Also called: AI visibility index, AI search visibility score, LLM visibility score, AI Visibility Score (Semrush)
ObservedClaimedEach vendor's formula is its own claim; the idea underneath, presence measured across a fixed panel of prompts, is observed practice.
An AI visibility score is a vendor-defined number that summarizes how present a brand is in AI-generated answers across a set of tracked prompts and engines. Each tool computes it differently, often blending appearance frequency with prominence, sentiment or competitor share, so a score is only readable alongside that tool's formula.
An AI visibility score compresses many sampled AI answers into one figure, usually on a 0 to 100 scale. There is no shared standard. The simplest defensible version is a presence rate: the share of sampled answers, over a declared panel of prompts, engines, locales and dates, in which the brand appears at least once. A measured study of answer variability (SparkToro, January 2026) concluded that a visibility percentage across dozens to hundreds of prompts, each run many times, is a reasonable metric.
Commercial scores usually go further. Semrush documents its AI Visibility Score as a 0 to 100 benchmark combining Topic Coverage (how many topics include the brand in AI answers, compared with other domains) and Mention Consistency (how often the brand is mentioned within those topics), computed over its own prompt database; the weighting is not published. The term covers such composite dashboard numbers. The panel-level primitives that feed them (citation share, brand mention rate, citation prominence) are defined separately so they can be compared across tools.
How it works
A tool runs a set of prompts on one or more engines, often several times per prompt, parses each answer for brand names and cited URLs, and aggregates the hits. Every step moves the number: which prompts are in the panel, how many runs per prompt, whether a mention without a link counts, whether position or sentiment adds weight, and whether the result is normalized against competitors. Two tools can watch the same brand in the same week and report very different scores.
Why it matters
Visibility scores are now the headline number in many AI search reports, so they shape budgets and verdicts on what worked. Read alone, they mislead: a score of 62 can mean the brand appeared in 62 percent of sampled answers, or a weighted blend with no percentage meaning at all. A team that relies on one should write down the vendor's formula, the panel and the sampling count, keep them fixed over time, and track a plain presence rate beside the composite so that any change can be traced to a cause.
Example
A dashboard reports "AI visibility: 62". Under a plain presence formula, that means the brand appeared in 62 of 100 sampled answers. Under a Semrush-style composite, it is a relative benchmark that also reflects topic breadth against competitors, so it cannot be compared with another tool's 62.
Also called: AI share of voice, share of AI answers, domain citation share
ObservedClaimedObserved as an industry metric; each vendor's exact formula is its own claim.
Citation share is the proportion of sampled AI answers, over a fixed set of prompts and engines, that cite a given domain at least once. It is the AI search counterpart of share of voice, and it is only comparable between reports that hold the prompt panel, engines, locale, sampling count and dates constant.
Computed precisely, citation share = (sampled answers that cite the domain one or more times) ÷ N, where N is every sampled answer in the panel. The indicator is binary per answer: three links to the same domain in one answer count once. N includes every repetition of every prompt, and the report should declare how errors, refusals and answers without sources are handled (kept in N, or excluded and counted). Counting raw citations instead of answers lets the ratio exceed 100 percent and rewards answers that repeat one source.
A related but different number is competitive share of voice: the domain's citations or mentions as a share of all those earned by a defined competitor set. Semrush, for example, describes its Share of Voice as the share of mentions a brand gets in AI answers measured against competitors, and a16z describes monitoring dashboards that report competitive share of voice. That figure answers how a brand compares; citation share answers how often a domain serves as a source. Neither is a universal standard.
Why it matters
Citation share is the most direct measure of whether engines use a site as evidence, and it is the number most often promised in AI search contracts. It only means something when five parameters are locked: the prompt set, the engines, the locale, the number of runs per prompt and the date range. Two tools both reporting 12 percent may be measuring different things. Teams should get the formula and the denominator in writing, and keep the panel fixed long enough to read a trend.
Example
A 50-prompt panel runs on 3 engines, once per prompt, giving 150 answers. The domain is cited in 30 of them, however many links each answer contains, so citation share is 30 ÷ 150 = 20 percent. Run with 5 samples per prompt, the denominator becomes 750 and the figure gains a usable error margin.
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Zach Cohen and Seema Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: page citation rate, per-query citation rate, citation frequency
ObservedCommon usage with no fixed denominator; the denominator is the known pitfall.
Citation rate is how often a specific page or domain is cited when a given question is asked, usually measured on one engine over repeated runs. It is narrower than citation share, and its meaning depends entirely on the denominator: per page or per domain, per query or per panel, per run or per day.
Citation rate = (runs in which the target page or domain is cited) ÷ (runs of the query or query group on that engine). The usual scope is one engine and one query, or a small group of queries on one topic, sampled many times so the rate stabilizes. Usage varies: some tools report it per page, others per domain, and some count citations per query without normalizing for how many source slots each answer offered.
First-party counts now exist for one engine family. Bing Webmaster Tools' AI Performance report, in public preview since 2026-02-10, shows Total Citations, Average Cited Pages and page-level citation activity across Microsoft Copilot, AI-generated summaries in Bing and some partner integrations. Microsoft states that these figures do not indicate ranking or placement. They are raw counts without a published denominator, so they cannot be turned into a rate without knowing how often the relevant questions were asked.
Why it matters
Citation rate is the working metric for one important page: the pricing page, the flagship guide, the comparison page. It tells a team whether changes to that page are followed by more or fewer citations on the engine that matters most for it. Because one run proves little, the rate needs dozens of runs per query, and it should always be reported with its denominator, engine and time window. Bing's citation counts add a free first-party cross-check for Copilot.
Example
A pricing page is cited in 12 of 50 runs of one query on one engine in one week: a 24 percent per-page citation rate for that query, on that engine, that week. The same page can show a very different rate on another engine or on a broader group of queries.
Also called: citation position, citation placement, position-adjusted word count, PAWC
MeasuredClaimedThe GEO paper defines and measures a text-visibility version (one paper); vendor prominence scores are claims.
Citation prominence describes where and how visibly a source appears inside an AI answer: an inline link early in the text, a footnote at the end, or a collapsed sources panel. The GEO paper formalized one version, position-adjusted word count, which measures how much of the answer's text is attributed to a source, weighted toward the top.
Prominence separates two answers that both cite a domain but give it very different exposure. Aggarwal et al. (the GEO paper, KDD 2024) defined two impression metrics. Word Count is the words of the sentences citing a source divided by all words in the answer, with a sentence's words shared equally among its citations. Position-Adjusted Word Count (PAWC) applies the same ratio with each sentence weighted by an exponential decay on its position, on the stated rationale that earlier sentences are more likely to be read. Impressions are normalized so that all citations in one answer sum to 1. The paper also scored a Subjective Impression with an LLM judge on seven dimensions, including the likelihood of a click.
These are measures of text visibility. The paper does not measure clicks, so any link from prominence to traffic is a plausible mechanism, not one of its findings. Vendors report their own versions, such as Semrush's Average Position of a domain's citation within answers.
How it works
For a source c in answer r, the paper writes PAWC as the sum, over sentences s that cite c, of |s| × e^(−pos(s)/|S|), divided by the sum of |s| over all sentences of r, where |s| is a sentence's word count and the exponent discounts later sentences. A tool applying it parses each sampled answer into sentences, attributes each sentence to its citations, and averages the source's share across runs.
