The working vocabulary of answer engine optimisation. Each term gets one entry, each entry opens with a single self-contained sentence, and the qualification that follows never smuggles in a new claim.
46 terms · 5 families · definitions only, no figures
The terms are grouped by the job they do rather than alphabetically, because the order they are usually needed in is: what this is called, what an engine produces, what decides whether you are in it, what you build so it can be read, and how any of it is measured.
Definitions here explain mechanism, so no figure appears on this page and none is needed. Where a term has a longer treatment, the entry links to it.
Answer Engine Optimization is the practice of getting your brand named and cited in the answers AI engines generate when a buyer asks them to recommend a provider.
The unit of success is citation share across prompts and engines, not the rank of a single link.
Generative Engine Optimization is the practice of optimizing to be named and cited in AI-generated answers, used interchangeably with AEO and covering the same underlying work.
The two terms describe one discipline. Which you use is preference, not method.
An answer engine is an AI system that responds to a question with a synthesized recommendation in prose, naming a few providers, rather than returning a page of links to scroll.
ChatGPT, Perplexity, Gemini and Google's AI Overviews are answer engines.
AI search optimization is the same discipline as AEO and GEO under a third name: making a brand retrievable and quotable so AI systems name it when a buyer asks for a recommendation.
The three labels arrived from different places. AEO came from the question-answering side, GEO from work on generative retrieval, AI search optimization from search marketers describing what changed in their own channel. None of them denotes a different method, so the choice between them is a decision about which words your buyer already types. It goes wrong when a provider sells the names as three services. Ask what differs in the work, and if nothing does, you are looking at page inventory rather than an offering.
LLM SEO is the informal name for optimising to be cited by large language models, borrowed from search vocabulary because that is the vocabulary buyers already have.
The borrowed word carries a wrong assumption with it. Ranking does matter, because the index an engine grounds against is largely a conventional search index and rank governs whether you enter the candidate pool at all. It stops mattering at the second step, where the engine picks which retrieved passage to quote and structure decides that rather than position. Reading the whole problem as ranking, which the name invites, buys authority work for what is often an extractability problem.
An AEO agency is a firm that measures how often AI engines name a brand, diagnoses why they do not, and ships the technical, content and entity work that changes it.
Which engines a firm claims to track separates nobody, since every dashboard claims all of them. The question that does is whether the firm will state its engine coverage including the gaps, the number of runs behind a figure, the window in dates, and the point below which it calls a movement noise rather than progress. A firm that answers those can be checked. One that will not is selling a chart, and the test applies to us exactly as it applies to anyone selling against us.
A citation is one instance of an AI engine naming or referencing your brand inside the answer it generates.
Citations are earned as a byproduct of being retrievable, authoritative and extractable. They are not bought directly.
An AI Overview is Google's AI-generated answer shown above the traditional results, synthesizing a response and citing a handful of sources.
Appearing in the overview is a separate goal from ranking a blue link beneath it, and the two move independently.
Zero-click describes a result where the engine answers the question itself and the reader never clicks through to a source.
The payoff is brand influence. An AI names you, and the buyer searches for you by name later, with no referral click in between.
Retrieval is the step where an answer engine fetches candidate sources from an underlying index before it synthesizes an answer.
If your content is not retrievable from the index a given engine draws on, it cannot be cited, however good it is.
Extractability is whether a page that an engine has already retrieved is written so a passage can be lifted from it: chunked cleanly, with the format matching the shape of the answer.
It is the second of two gates and has the opposite fix to the first. Authority decides whether you enter the pool of candidate sources, extractability decides whether you are chosen from it.
An entity is the structured, machine-resolvable identity of an organization, person or thing: a consistent name, attributes and credentials an engine can pin down and trust.
A coherent entity reads as authoritative. A thin or contradictory one reads as risky, and a cautious engine names someone else.
A knowledge graph is a structured network of entities and the relationships between them that engines use to work out who you are and how you connect to things they already know.
Being a well-formed node in that graph, consistently described and corroborated elsewhere, is what lets an engine cite you with confidence.
Third-party consensus is the agreement of independent sources, reviews, directories, editorial mentions and peer references, that an engine treats as corroboration it can repeat.
Your own site can claim you are excellent. An independent source saying it is what an engine is comfortable echoing.
Retrieval-augmented generation is the pattern behind every answer engine: the system searches an index for relevant documents, then writes its answer conditioned on what it retrieved rather than on memory alone.
This is why AEO is possible at all. Fixed model weights cannot be edited from outside, but the retrieval step in front of them reads live documents, and those you control. The common error is treating retrieval as a separate AI channel with its own rules. The index being searched is mostly a conventional search index, so what decides whether you are found in it is what always did. Only the selection step afterwards is new.
