AEO glossary: AI search terms explained
Plain definitions of the terms used in answer engine optimization and AI visibility, from fan-out and grounding to share of voice and confidence ranges.
AEO
Also called Answer engine optimization
The work of getting a brand named, recommended and cited in AI-generated answers from ChatGPT, Claude, Gemini, Perplexity and Google's AI features. It overlaps heavily with SEO, because most engines search an index and read pages before they answer.
AI crawler
Also called AI bot, AI user-agent
A bot an AI company runs to fetch web pages, either to collect training data, to build a search index, or to read a page live because a user asked. Most operators give each job its own user-agent, so a site can opt out of training and still be found in AI search.
AI Mode
Also called Google AI Mode
Google's conversational search mode, which breaks a question into subtopics and runs many searches at once before writing an answer with links. Ahrefs found AI Mode and AI Overviews cite different sources about 87% of the time, so the two are tracked as separate surfaces.
AI Overviews
Also called AIO, Google AI Overviews
The AI-written summary Google shows above some search results, with links to pages it drew on. Google says a page needs to be indexed and eligible for a search snippet to appear, and that no special schema or file is required.
AI visibility
How often, and how favorably, AI answers bring up a brand for the questions its buyers ask. Because answers change from run to run, it is measured as rates across many sampled answers rather than as a single position.
Answer-first
Writing a page or section so its opening sentence directly answers the question, before background or caveats. In one vendor dataset of 18,012 ChatGPT citations (Growth Memo, Feb 2026), 44.2% of cited text came from the first 30% of a page; there was no base rate, so treat the effect as plausible, not proven.
API vs app
The gap between answers an engine gives through its developer API with web search on and the answers people see in its consumer app. An Oct 2026 paper (arXiv 2610.15382) found the ChatGPT API's sources covered only about 23% of the pages the app cited, so every number should say which surface it came from.
Buyer question
Also called Tracked prompt
A question a client's real customer would plausibly ask an AI assistant before buying, such as "best payroll software for a 30-person company". A tracked set mixes branded and non-branded questions and stays fixed for a given market and language, so rates can be compared week to week.
Citation
A link an AI answer attaches to its text as the source for what it says. Citations churn quickly: in Ahrefs' 20-day study (Oct 2026), about 60% of ChatGPT and AI Mode citations were gone the next day, and about three quarters of those that dropped out returned within the 20 days.
Citation coverage
Also called Citation rate
The share of sampled answers that link to at least one of the brand's own pages as a source. A brand can be named without being linked, and linked without being named, which is why citation coverage and mention rate are reported side by side.
Confidence range
Also called Confidence interval, Likely range
The span a measured rate probably falls in, given how many answers it rests on; Optimz shows it as "likely between" two numbers, which holds the true value 95 times out of 100. With only 10 answers, a 50% mention rate could plausibly sit anywhere from about 24% to 76%.
Consulted source
Also called Retrieved but not cited
A page the engine retrieved or read while building an answer but did not cite. The OpenAI and Anthropic APIs can return these pages, which separates "not found" (an access or indexing problem) from "found but not cited" (a content problem).
Control group
Questions a fix did not touch, measured on the same engines over the same weeks, so drift across the whole platform can be subtracted from the result. In Ahrefs' schema study (May 2026), a raw gain of about 43% on AI Mode shrank to about 2% and was no longer a real change once control pages were compared.
Detectable change
Also called Minimum detectable effect, MDE
The smallest real change a measurement can reliably pick up, given its number of questions, engines and samples. Working it out before a fix ships sets honest expectations: a small test may only be able to see large swings within a month.
Difference test
A statistical test on the change between two rates, used to decide whether a week-over-week or before-and-after move is real or noise. It is more sensitive than checking whether two confidence ranges overlap, a common shortcut that reads real changes as "no change".
Fan-out
Also called Query fan-out, Fan-out queries
The separate web searches an engine runs behind a single question, often reworded with terms like "2026", "best" or "vs". Ahrefs found fan-outs in 14.4% of 66.6 million logged-out free ChatGPT responses (Oct 2026), about 1.6 per response, while Google says AI Mode issues "a multitude of queries" at once.
GEO
Also called Generative engine optimization
Another name for AEO, taken from a 2024 research paper (KDD 2024) that tested page edits against AI-generated answers. A 2026 re-run of its tactics, such as adding quotations or statistics, on modern engines found gains indistinguishable from zero or negative (a vendor-affiliated preprint).
Grounding
Feeding live search results to a model at answer time, so the answer rests on retrieved pages rather than on memory alone. Google's Google-Extended robots.txt token controls whether a site's content can ground Gemini Apps and the Vertex AI API, separately from Google Search.
How sure is this number?
Also called Reliable, Good estimate, Early estimate, Collecting
Every Optimz number says how many answers it is based on and how sure it is. Reliable: the true value is likely within 5 points. Good estimate: within 10. Early estimate: within 20, so more answers will firm it up. Collecting: too few answers to say yet. A change marked "real change" passed a direct test on the difference; one marked "too small to call yet" could still be chance.
Known vs found
Also called Parametric vs retrieved visibility
The split between a brand an engine names from memory (known) and one it names only after searching the web (found). Asking the same questions with search off and on shows which is missing, and the fixes differ: off-site footprint for known, crawlable answer pages for found.
llms.txt
An optional Markdown file at /llms.txt that lists a site's key pages for language models. Ahrefs found 97% of the llms.txt files across 137,210 domains received no requests in May 2026, and Google says its AI features need no special files, so it is not a visibility lever.
Mention rate
The share of sampled answers that name the brand at all, for a set of questions on a given engine. It is the broadest measure of presence and should always carry its sample size and confidence range.
Parametric knowledge
Also called Training-data knowledge
What a model knows from its training data, without searching. A brand gets there by being written about across many sources before the model's training cutoff, so this knowledge only changes when a new model is released.
Paraphrase group
Also called Intent cluster
Three to five wordings of one buyer intent, tracked and reported together. A May 2026 paper (arXiv 2605.27440) found that rewording a question changed the recommended brands more than re-asking it did, so a single wording gives a biased reading.
Recommendation rate
The share of sampled answers that actively suggest the brand for what was asked, not just name it in passing. Deciding whether an answer recommends is a judgement on each answer, so this rate is stricter than mention rate and harder to classify consistently.
Retrieval
The step where an engine searches an index, picks candidate pages and passes short passages from them to the model. A page that is blocked, unindexed or empty in its server HTML can't be retrieved, however good its content is.
Sample size
Also called n
The number of answers behind a rate. Below about 20 answers a rate is too uncertain to act on, which is why Optimz shows such cells as "Collecting" instead of a percent.
Snippet window
The roughly 202 characters at the start of a page's rendered body that ChatGPT's own search index keeps as the page's snippet, per RESONEO's 2026 tests. It is not the meta description, so a kicker, date or breadcrumb placed before the H1 uses up the window.