Why it matters
A single cited checkbox hides most of what matters for traffic and persuasion. Being the first inline source in the opening paragraph and being footnote nine in a collapsed panel both count as a citation, yet they offer very different chances of being read or clicked. Teams comparing visibility over time should log where each citation appears as well as whether it appears. Microsoft states that Bing's citation counts do not indicate placement, so prominence still has to be measured from sampled answers.
Example
Two answers each cite a domain once. In the first, the cited sentence opens the answer and holds 30 of its 120 words; in the second, it is the last of eight sentences. Both count as one citation, but the first earns a far larger position-adjusted word count.
Also called: mention rate, brand mention frequency, AI brand mentions
Observed
Brand mention rate is the share of sampled AI answers that name a brand, whether or not they link to it. It captures awareness that citation counts miss, but an uncited mention carries no click path, and the answer alone cannot show whether the name came from training data or from retrieved pages.
Brand mention rate = (sampled answers that name the brand at least once) ÷ (all sampled answers in the panel). Like citation share, it is binary per answer and needs a fixed panel. Detection is harder than it looks: a tool must match product names, sub-brands, misspellings and translations. Semrush, for instance, says it identifies mentions with its own AI brand-extraction system that recognizes sub-brands and products, and its Mentions metric counts the prompts whose answers include the brand.
A mention differs from a citation. An engine can draw on a retrieved page and still not link it, so an uncited mention does not prove that the name came from training memory; its provenance cannot be settled from the answer text alone. Mention rate and citation share are best reported side by side, because the gap between them is informative in itself.
Why it matters
Many category questions, such as "best project management tools", are answered with lists of names and few links. A brand can be absent from citation dashboards and still be recommended in a large share of answers, or the reverse. Mention rate measures that recommendation layer directly. Teams use it to check whether a brand belongs to the default shortlist an assistant produces, and the gap between mentions and citations points either to owned pages (to earn the link) or to third-party coverage (to earn the mention).
Example
Asked for monitoring tools, an engine names three products without linking any of them: three mentions, zero citations. Across 200 sampled answers, a brand named in 46 has a 23 percent mention rate, whatever its citation share.
Also called: AI sentiment, brand sentiment in AI answers, LLM sentiment tracking, AI brand perception
ObservedClaimedOffered by many monitoring tools (observed); the accuracy of each classifier is the vendor's claim.
Sentiment of AI answers is the automated classification of how AI-generated answers describe a brand: positive, neutral or negative overall, and often per attribute such as price, quality or support. Monitoring tools offer it to catch stale, inaccurate or unfavorable narratives early; the direction of change is more reliable than the exact score.
Sentiment tracking applies a classifier, usually itself a language model, to the passages of sampled answers that mention a brand. Outputs range from a single polarity score to attribute-level labels with the themes driving them. a16z's 2025 essay describes tools such as Profound, Goodie and Daydream as letting brands "track sentiment across model outputs". Semrush documents an Overall Sentiment figure, Key Sentiment Drivers (the themes behind positive or negative perceptions) and the sentiment attached to cited pages.
The scope is how answers frame a brand, which is separate from whether the brand appears (mention rate) or is cited (citation share). Scores depend on the classifier, the prompts and the sampling, and neither source reviewed here publishes the accuracy of its classifier, so the trend within one tool is more trustworthy than the absolute figure.
Why it matters
An assistant that calls a product expensive, outdated or discontinued can shape a buyer's shortlist before any visit to the site. Sentiment tracking surfaces those narratives, including ones built on old pricing pages, stale reviews or plain errors. The useful practice is to read the flagged answers, trace each negative claim to the sources the engine cites, and fix what can be fixed: update the page, correct the third-party listing, publish the missing fact. The score works best as an alarm that sends someone to read.
Example
After a price increase, weekly sampling shows answers drifting from "affordable" to "expensive for what it offers", and the cited source is an outdated comparison article. The team updates its pricing page and asks that publisher for a correction.
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Zach Cohen and Seema Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: AI referrals, LLM referral traffic, AI Assistant channel, chatbot referral traffic
DocumentedObservedReporting behavior is documented by Google and Microsoft for their own products; referrer-based attribution of other engines is observed practice with known blind spots. ChatGPT's `utm_source` tag is reported by a third party (LLM Pulse, 2026-10-13); OpenAI's publisher FAQ could not be read at the source.
AI referral traffic is the visits a site receives when people click links inside AI answers, from referrers such as chatgpt.com, perplexity.ai, copilot.microsoft.com or gemini.google.com. Analytics tools now group much of it separately, but clicks from Google's AI Overviews and AI Mode stay mixed into ordinary Google organic traffic.
AI referral traffic is the click side of AI visibility: sessions that arrive from an answer engine with an identifiable referrer or tag. Since 2026-05-13, Google Analytics 4 includes an AI Assistant channel in its default channel group. Google defines it as traffic from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, given the medium ai-assistant when the referrer matches Google's list of assistants. Google's own AI Overviews and AI Mode are excluded and count as Organic Search.
On the Google side, Search Console counts AI Overviews and AI Mode traffic inside the Web search type of the Performance report. A separate Generative AI performance report, announced on 2026-06-03 and rolled out to all sites as of 2026-08-31, shows impressions in AI Overviews and AI Mode by page, country, device and date; its documentation describes impressions only, so clicks from these features still cannot be isolated. Bing Webmaster Tools reports citations in Copilot and Bing's AI summaries (public preview since 2026-02-10), and its announcement does not mention clicks.
How it works
Attribution depends on what reaches the site. A click from a web interface usually carries a referrer header, and some engines also tag links; ChatGPT links have commonly carried utm_source=chatgpt.com. Clicks from mobile apps, links opened with noreferrer, and copied URLs often carry neither, so they land in Direct. In October 2026 the monitoring vendor LLM Pulse reported that links in many ChatGPT answers stopped carrying both the tag and the referrer from about 2026-10-09, cutting ChatGPT-attributed GA4 sessions on the sites it monitors by about 46 percent while Direct rose; OpenAI had not commented at the time of that report.
Why it matters
Referral sessions are the most concrete business outcome of AI visibility, since they can be tied to sign-ups, demos and revenue. They are also undercounted, and a falling AI channel can reflect a tracking change rather than lost interest, as the October 2026 ChatGPT episode suggests. Teams should report the GA4 AI Assistant channel or an equivalent custom channel, watch Direct for unexplained jumps on the same pages, use Search Console's impressions report for Google's AI features, and read all of it next to citation data from sampled answers.
Example
A team sees AI Assistant sessions fall from 1,800 in September to 1,000 in October. Before concluding that ChatGPT sends less traffic, it checks Direct on the same landing pages and finds an offsetting rise starting October 9, consistent with a tagging change.
Also called: AI bot log analysis, log file analysis for AI crawlers, crawler log monitoring
DocumentedObservedIdentification is documented when request IPs are checked against the operators' published ranges; user-agent strings alone can be spoofed, so readings based on them are observed only.
AI crawler log analysis is reading a site's server or CDN logs for the user agents of AI crawlers and fetchers, such as GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot or ClaudeBot, to see what they actually request, how often, and with what responses. It shows access, never whether content was later used or cited.