Chunking is the splitting of a page into passage-sized pieces before indexing, so an engine retrieves and quotes a fragment of your page rather than the page as a whole.
What competes for a citation is therefore a passage, not a document, and a passage is read with none of the context that surrounded it. An argument built carefully across four paragraphs can be the better one and still lose to a worse but self-contained rival. Writers assume the boundaries are theirs to set; they are not, and the engine cuts where it cuts. What you control is whether a piece survives being cut, which means each section answering its own heading without leaning on the sentence before it.
Passage retrieval is the step where an engine scores individual passages rather than whole pages and carries only the highest-scoring few into the context it writes from.
Page-level and passage-level outcomes come apart because of it. A page can be retrieved on its authority and still contribute nothing, because none of its passages scored well enough to travel. That mechanism produces the most confusing pattern in this work: visible in classic search, absent from the answer sitting above it. The fix there is structural, and rewriting for authority when the problem is passage competition spends a quarter on the wrong gate.
Grounding is the constraint that ties a generated answer to retrieved sources, so the claims in it trace back to documents rather than being produced from the model's weights.
It is what makes a citation mean anything, and its absence is what produces confident invention. Verifying it is harder than vendors imply. Of the engines we read as ground truth, one returns no list of sources alongside its answer at all, so grounding there is derived from the text rather than confirmed against a source set, and we describe it that way instead of calling it verified. When a provider says an answer was grounded, the useful follow-up is how they know.
Co-occurrence is how often your brand appears alongside a category, a service or a competitor inside the same document, which is part of how a model learns that you belong in that set.
A roundup, a comparison or an industry list places you next to firms an engine already associates with the category, and repetition of that pattern across independent documents is what makes the association hold. Running it as an on-page tactic is the mistake. Writing competitor names into your own copy teaches an engine nothing, because the documents that carry weight are the ones you did not publish. Earning the mention is the work, and it accumulates slowly.
A fan-out query is one of the several searches an engine issues internally after rewriting a single user question, with the answer assembled from the union of what all of them returned.
One typed question becomes a handful of narrower retrievals, and a source that wins any one of them can reach the answer without ever matching the question as asked. That reframes what you are optimising for. Buyers never type the sub-questions, so a keyword list built from what buyers type misses most of the surface being searched. Where an engine exposes its fan-out, reading it is the cheapest available look at what was actually asked on your behalf.
Schema markup is structured data, usually written in JSON-LD using the schema.org vocabulary, that labels your content so machines read your facts directly instead of inferring them from prose.
Markup represents what a visitor can see. An answer that appears only in the markup and nowhere in the visible copy is a guideline breach, not a shortcut.
llms.txt is a concise, machine-readable file at the root of a domain that points AI assistants at a site's most important pages and states its key facts in curated form.
It is an emerging convention, not a confirmed ranking input. What it does reliably is remove ambiguity by handing a model the authoritative version of your facts.
The agent-ready layer is the set of machine-readable affordances, llms.txt and structured data among them, that let an AI assistant acting for a user cite you accurately rather than guess.
Its payoff is accuracy. Clean facts in a machine-readable place leave an agent no reason to invent a wrong one.
Answer-first content opens with a direct, self-contained answer to a real question and supports it afterwards, which is the shape an engine can lift cleanly into a response.
Question-shaped headings, a summary of takeaways and FAQ structure all make expertise extractable rather than buried.
sameAs is the schema.org property that links your entity to the profiles describing it elsewhere, letting a machine resolve those separate records to one organisation.
Identity resolution is the whole job. A firm described consistently across its own markup, its listings, its registry entries and its public profiles reads as one confident node; the same firm described inconsistently reads as several uncertain ones. Every URL in the property has to be verified before it ships, because a guessed profile that turns out to belong to a different company teaches a resolver the wrong thing and is worse than leaving the property out. Entity hygiene, not a lever.
GPTBot is OpenAI's crawler, declared in robots.txt under that user agent, and disallowing it removes your pages from the corpus it collects.
OpenAI runs more than one agent and they do different jobs, so a blanket rule can produce an outcome nobody chose: blocking the crawler that gathers content while leaving the one that fetches pages for a live answer, or the reverse. Check which agents your robots file actually names before concluding you have made a decision. Whether to allow a training crawler is a genuine judgement with arguments on both sides. Making it by accident, in a file nobody has opened in a year, is not.
ClaudeBot is Anthropic's crawler, and it is the user agent to allow or disallow in robots.txt if you want a position on Claude reading your site.
