publish and useful where a known agent or developer tool fetches it, such as a coding assistant reading documentation. It is weak as an AI search lever: no answer engine has publicly documented using it to choose sources, and Google says its Search ignores the file. Teams should treat it as documentation hygiene for agents, keep it accurate if they publish one, and put their main effort into crawlable, well-structured HTML, which is what the engines' documented crawlers read.
Example
A developer tools company publishes /llms.txt listing its quickstart, API reference and pricing pages with Markdown twins. A coding assistant pointed at the docs loads the file and answers integration questions from the clean Markdown instead of parsing the HTML.
Also called: GEO glossaries, AI search glossaries, AEO glossaries
ObservedSite-stated counts and counts of visible entries as of 2026-11-07; these pages change often.
The glossary race is the wave of GEO and AI search glossaries published by vendors, agencies and independent authors as the field grew, ranging from lists of about 45 terms to catalogs claiming more than 300. They differ widely in size, in whether their stated counts match the entries a reader can see, and in whether those entries carry dates.
Counts as checked on 2026-11-07, giving each site's stated figure and the entries visible in the fetched page (pages can load partly, so visible counts may understate). Promptwatch's AI Search & GEO Glossary states 324 terms, with entries dated October 15 and 29, 2026. iPullRank's glossary in The AI Search Manual lists 246 terms. Digital Applied's GEO glossary (2026-04-30) claims 150 terms, with 73 entries visible in a fetch that looked complete. aisearchglossary.com states 94 terms (its FAQ says 60 or more), was updated 2026-11-06, and tracks whether each term is cited by five AI engines. Rankscale states 90 terms; LumenGEO states and shows 88; geodocs.dev claims more than 80, with 45 visible in an incomplete fetch; Reboot Online's GEO Playbook glossary shows 62; AEO.dev shows 46.
What the check showed: stated counts and visible entries often diverge, so a headline number is a weak guide to coverage. Dating is uneven: Promptwatch dates each term and aisearchglossary.com shows a site update date, while LumenGEO, iPullRank, Reboot Online and AEO.dev showed no publication date. Distinctive features vary too, from per-engine citation tracking at aisearchglossary.com to short definitions tagged with course modules at Rankscale.
Why it matters
Glossaries are among the pages AI engines cite when someone asks what a term means; aisearchglossary.com reports that 76 of its terms have been cited by at least one engine it tracks. Whoever defines the field's vocabulary therefore shapes how assistants explain it. For readers, the differences matter more than the counts: a definition with a source and a date can be checked, while an undated one-liner cannot. Anyone citing a glossary definition, this one included, should check its sources, its date and whether vendor terms are labeled as claims.
Example
Before adopting a new metric, an analyst compares how four glossaries define it, finds three different formulas and no cited source in two of them, and uses the definition that links to the vendor's own documentation, labeled as that vendor's term.
Also called: Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), arXiv:2607.14035, Martinez 2026, critical survey of GEO
ObservedA single-author preprint that synthesizes 45 studies; it runs no measurement of its own and is not peer-reviewed.
The critical GEO survey (Martinez 2026) is a July 2026 arXiv preprint by Olivier Martinez that reviews 45 GEO studies from November 2023 to July 2026 and concludes the evidence is narrow: content changes can alter how already-retrieved sources are cited, but no reviewed technique shows a stable, cross-platform effect on discoverability.
Reference: Olivier Martinez, "Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)", arXiv:2607.14035, v1 2026-07-15; 18 pages, filed under Information Retrieval and Digital Libraries, with a literature matrix and search protocol as ancillary files.
What it establishes, within its corpus: GEO is a "stochastic, partially observable pipeline" whose stages run from search activation, crawling and indexing through retrieval, reranking and context allocation, citation, prominence, factual absorption and fidelity to user behavior. The founding paper's gains hold in its setting, where the source is already in a fixed context, and show neither organic discoverability nor lasting traffic effects. Topical relevance and context position are the most reproducible levers; competition can erode individual gains; citation-oriented rewrites can impair retrieval; commercial audits report low source overlap, substantial run-to-run variability and persistent fidelity gaps.
What it does not establish: pooled effect sizes. It is a single-author survey, not a peer-reviewed meta-analysis, and its "most reproducible" findings are bounded by the 45 studies it reviewed.
Why it matters
The survey is a useful corrective to tactic lists that quote the 2023 GEO paper as if it described today's engines. It gives teams a pipeline to diagnose against: a missing citation can fail at crawling, retrieval, context allocation or citation, and each stage calls for different work. Its findings argue for relevance-first content, measurement that samples answers repeatedly, skepticism toward any vendor promising durable citation gains, and caution with citation-oriented rewrites that could weaken how a page is retrieved.
Example
A team whose rewritten pages gained citations in a lab test but not in live engines uses the survey's pipeline to look upstream, and finds the pages were rarely retrieved for the target questions at all.