Log analysis is the one AI visibility signal a site owns outright. Each request line gives the path, status code, user agent, IP address and time, so a team can see which sections each AI system fetches, how often, and whether it meets errors, redirects or pages that need JavaScript to render. Operators document their agents and purposes. OpenAI lists OAI-SearchBot (search), GPTBot (training), ChatGPT-User (user-initiated actions) and OAI-AdsBot (ad landing pages), each with a published IP range file. Perplexity lists PerplexityBot and Perplexity-User, also with IP files.
The limits are strict. Logs record fetches, and a fetch is a separate event from indexing, retrieval or citation. Content can reach answers through other fetchers, partner indexes or caches, so absence from logs does not prove absence from answers. A user-agent string proves nothing until the IP validates against the operator's published ranges or the request carries a verifiable signature.
How it works
Filter logs by the documented user-agent tokens, then verify each hit's IP against the operator's JSON range file (for example openai.com/searchbot.json or perplexity.ai/perplexitybot.json). Group requests by section and status code, and compare them with the sitemap to find sections never fetched. Cloudflare's pay-per-crawl program requires participating crawlers to sign requests with HTTP Message Signatures (Web Bot Auth), a second verification path where it applies. Policies differ by agent: OpenAI says robots.txt rules may not apply to user-initiated ChatGPT-User fetches, and Perplexity says Perplexity-User generally ignores robots.txt.
Why it matters
Logs answer questions no dashboard can: is an engine's search crawler reaching the documentation, does a training crawler respect a disallow, do fetchers hit 404 errors or empty JavaScript shells? These are fixable problems with direct effects on whether content can be retrieved at all. Teams should check logs after every robots.txt or CDN rule change, alert on sudden drops in verified search-crawler activity, and read findings carefully: a section fetched daily is eligible to be used, and that is all a log can show.
Example
Verified logs show OAI-SearchBot fetching /blog daily but never /docs, which sits behind a JavaScript-only router that returns an empty shell. The finding becomes a concrete task: render /docs on the server and recheck the logs two weeks later.
Also called: difference-in-differences, DiD, diff-in-diff, controlled before-and-after test
DocumentedA documented econometric method; its causal reading depends on assumptions that published GEO case studies rarely test.
Difference-in-differences is a test design that estimates the effect of a change by comparing the before-and-after shift on treated pages with the shift on similar untreated pages over the same period. Subtracting the control group's change removes shared background drift, but the result is causal only if the groups would otherwise have moved in parallel.
The estimate is (treated after − treated before) − (control after − control before), in the metric's own units: citations, citation share, sessions. In AI search it answers questions such as whether adding answer capsules to 20 pages changed their citation rate, at a time when engines, models and competitors are all shifting.
It supports a causal claim only under assumptions: parallel trends (without the change, treated and control pages would have moved alike), no spillover between groups, no other change hitting one group only, and stable group membership. The econometrics literature treats violations of parallel trends and staggered rollouts as central problems (Roth, Sant'Anna, Bilinski and Poe). Bertrand, Duflo and Mullainathan showed a separate trap: with outcomes correlated over time, conventional standard errors are badly understated, and in their placebo tests on US wage data, fake laws came out significant at the 5 percent level in up to 45 percent of cases.
How it works
Choose treated and control pages that are similar in topic, traffic and baseline citation level, ideally assigned at random. Measure the metric for several periods before the change to check that both groups moved together, apply the change to the treated group only, and keep measuring both on the same prompt panel and sampling schedule. Report the gap with an uncertainty interval that respects correlation over time, for example by collapsing the data into before and after averages or by randomization inference, two of the fixes Bertrand, Duflo and Mullainathan recommend.
Why it matters
Most published GEO case studies are before-and-after stories without a control group, so they cannot separate a tactic's effect from model updates, seasonality or competitor moves. A 2026 critical survey of GEO research found no reviewed technique with a stable, longitudinal, cross-platform causal effect. Difference-in-differences is the cheapest design that can do better. A team testing a content change should hold back comparable pages as controls, freeze the measurement panel and state its assumptions. A significant gap shows the difference is unlikely to be noise; whether parallel trends held is a separate question.
Example
Answer capsules are added to 20 guides while 20 similar guides stay unchanged. Over six weeks, citations rise by 8 on the treated set and by 1 on the controls, a gap of 7. If the two sets tracked each other during the prior six weeks, the gap is evidence that the capsules helped; otherwise it remains a descriptive difference.
Bertrand, Duflo and Mullainathan, How Much Should We Trust Differences-in-Differences Estimates? (NBER Working Paper 8841; published in the Quarterly Journal of Economics, 2004), published 2002-03, read 2026-11-07: https://www.nber.org/papers/w8841
Roth, Sant'Anna, Bilinski and Poe, What's Trending in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature (arXiv:2201.01194v3), published 2023-01-09, read 2026-11-07: https://arxiv.org/abs/2201.01194
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
ClaimedA vendor and community coinage with no published cross-tool formula and no independent measurement.
Citation decay, sometimes expressed as a citation half-life, is a vendor term for the rate at which a page's citations in AI answers fall away over time when the content is left unmaintained. The direction is plausible, since answers regenerate and sources rotate, but published half-life figures are tool-specific and cannot be compared.
The term borrows from radioactive decay: a citation half-life would be the time it takes for a page's citation count or citation share, on a fixed panel, to fall to half its starting level. Benchmark glossaries such as Lumen GEO and Promptwatch list citation decay among their terms, but no vendor reviewed for this entry publishes a formula another tool could reproduce, and no independent study has measured a half-life across engines.
A computable version is easy to state. Fix a prompt panel and a sampling schedule, measure a page's citation share each week after it peaks, and fit the decline; the half-life is the number of weeks until the share reaches 50 percent of its peak. Without that discipline, a decline can reflect panel changes, model updates or sampling noise as easily as any loss of value in the content.
Why it matters
The idea behind the term is useful even when the numbers are not: AI answers are regenerated continuously, engines refresh their indexes, and newer or better sources displace older ones. A page that earned citations at launch can lose them silently. Teams can turn the claim into a check by tracking citation share for key pages on a fixed panel and scheduling content reviews when a page's share falls. A vendor's half-life figure is that vendor's model output, never a property of the web.
Example
A guide peaks at a 30 percent citation share on a fixed 40-prompt panel in March. Measured weekly, it reaches 15 percent in early May, about eight weeks later. That is the page's observed half-life on that panel, and it does not transfer to other pages, panels or engines.
ClaimedA vendor and community coinage with no published cross-tool formula and no independent measurement.
Citation velocity is a vendor term for how fast a page gains citations in AI answers after it is published or updated. It mirrors citation decay and carries the same caveat: there is no shared formula, so a velocity figure means something only inside the tool and the prompt panel that produced it.
In the benchmark glossaries that use it, citation velocity describes momentum: citations gained per unit of time over an early window. Lumen GEO lists the term; no vendor reviewed for this entry publishes a reproducible definition, and no independent study has measured it.
A computable version: on a fixed prompt panel with a fixed sampling schedule, velocity = (citation share at the end of a window − citation share at its start) ÷ the window's length, for example percentage points per week over the first four weeks after publication. It must be read against a baseline, because engines differ in how quickly they discover and retrieve new pages; Bing, for instance, presents IndexNow as a way to notify participating engines of new or changed URLs so AI systems can reference current versions. A velocity with no stated panel, window and baseline cannot be compared with anything.