Access and measurement are separate questions, and the second has an uncomfortable answer. No ground-truth path to Claude's consumer product exists for anybody, ourselves included, so allowing the crawler is a decision whose effect you cannot verify. A vendor reporting Claude citation share is reporting an API surface relabelled. We would rather record the coverage as missing than fill the hole with a number from somewhere else, and the crawler decision stands on its own merits meanwhile.
PerplexityBot is Perplexity's crawler, named in robots.txt like any other, and it governs whether your pages are available to the index its answers draw on.
Its retrieval behaviour raises the stakes on that rule, because answers frequently cite pages fetched at query time rather than months earlier. The reverse direction matters too: Perplexity's own robots file disallows automated access to its search paths, which is why we do not scrape it, and technical capability is never permission. Reading a public robots file in both directions, theirs and yours, is the shortest check in this glossary and it settles arguments people otherwise have for months.
JSON-LD is the syntax structured data is usually written in: a block of JSON inside the page that describes your facts as linked data, kept separate from the HTML that renders them.
Separating the description from the template is why it won. One script tag can carry an organisation, its services and its identifiers without anybody touching a layout, which is also why the graph drifts away from the page it describes. The format is not what gets judged. A perfectly valid block full of claims that appear nowhere in the visible copy is worse than no block at all, because it validates cleanly and still breaches the rule that markup represents what a visitor can see.
Structured data is the umbrella term for machine-readable descriptions of a page's facts, of which schema.org is the vocabulary and JSON-LD the usual syntax.
Three words get used as though they named one thing. The vocabulary decides what you are allowed to say, the syntax decides how you write it down, and structured data is the category both sit inside. On what it buys you: we ship it as entity hygiene, we test it as an experiment, and we do not bill it to a client as a proven citation driver. No controlled public test exists that would justify that, ours included, and saying so costs less than being caught.
Crawlability is whether a machine that is allowed to fetch your pages can also read them, which is a different question from whether your robots file permits the fetch.
Permission and legibility fail separately. A page can be fully allowed and still arrive as an empty shell, because its content is assembled by JavaScript after load and a crawler that does not execute scripts sees nothing worth keeping. Firewall rules, bot management and CDN protections add a third failure that never appears in robots.txt at all. What settles it is fetching your own page as the agent in question and reading what comes back, rather than confirming a rule permits something and assuming the rest.
Raw HTML is what the server returns; rendered HTML is what exists after JavaScript has run, and the gap between them is content only a client executing scripts can see.
Assume the two are identical and every conclusion from one applies wrongly to the other. Text a browser shows a human can be absent from the raw response an engine parsed, so a check run against the wrong version reports content as missing when it is present, or present when it is not. This is also the most common route to a false finding: absence according to your instrument is a claim about your instrument until a second one agrees.
A prompt set is the fixed collection of buyer-style questions run across engines on a schedule to measure how often, and where, your brand is named.
Every figure a prompt set produces is pinned to that cohort, so the set is stated alongside the number and changes to it are treated as a new baseline.
Citation share is how often your brand is named across a representative set of prompts and engines, benchmarked against the competitors named instead of you.
It is the headline metric of AEO, read against firms you can name rather than estimated from traffic.
Share of voice, in AEO, is your portion of every brand mention an engine makes across a category's prompts: your visibility relative to the full competitive set.
It frames citation share as a split. Of every name the engine could give, how many are yours.
AI visibility is how often an engine names your brand across the prompt runs you measured, expressed as a rate over runs rather than as a count of prompts you have ever appeared in.
The phrase is used loosely enough to be worthless without its definition attached, so ours is fixed: visibility is a per-run rate, and the cumulative union of prompts where a brand appeared at least once is called reach and is never labelled visibility. We shipped that confusion in our own product before catching it. The reason it matters is arithmetic: a union can only rise as runs accumulate, so a product reporting one under the visibility label improves every month whether or not anything did.
Reach is the union of prompts where a brand was named at least once across every run, measuring the breadth of questions where you can appear rather than how often you do.
Under its own name it answers a real question: which parts of a category are open to you at all. What it cannot do is trend. Each additional run offers another chance of a hit and no chance of a miss, so the figure climbs as a function of how long the tool has been running. Read a rate to see whether work moved something, read reach to see where the ceiling sits, and treat a product that reports the union as visibility as one whose chart cannot go down.
Appearance rate is the share of a fixed set of tracked questions where a brand was named at least once, which makes it a union expressed against a stated denominator.
That denominator does two jobs. It fixes the meaning, since a rate over one set of questions is not comparable to a rate over another, and it fixes the granularity, because the only values available are multiples of one divided by the question count. Which gives a published table a property any reader can check without access to the data. A figure landing between two of those steps is not a smaller number, it is a number the stated method cannot produce, and it is fair to ask about.