Why it matters
Teams want to know how soon a new page can appear in answers, and velocity is the word vendors use to sell that answer. Measured on a fixed panel, early gains show whether a page is being discovered and retrieved at all, which is actionable: a page with zero velocity after several weeks may have a crawl, rendering or indexing problem. Measured loosely, velocity rewards noise. Treat any figure as a within-tool trend, and confirm discovery with server logs and index checks.
Example
A new comparison page goes from 0 to 12 percent citation share on a 30-prompt panel over its first four weeks: 3 percentage points per week. A sibling page stays at zero, and server logs show no search-crawler fetch, which points to a discovery problem before any content problem.
Prompt tracking is the practice of running a fixed set of questions, a query panel, through chosen AI engines on a schedule and recording each answer, its mentions and its citations. The panel is the sampling frame under every AI visibility metric, so its design determines what those metrics can and cannot show.
A query panel specifies the prompts (wording, intent mix, branded and unbranded), the engines and surfaces, the locale and language, any persona or account state, the number of runs per prompt, and the schedule. Each element changes the results. SparkToro's January 2026 study found that people asking for the same thing write very different prompts: 142 volunteer-written prompts about one need had an average semantic similarity of 0.081. A panel written by a marketer can therefore miss how buyers actually ask.
Vendors build panels differently. a16z describes tools that fine-tune models to mirror brand-relevant prompt language and run synthetic queries at scale. Semrush offers prompt tracking on ChatGPT and Google AI Mode and reports, as Visibility, the share of top citations a domain holds across tracked prompts. Comparability over time requires keeping the panel fixed and versioning every change to it.
Why it matters
Every visibility number in this section is a statement about a panel. A panel of easy branded questions flatters a brand, a panel of competitive category questions humbles it, and swapping prompts every week makes trends meaningless. Teams should write the panel down, mix branded and unbranded questions that resemble real buyer wording, run each prompt several times, and change the panel only in dated versions. When comparing vendors, ask to see the panel before the dashboard.
Example
Fifty fixed prompts such as "best AI search monitoring tools" and "how to measure AI search visibility", run weekly on ChatGPT, Perplexity and Google AI Mode with five runs each, form a panel. Replacing ten prompts in June starts a new panel version, reported separately.
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Zach Cohen and Seema Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: answer non-determinism, answer variability, AI answer volatility, run-to-run variability, sampling variance
MeasuredObservedMeasured for brand recommendations in one named study (SparkToro, 2026) and reported across commercial audits in a 2026 survey; a known property of generative systems.
Answer non-determinism is the tendency of AI engines to give different answers, and different citations, to the same question asked repeatedly. Generation is sampled and retrieval changes over time, so a single check proves little: visibility has to be measured as a rate over many runs of each prompt, with an error margin.
Two sources of variation combine. The model samples its words from probability distributions, so phrasing, lists and order shift between runs, and the retrieval layer can return different pages as indexes, rankings and query rewrites change. The result has been measured. In SparkToro's study (published 2026-01-28, with Gumshoe.ai), 600 volunteers ran 12 brand-recommendation prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. ChatGPT and Google's AI had less than a 1 in 100 chance of returning the same list of brands in any two responses, and identical ordering appeared about once in 1,000 runs. Martinez's 2026 critical survey of GEO likewise reports "substantial run-to-run variability" in commercial audits.
The term covers variation for one fixed prompt. Variation caused by different wordings of the same need is a separate issue of panel design.
How it works
Treat each answer as one draw. For a prompt run n times with the brand appearing in k of them, the visibility rate is p = k ÷ n, and a simple binomial standard error is √(p(1 − p) ÷ n). With p = 0.4 and n = 10, the error is about 0.15, so the true rate could plausibly sit anywhere from about 0.1 to 0.7; with n = 100 it narrows to about 0.05. Repeated runs spread over several days are what make a rate trustworthy.
Why it matters
Non-determinism is why a screenshot of one answer, flattering or damning, is poor evidence, and why rank tracking imported from SEO breaks down. SparkToro's authors concluded that a visibility percentage across many prompts run multiple times is a reasonable metric, while a single ranking position in AI answers is not. Teams should budget for repeated sampling, report rates with error margins, and distrust week-to-week swings that fall inside the noise.
Example
The same prompt, run 10 times on one engine in one afternoon, cites a site 4 times. Reporting "we are cited" or "we are not cited" from any single run would be noise; 40 percent, with a wide margin, is the honest reading.
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: citation gap, source gap analysis, missing sources analysis, AI content gap analysis
Observed
Citation gap analysis compares the sources AI engines cite for a set of target questions with a brand's own presence, to find the questions where competitors or third-party sites are cited and the brand is absent. The resulting list of gaps, by question and by cited source, becomes a plan for content and outreach work.
The analysis starts from a fixed prompt panel. For each prompt it records every cited domain and page across repeated runs, then sorts prompts into those where the brand is cited, those where it is mentioned without a citation, and those where it is absent. For the absent ones, it lists the sources engines rely on instead: a competitor's comparison page, a review site, a forum thread, a news article. Vendors package this under their own names. Semrush, for instance, reports Source Opportunities or Missing Sources, meaning frequently cited sites that mention competitors and not the brand, and a16z describes dashboards that show which publishers shape model behavior.
The scope covers missing owned pages and third-party sources where the brand is not represented. A gap describes the sampled answers; filling it is no guarantee of citations, since engines weigh many factors they do not publish.
Why it matters
Gap analysis turns abstract visibility numbers into a work list. If engines answer a buying question by citing two competitor guides and a Reddit thread, the options are concrete: publish a better primary source, earn a place in the third-party sources engines already use, or both. It also stops teams from polishing pages that engines never consider for the question. The output should be re-measured on the same panel after the work ships, ideally with control prompts, before anyone claims success.
Example
For 20 "AI search analytics" prompts run five times each, engines cite two competitor blogs and one Reddit thread, and never the brand's documentation. The gap list assigns a new methods page for the first cluster of prompts and community participation in the cited forum for the second.
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Zach Cohen and Seema Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: share of model, SoM, Share of Model (Jellyfish), LLM share of voice
ClaimedIntroduced and trademarked by the agency Jellyfish for its platform; no public formula, so every figure is proprietary.
Share of model is a term introduced by the marketing agency Jellyfish in 2024 for how large language models perceive and present a brand within its category, and the name of its trademarked platform. Others now use it loosely for a brand's share of AI mentions. With no published formula, each share-of-model figure is proprietary.
Jellyfish, part of The Brandtech Group, promoted share of model as a new marketing metric in July 2024, in a piece by Jack Smyth published through the ANA, and launched its Share of Model platform on 2024-12-05. The launch announcement describes "a single-view interface" for analyzing how different LLMs perceive brands, covering ChatGPT, Gemini and Llama at launch, and for judging whether a brand's digital presence earns their recommendations. Neither page publishes a formula.
In wider use, share of model has become a loose synonym for a brand's share of mentions or citations in AI answers within a category, by analogy with share of voice and share of search; benchmark glossaries such as Lumen GEO and Promptwatch list it. Because each implementation blends mentions, citations, prominence or perception differently, a share-of-model number is a proprietary dashboard figure. It can show a trend inside one tool and cannot be compared across tools.
Why it matters
The term travels well in boardrooms because it echoes share of voice, so it often anchors AI search budgets and agency reports. That is exactly why it needs translating. Before accepting a share-of-model figure, a team should ask which models, prompts, locales and runs it covers, whether it counts mentions, citations or both, and how competitors are defined. Restated as panel-based primitives (mention rate, citation share, competitive share of voice), the number becomes checkable and comparable over time.
Example
An agency reports "share of model: 18 percent" for a running-shoe brand. Asked for the method, it explains that the figure blends unprompted mentions across three models with a prominence weight. The client adds a plain mention rate on its own panel to keep a number it can audit.
Also called: AI search monitoring tools, AI visibility tools, GEO tools, AI rank trackers, LLM monitoring platforms
ObservedClaimedThe category's existence is observed; every tool's metrics and methods are its vendor's claims, and none has been validated across vendors.
Monitoring tools are platforms that run panels of prompts through AI engines on a schedule and report how a brand appears: mentions, citations, prominence, sentiment and competitor share. The category includes specialist startups and AI modules added to established SEO suites; their methods are proprietary, so their numbers are not directly comparable.
a16z's May 2025 essay named Profound, Goodie and Daydream as tools that let brands see how they appear in AI responses, track sentiment and identify the publishers that shape model behavior, and described their method as fine-tuning models to mirror brand-relevant prompt language and running synthetic queries at scale. It also named Ahrefs Brand Radar, for brand mentions in AI Overviews, and the Semrush AI toolkit, now presented as Semrush AI Visibility. Other specialist products with live sites on 2026-11-07 include Otterly.ai, Peec AI, AthenaHQ, Scrunch and Evertune.
This is a category map, not a recommendation. The tools differ in where prompts come from (written, synthetic or licensed conversation data), which engines they cover, how many runs they make per prompt, how they detect brands and how they compute each metric. None of these methods has been validated across vendors.
How it works
A typical setup defines a few hundred prompts, selects engines (for example ChatGPT, Perplexity, Gemini, Google AI Overviews and AI Mode), runs each prompt several times per week, extracts brand mentions and cited URLs, and aggregates them into visibility, citation, sentiment and competitor dashboards. Some tools add estimated prompt volumes or recommendations for content and outreach.
Why it matters
Most teams will measure AI visibility through one of these tools, so the tool's methodological choices become the team's view of reality. Tools can disagree about the same brand in the same week because their panels and formulas differ, and SparkToro's 2026 study shows answers vary so much between runs that thin sampling produces noise. Judge a tool by its transparency: can it export the prompts, raw answers and citations, state the runs per prompt, and explain each metric's formula? A headline score without those is hard to act on.
Example
Two vendors trial on the same brand for a month. One reports 41 percent visibility on 120 prompts run three times a week; the other reports 23 percent on 500 prompts run once. Exporting both panels shows the first is mostly branded questions, which explains most of the gap.
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Zach Cohen and Seema Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: unaided awareness, spontaneous brand mention, unprompted brand recall in AI answers
ObservedA measurement concept carried over from brand research; no standard protocol for AI answers.
Unaided awareness in AI answers is whether a model names a brand spontaneously when asked a category question that does not mention it, such as "best winter jackets". a16z called it an indicator of unaided awareness in the AI era. It measures presence under given conditions, and says nothing about why the model knows the brand.
The concept comes from brand research, where unaided recall asks people to name brands in a category without prompting. a16z's 2025 essay carried it to language models with a Canada Goose example: the brand used a GEO tool to see how LLMs referenced it, and the key question was "whether the model spontaneously mentioned the brand at all". The essay does not report what the tool found.
Measured properly, unaided awareness is a mention rate on unbranded category prompts: the share of sampled answers that name the brand when the prompt names only the need. It depends on the model, its settings, the locale, the date and whether the engine searched the web. It does not isolate training data as the source, since grounded answers can name a brand unprompted too. Aided awareness, by contrast, asks about the brand directly, as in "what do you know about this brand?"
Why it matters
Unbranded category questions are where new buyers start, and being named there is the AI equivalent of entering the consideration set. A brand with strong branded answers and low unaided awareness is invisible to people who do not already know it. Teams should separate branded and unbranded prompts in their panels, track the unaided mention rate per engine, and work on the inputs engines draw on for category questions: third-party reviews, comparisons, lists and authoritative coverage.
Example
Across 30 unbranded prompts such as "warmest parka for city winters", run five times on two engines, a brand is named in 51 of 300 answers: a 17 percent unaided mention rate for that panel, model mix and month.
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Zach Cohen and Seema Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: AI dark traffic, dark AI traffic, hidden AI traffic, AI agent traffic
ObservedChatGPT's `utm_source` tag is reported by a third party (LLM Pulse, 2026-10-13); OpenAI's publisher FAQ could not be read at the source.
AI dark traffic is the visits influenced or generated by AI systems that analytics cannot attribute to them: clicks from assistant apps without a referrer, copied links, and pages fetched by agents on a user's behalf. Much of it lands in Direct or never reaches analytics at all, so referrer reports undercount AI's role.
Two flows hide here. The first is human clicks that lose their attribution: links opened from mobile apps, links that use noreferrer, and URLs copied from an answer into a browser arrive without a referrer and are counted as Direct. GA4's AI Assistant channel depends on the referrer matching Google's list of assistants, so it cannot recover these visits. In October 2026 the vendor LLM Pulse reported that many ChatGPT answer links had stopped carrying a referrer and the utm_source=chatgpt.com tag, moving sessions from the ChatGPT line to Direct.
The second is machine visits. User-initiated fetchers such as ChatGPT-User and Perplexity-User retrieve pages to answer a question, and browsing agents navigate sites for people. These requests appear in server logs but usually not in JavaScript analytics, and OpenAI and Perplexity both say robots.txt may not govern such user-initiated fetches.
Why it matters
If AI's contribution is measured only through referral sessions, it will be systematically undercounted, and content and budget decisions will follow the undercount. The correctives are partial but cheap: watch Direct traffic on pages that AI answers cite, analyze verified fetcher activity in server logs, add a "how did you hear about us" field that lists AI assistants, and compare these signals over time. None proves causation alone; together they bound the hidden share.
Example
A user asks the ChatGPT iOS app for vendors, taps a cited link and lands with no referrer, so GA4 records Direct. The same week, the server logs a ChatGPT-User fetch of that page from a verified OpenAI IP address, which the JavaScript analytics tag never sees.
Also called: AI reference rate, source reference rate
ObservedA concept named in a16z's 2025 essay, with no standard formula.
Reference rate is a16z's name for how often a brand or its content is cited or used as a source in AI-generated answers, a measure the firm's 2025 essay set against click-through rate as the one that matters in generative search. It is a concept rather than a formula, usually computed as citation share or mention rate.
a16z partners Zach Cohen and Seema Amble used the phrase in their May 2025 essay on generative engine optimization, describing "reference rates: how often your brand or content is cited or used as a source" in model-generated answers. Their argument: when answers replace lists of links, the question shifts from where a page ranks to whether the model chooses to reference it.
The essay gives no formula. Two computable readings fit the wording. A citation reading equals citation share: the share of sampled answers that cite the domain. A usage reading is broader, counting answers that name or paraphrase the brand's content even without a link, which approaches mention rate and the absorption measures of the research literature. Any report using the term should say which reading it applies, and over which prompt panel.
Why it matters
The term works as a framing: it moves reporting away from average position and click-through rate, which describe lists of links, toward a measure of being used as evidence. Without a formula, though, it can mean whatever a vendor wants. Teams adopting it should define it explicitly, usually as citation share on a fixed panel, add mention rate for the uncited layer, and keep reporting clicks separately so the framing does not hide traffic losses.