A measurement window is the pair of dates a figure was captured between, printed on the figure itself rather than implied by the month the report was sent.
Answers move week to week, so a number without dates cannot be compared with anything, including its own earlier version. The window also has to sit inside a frozen prompt set. A figure computed against today's prompt list while scanning last month's runs is a rewrite of history rather than a trend, which is why our metrics are stamped to the cohort that was live when the run executed. A named month is the usual tell. Ask for dates, then ask what the set was on each of them.
Run-to-run variance is the spread you get by asking an engine the same question repeatedly, and it is wide enough that a single run cannot establish whether you are cited for that question.
Nothing is fixed between runs. The same prompt can name a different set of companies an hour later, and AI Overview presence itself comes and goes across passes on identical queries. So the number of runs behind a figure is a required disclosure, not a technical footnote. A screenshot is the standard failure: one capture showing a competitor named and you absent is a sample of one, which is why we hold any citation figure that has not run often enough to separate a reading from noise.
Ground truth, in AI visibility measurement, means reading the consumer product a buyer actually uses instead of calling that vendor's API and reporting what came back.
An API with web search switched on is a different surface from the app. It retrieves differently, formats differently and cites differently, so a figure taken from one and labelled AI visibility describes something no buyer will ever see. Our sweeps filter model-level API runs out structurally rather than by configuration, so nobody can quietly re-enable one. The cost of that rule is honest and worth stating: it is why some engines have no coverage at all, and naming which ones is part of publishing a method.
Citation rate is the fraction of measured prompt runs in which your brand was named, counting only your own presence and leaving competitors out of the denominator.
It differs from share of voice in what it counts against. Citation rate asks how often you appeared across the opportunities measured; share of voice asks what portion of every company name produced was yours. The two move independently, so a brand can hold a strong rate on a narrow set of questions while its share of the naming stays small, which means the shortlist a buyer reads still has four other firms on it. Neither figure is wrong. Reporting one without the other is.
Engine coverage is the list of engines a measurement actually reads, published together with the engines it does not read and the reason each one is missing.
Omitting it lets a coverage claim quietly expand to whatever a reader assumes. Ours: ChatGPT and Gemini through ground-truth paths, plus Google AI Overviews. Claude is not covered, because no ground-truth path to it exists and an API is not the product. Perplexity is not covered. Technical reach is not permission either, since several consumer surfaces disallow automated access in their own robots files, and we do not read what we have been told not to.
An AEO audit is a diagnosis of why AI engines name someone else, run against a frozen prompt set for your category and benchmarked against competitors you can name.
Its output is a decision about which of two gates you are failing, because retrieval problems and extractability problems take opposite fixes and buying the wrong one costs a quarter. What disqualifies an audit is not its length. It is a single run, an unnamed prompt set, an unstated window, or a composite score standing in for a diagnosis, since fusing two problems with opposite remedies into one figure destroys the only information the exercise was for.
An AI visibility report is the deliverable stating, for one brand, how often named engines cite it against named competitors, with the prompt set, run count and capture window printed beside every figure.
Structure carries the credibility here. A reader who can see the questions asked, how many times each ran and the dates can repeat the exercise, and a figure that can be repeated is a different class of object from one that cannot. Read the limits section first when one arrives. A report that never says what its instrument could not see has either measured nothing awkward or has not gone looking.
A citation gap is the set of prompts where an engine names competitors and does not name you, read off the same frozen prompt set that produced your citation figures.
It is more useful than the headline figure because it is specific: these questions, these engines, these firms named instead. Worked backwards, it shows whether you were absent from the sources behind the answer entirely or present in them and passed over. Treating the gap as a content to-do list is the standard error. Much of it is not a content problem, and writing pages against prompts you never entered the candidate pool for changes nothing.
Most of the vocabulary above resolves to one question: which of the two gates are you failing. A single visibility score cannot tell them apart, which is why a score on its own is not a diagnosis, and why work chosen from one can be aimed at the wrong gate for months.
Citation share and share of voice are not estimated from traffic. They come from a prompt set written once for a category and then frozen, run against every engine in clean sessions, with each answer stored unmodified so any figure can be traced back to the answer that produced it.
A set that moves cannot produce a trend. Every figure stays pinned to the cohort of prompts that produced it, and a change to the set is treated as a new baseline rather than as movement.
A definition is not evidence about your category. The disclosure sits on the page rather than in a footnote, on this page as on every other.
The audit runs a frozen prompt set for the questions your buyers actually ask, benchmarked against competitors you name, and the findings are explained on a call rather than emailed. You keep the report and the underlying data whatever you decide afterwards.