Example
A quarterly report replaces "average position 3.2" with "cited in 22 percent of sampled answers across our 80-prompt category panel on three engines, up from 17 percent". That is reference-rate thinking with an explicit formula.
a16z, How Generative Engine Optimization (GEO) Rewrites the Rules of Search (Zach Cohen and Seema Amble), published 2025-05-28, read 2026-11-07: https://a16z.com/geo-over-seo/
Also called: AI Overview trigger rate, AI Overview prevalence, AIO presence rate
MeasuredObservedOne study (Pew, US, March 2025) measured it; current rates come from trackers with their own panels.
AIO trigger rate is the share of searches in a defined set for which Google shows an AI Overview. Pew Research measured 18 percent of searches in a US panel in March 2025, with rates rising sharply for long and question-style queries. Rates for a site's own keywords have to be tracked on its own query panel.
AIO trigger rate = (queries in the panel that return an AI Overview) ÷ (queries checked), for a stated locale, device and date. The measured anchor is one study. Pew Research Center tracked the browsing of 900 US adults in its KnowledgePanel Digital panel during March 2025; a scraping service re-ran their 68,879 unique Google searches between April 7 and 17, 2025, and 12,593 produced an AI summary. Some 18 percent of searches generated one. Only 8 percent of one- or two-word searches did, against 53 percent of searches of ten words or more, and 60 percent of queries starting with who, what, when or why.
The scope is Google's AI Overviews. AI Mode, where every response is generated, needs different measures. Commercial rank trackers publish their own rates by industry or keyword set; those depend on their panels and cannot be compared with Pew's browsing-based sample.
Why it matters
Whether an AI Overview appears decides whether a query's traffic follows classic organic patterns or the AI-summary pattern, in which Pew found people clicked a traditional result in 8 percent of visits, against 15 percent when no summary appeared. Teams should measure the trigger rate on their own target queries by locale and device, segment by query type, and focus AI search work where Overviews fire. Search Console's Generative AI performance report now adds page-level impressions in AI Overviews and AI Mode.
Example
A team checks its 200 target queries monthly in the US on mobile. AI Overviews fired on 60 in March and 95 in October, so its measured rate went from 30 to 48 percent. That describes its panel's readings; locale, panel or method changes can move the number without the feature itself expanding.
Also called: prompt volume, AI search volume, Prompt Volumes (Profound), AI Topic Volume (Semrush)
ClaimedEngines publish no prompt counts; every prompt-volume figure is a vendor's modeled estimate.
Prompt volume is a vendor term for an estimate of how often a question, or a topic, is asked to AI assistants: the hoped-for counterpart of keyword search volume. AI engines publish no such counts, so every figure is modeled by a vendor from its own data, and figures differ in method, coverage and granularity.
In classic search, Google Ads and Search Console provide metered query data. AI assistants publish nothing comparable, so vendors estimate. Profound describes its Prompt Volumes as "a dataset built by licensing AI answer engine conversations", anonymized, cleaned and then subjected to probabilistic modeling, covering ChatGPT, Gemini and Perplexity with coverage that varies by region. Semrush says individual prompts are often too specific to measure directly, so it reports AI Topic Volume at the level of topics, drawn from a database of over 317 million prompts and responses.
The term therefore names a family of modeled estimates with different units: per prompt, per topic, per region, per engine. None is a count reported by the engines themselves, and none of the vendors reviewed here cites an independent validation against engine logs.
Why it matters
Teams need a way to choose which questions to track and which content to build, and prompt volume promises a demand signal. Used carefully, it ranks topics by relative interest and shows how people phrase needs to assistants. Used carelessly, it imports false precision from keyword tools. Treat each figure as a vendor's directional estimate, compare topics within one tool only, and check priorities against other signals such as Search Console queries, sales conversations and support tickets.
Example
A tool estimates 2,400 monthly prompts about "AI search platforms" in the US. That modeled figure is useful against the same tool's estimate of 600 for a neighboring topic, and it is nothing like Search Console's metered impressions.
Also called: click-through to source, CTS, AI citation click-through rate, source click-through rate
MeasuredClaimedA vendor metric (claimed); the underlying behavior has one measured anchor, Pew's 2025 study.
Click-through to source is the rate at which people who see an AI answer click one of its cited source links, sometimes measured for a single cited domain. The measured anchor is small: in Pew Research's 2025 study, users clicked a link inside a Google AI summary in about 1 percent of visits to pages showing one.
As a metric, CTS = (clicks on cited source links) ÷ (impressions of answers containing those links), for all sources or for one domain. Engines rarely give site owners both the numerator and the denominator, so most CTS figures come from vendors' own panels or modeling, and benchmark glossaries name variants such as AI citation click-through rate.
The independent anchor is Pew Research Center's browsing study of 900 US adults in March 2025. Pew reported that clicking a link inside an AI summary "occurred in just 1% of all visits to pages with such a summary". That figure covers any cited source, on Google only, at that time; it says nothing directly about one brand's click rate, and it cannot be converted into one without conditional data. On Google's side, Search Console's Generative AI performance report shows impressions in AI Overviews and AI Mode, and its documentation describes no click metric.
Why it matters
A citation is worth less than it looks if almost nobody clicks. CTS is the bridge between citation metrics and traffic, and Pew's 1 percent figure warns against forecasting visits from citation counts. Teams should treat citations mainly as exposure and influence, measure actual clicks in analytics (the GA4 AI Assistant channel, and Search Console for Google), and be skeptical of any vendor CTS figure that does not disclose how impressions were counted.
Example
A forecast assumes every citation brings visits. Pew's anchor suggests caution: on pages with an AI summary, a link inside it was clicked in about 1 percent of visits, across all cited sources, so 10,000 such visits might produce on the order of 100 source clicks in total.
Also called: citation-metric zoo, vendor citation metrics, GEO metric coinages
ClaimedA class of vendor coinages, each defined only inside its own tool's methodology.
The citation-metric zoo is a collective label for the many metric names vendors have coined for AI search, such as citation density, citation halo, citation readiness or entity salience. Each is defined only inside its own tool's methodology, none has a cross-vendor definition or independent validation, and each should be translated into panel-based primitives before comparison.
Benchmark glossaries reviewed for this entry list, among others: citation decay and citation half-life, citation velocity, citation density, citation halo, citation absorption, citation pipeline, citation signal and entity density (Lumen GEO); citation readiness (geodocs.dev); entity salience and AI answer volatility (Promptwatch). The names suggest precise quantities, but none has a published cross-vendor definition, and no independent study reviewed here validates any of them.
Some names echo research vocabulary. Absorption is studied in the academic literature on generative engines, where Martinez's 2026 survey treats it as one component of a source's visibility, and citation probability appears there as a modeled quantity decomposed into activation, retrieval and citation given retrieval. The vendor scores that borrow these words remain proprietary. The label is descriptive: it records the vocabulary so it can be recognized in sales decks and implies nothing about any one vendor's accuracy.
Why it matters
Vendor decks and agency reports increasingly lead with these names, and a buyer who cannot decode them cannot compare offers or hold a provider to account. In contracts and reports, ask for every coined metric to be restated in the primitives defined in this section: citation share, citation rate, citation prominence and mention rate, over a stated prompt panel with stated sampling. If a vendor cannot restate a metric that way, treat it as a proprietary index and give it little weight.
Example
A proposal promises to raise "citation readiness" from 54 to 80. Asked to restate the goal, the agency commits instead to raising citation share on an agreed 60-prompt panel from 9 to 15 percent, measured with five runs per prompt on three engines.
Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), arXiv:2607.14035v1, full text, published 2026-07-15, read 2026-11-07: https://arxiv.org/html/2607.14035v1
DocumentedA peer-reviewed framework (EACL 2024 System Demonstrations) with open-source documentation.
RAG evaluation is the technical scoring of a retrieval-augmented generation system's own quality: whether retrieval returns relevant, focused context, whether answers stay faithful to that context, and whether they address the question. The RAGAS framework popularized automated, LLM-graded versions of these metrics, run by a system's builders on their own pipeline.
The RAGAS paper by Es, James, Espinosa-Anke and Schockaert (arXiv:2309.15217, posted 2023-09-26 and revised 2025-04-28; published in the EACL 2024 System Demonstrations) proposed reference-free evaluation along three dimensions: whether the retriever finds relevant, focused passages, whether the model uses them faithfully, and the quality of the generated answer. The open-source Ragas library defines the metrics precisely. Faithfulness = claims in the response supported by the retrieved context ÷ total claims in the response, from 0 to 1. Answer (response) relevancy is the mean cosine similarity between the user's question and several questions an LLM generates back from the answer. Context precision averages precision@k over the ranked retrieved chunks, so relevant chunks ranked higher score better. Context recall = claims in a reference answer supported by the retrieved context ÷ claims in that reference, so unlike the paper's core metrics it requires a reference.
The scope is a system its operator controls. Measuring how other companies' engines cite a brand is a different discipline.
Why it matters
Keeping this boundary clear prevents a common confusion. RAG evaluation tells a team how well its own assistant, support bot or site search answers from its own documents, and it is tuned through chunking, retrieval and prompts. AI search visibility measurement looks at engines the team does not control, from outside, with no access to their retrieval. The ideas overlap, since faithfulness resembles the fidelity measures applied to commercial engines, but the metrics, tools and levers differ, and mixing them produces meaningless dashboards.
Example
Before launch, a team scores its documentation assistant on 300 test questions: context precision 0.81, faithfulness 0.92, answer relevancy 0.88. It then improves chunking and reranking until faithfulness passes 0.95, work with no direct effect on how outside engines cite its site.
Also called: absorption, fidelity, source absorption, citation fidelity
MeasuredEmergingFidelity gaps are measured (Liu et al., 2023, and later audits summarized in Martinez's 2026 survey); absorption as a named measure is recent research vocabulary with no settled method.
Absorption and fidelity describe whether an AI answer actually uses a source and represents it accurately, beyond citing it. In Martinez's 2026 survey of generative engine optimization, absorption is a source's effective contribution to an answer's facts, wording or structure, and fidelity is the extent to which attributed claims are supported and rendered faithfully.
Martinez's critical survey (arXiv:2607.14035, 2026-07-15) proposes a visibility vector for a source with seven components: discoverability, exposure in context, citation, observable prominence, absorption, fidelity, and behavioral or economic outcome. Separating them makes plain that a source can be cited without being used, or used without being cited. The survey notes that published absorption scores combine position, repetition, coverage and textual similarity without revealing a causal trace, and that a real measure needs a matched ablation: does the answer change when the source is removed?
Fidelity is measured with citation recall (the share of verifiable sentences fully supported by their citations) and citation precision (the share of citations that support their sentence). Liu, Zhang and Liang's 2023 audit of Bing Chat, NeevaAI, Perplexity and YouChat found that 51.5 percent of generated sentences were fully supported by citations, and that 74.5 percent of citations supported their sentence.
Why it matters
Citation dashboards count links, and links are only part of influence. A citation without absorption is decoration, absorption without citation is invisible to every citation metric, and a citation with low fidelity can attach a brand's name to claims its pages never made. Measuring these dimensions means reading answers against source text: sampling answers, extracting claims, and checking which trace to the brand's content and whether they are stated correctly. Teams can start with the pages that matter most, such as pricing, specifications and research findings.
Example
An engine states a study's headline figure without linking the study. If the figure is distinctive and the brand was its only publisher at answer time, absorption without citation is a reasonable suspicion, short of proof, since copies and model memory can reproduce it. Elsewhere, an answer cites the brand's pricing page but states last year's price: a citation with a fidelity failure.
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
Martinez, Optimizing Visibility in Generative Engines (arXiv:2607.14035v1), full text, published 2026-07-15, read 2026-11-07: https://arxiv.org/html/2607.14035v1
Liu, Zhang and Liang, Evaluating Verifiability in Generative Search Engines (arXiv:2304.09848, Findings of EMNLP 2023), published 2023-04-19, read 2026-11-07: https://arxiv.org/abs/2304.09848
Also called: Search Generative AI performance report, Search Console AI report, AI Overviews impressions report, Search Console generative AI insights
Documented
Generative AI performance report is a Search Console view, introduced on 2026-06-03 and available to all sites since 2026-08-31, that shows how often a site's links appeared in Google's generative AI features: AI Overviews and AI Mode in Search, with a separate report for Discover. Its only metric so far is impressions.
Before this report, AI Overviews and AI Mode activity was counted inside the overall Web search performance data with no way to isolate it. The generative AI performance report adds a dedicated view, while the same impressions stay in the main report. Google defines an impression as a link to the site shown to a user in a generative AI feature, aggregated by property, so two links from the same site in one response count once in the chart total.
Dimensions are pages (final URL after redirects), countries, devices for Search, and dates at hourly, daily, weekly or monthly granularity. A search type filter separates 'Web: text-based' queries from 'Web: multimodal' image-based searches such as Google Lens and Circle to Search. Search Labs experiments are excluded, the newest data may be preliminary, and the usual 1,000-row table limit applies. Google lists impressions as the only metric and says more will be added over time; the report does not show clicks or the queries behind an impression.
Why it matters
It is the first data Google itself provides on visibility in its AI features, so teams can see which pages actually appear in AI Overviews and AI Mode, in which countries, and how that changes after content or technical work. Its limits shape its use: impressions without clicks or queries cannot show which questions triggered a page or whether users visited, so it complements prompt tracking and analytics instead of replacing them. Google also points site owners to it for gauging the traffic at stake before using the Search generative AI control to opt out.
Example
A recipe site filters the report to the last 28 days and finds that 40 pages account for most AI feature impressions, nearly all on mobile, while its long technique guides barely appear. It compares the structure of the two groups before rewriting the guides.
Also called: Bing AI Performance, Bing AI citations report, grounding queries, Bing citation share
DocumentedEmergingPublic preview since 2026-02-10, with further preview features from 2026-06-16; grounding-query data is sampled.
AI Performance report is a Bing Webmaster Tools view, in public preview since 2026-02-10, that shows how often a site's pages are cited as sources in Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. It reports total citations, cited pages and grounding queries, the phrases the AI used when it retrieved the site's content.
Microsoft's report counts citations rather than visits. Total citations is the number of times a site's content was shown as a source in AI answers during a period; average cited pages is the daily average of unique pages cited; page-level activity breaks citations down by URL; and a timeline shows trends. Grounding queries are the key phrases the AI system used to retrieve content that was later cited: a sampled, grouped view of retrieval, distinct from the users' exact prompts.
On 2026-06-16 Bing added preview features. Intents classify grounding queries into categories such as informational, commercial or local; Topics cluster them into themes; Citation Share gives the site's percentage of all citations shown across sites for a grounding query; and Compare overlays a previous period. Microsoft states that citation counts reflect how often pages are cited, not their importance, ranking or placement in the answer. Microsoft's announcement lists no click or traffic metric.
Why it matters
Unlike Google's generative AI report, it exposes the retrieval side of an answer engine: grounding queries show which phrases actually pulled a site's pages into Copilot answers, which comes close to observing query fan-out directly. Teams can use it to find pages that earn citations, compare topics where they are cited with topics where they are absent, and check whether content changes move citation counts. Its scope is Microsoft's surfaces only, so it says nothing about ChatGPT, Gemini or Perplexity, and without clicks it has to be paired with analytics.
Example
A solar installer's report shows 1,200 citations in a month, most under grounding queries grouped around solar panel cost and net metering, while its battery storage pages are never cited. It adds a cost breakdown and a plain-language battery FAQ to those pages and compares the next month's counts.
Also called: GA4 AI Assistant channel, ai-assistant medium, AI assistant traffic channel, (ai-assistant) campaign
DocumentedPer Google Analytics help pages; the 2026-05-13 announcement names ChatGPT, Gemini and Claude, while the channel definition names ChatGPT, Gemini, Deepseek, Copilot and Grok as examples.
AI Assistant channel is a default channel in Google Analytics 4, added on 2026-05-13, that groups visits arriving from AI chatbots such as ChatGPT, Gemini, Copilot, Deepseek and Grok. Matching sessions get the medium 'ai-assistant'. Clicks from Google's own AI Overviews and AI Mode remain in Organic Search.
Before May 2026, chatbot visits typically appeared in GA4 as Referral traffic, and analysts built custom channel groups with referrer rules to isolate them. Google Analytics now assigns visits whose referrer matches a recognized AI assistant to an AI Assistant channel in the Default Channel Group, with the medium set to ai-assistant and the campaign to (ai-assistant). Google's channel definitions describe it as covering users who arrive from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, and state that it excludes Google's AI Overviews and AI Mode, which the Organic Search definition explicitly includes.
The channel depends on a referrer being passed, so visits from apps that strip the referrer still land in Direct. Google notes that default definitions may evolve as the market changes, and they cannot be edited, so teams that need a different list of assistants still build a custom channel group placed above Referral.
Why it matters
A default channel makes AI referrals visible to every GA4 user without setup, so assistant traffic can be compared with search, social and email in standard reports and tied to conversions. It also creates a split teams must understand: clicks from Google's AI features count as Organic Search while clicks from other assistants count as AI Assistant, so the channel understates total AI-driven traffic. Teams should annotate the change date, avoid comparing the new channel with earlier Referral data without adjustment, and check which assistants their custom groups catch that the default list misses.
Example
A SaaS company sees the new AI Assistant channel deliver 3% of sessions but 7% of trial sign-ups in June, mostly from chatgpt.com landing on comparison pages, and starts reporting the channel weekly beside Organic Search.
Also called: crawl-to-referral ratio, crawl-to-click ratio, crawl-to-click gap
DocumentedMeasuredMetric defined by Cloudflare and measured on its network only; Cloudflare notes ratios are overstated where native apps send no referrer.
Crawl-to-refer ratio is a metric that compares how many pages a search or AI platform's bots fetch from websites with how many visitors that platform sends back through links. Cloudflare has published it on Radar since 2025-07-01; a ratio of 1,000:1 means a thousand crawl requests for each referred visit, a rough gauge of what crawling returns.
Cloudflare calculates the ratio by dividing the HTML requests made by a platform's relevant crawler user agents by the HTML requests whose Referer header points to that platform's hostnames, normalized to one referral. The data covers sites on Cloudflare's network. Radar shows it per platform on its AI Insights page, with time series in the Data Explorer and through an API.
The metric has known distortions. Native apps, including Claude's according to Cloudflare, often send no Referer header, so their referrals go uncounted and their ratios look worse than reality. Training crawlers send no visitors by design, so a platform that crawls heavily for training shows a high ratio whatever its search product does. In Cloudflare's first published week, 2025-06-19 to 2025-06-26, Anthropic's ratio was about 70,900:1 while Mistral's was 0.1:1. Figures move week to week and should always be read with their dates.
Why it matters
The ratio puts a number on the exchange behind AI crawling: content taken against traffic returned. Publishers use it to decide which bots to allow, which to rate-limit, and which to route to licensing or pay-per-crawl arrangements, and it frames negotiations with AI companies. For a marketing team the more useful version is its own: comparing crawler hits in server logs with AI referral sessions per platform shows which assistants read the site and which send buyers. A high ratio does not prove a bot is useless, since a citation without a click can still build awareness.
Example
A publisher's logs show one AI platform's crawlers fetching 400,000 pages in a month while analytics records 2,000 sessions referred by that platform, a site-level ratio of 200:1. It weighs that against the brand mentions it sees in the platform's answers before changing its robots.txt.
Also called: Is It Agent Ready, isitagentready.com, agent readiness check, agent-readiness audit
DocumentedClaimedThe checks and scan results are documented by Cloudflare; no study shows that a higher score changes how often agents use or cite a site.
Agent Readiness score is a free check from Cloudflare, launched on 2026-04-17 at isitagentready.com, that scans a website for standards AI agents can use to find, read, access and act on it, such as robots.txt rules, Markdown negotiation, Content Signals, Web Bot Auth, Agent Skills and MCP Server Cards, then returns a score with fixes.
Cloudflare frames the score as a Lighthouse-style audit for agents. It groups checks into four dimensions: Discoverability (robots.txt, sitemap.xml, Link response headers), Content (Markdown content negotiation), Bot Access Control (Content Signals, AI bot rules in robots.txt, Web Bot Auth) and Capabilities (Agent Skills, API Catalog, OAuth server discovery, MCP Server Card, WebMCP). It also checks agentic commerce standards such as x402, the Universal Commerce Protocol and the Agentic Commerce Protocol without counting them toward the score, and it leaves llms.txt out of the default scan. Each failed check comes with a prompt a coding agent can use to implement the fix.
Cloudflare's scan of the 200,000 most visited domains found that 78% had a robots.txt, 4% declared AI usage preferences through Content Signals, 3.9% passed Markdown negotiation, and fewer than 15 sites had MCP Server Cards or API Catalogs. The score measures adoption of a chosen set of standards; it says nothing about content quality or citation.
Why it matters
It gives teams a quick, shared checklist for the agent-facing layer of a site, a layer classic SEO audits do not cover, and its low baseline numbers show how early that layer is. The risk is treating the number as a goal: several checks (OAuth discovery, MCP Server Cards, API catalogs) only matter for products that expose actions or APIs, and the selection reflects Cloudflare's view of which standards matter, including several it created or co-authors. Use it to find cheap wins such as explicit bot rules or Markdown delivery, and rank the rest by whether agents actually visit and transact on the site.
Example
A B2B SaaS site scores low because it has no sitemap reference, no AI bot rules and no Markdown version of its docs. The team fixes those three in a day and leaves the MCP Server Card check for when its public API is ready for agents.