The AI search visibility guide

What this guide answers

How do you get an AI assistant to name your company when someone asks it which tool to use?

That question has a real answer, and the answer is not one thing. It is four conditions that have to hold at once, each with its own failure mode, its own owner inside a company, and its own way of being measured. A page can be perfectly written and never fetched. It can be fetched daily and never retrieved. It can be retrieved and still contribute no sentence to the answer, because nothing in it survives being pulled out of context. And a brand can do all three well on its own domain and still lose, because most of what a model says about a category is assembled from pages the brand does not own.

The four conditions are:

1. Access - the right crawler can fetch the page at all.
2. Retrieval - the page is selected for the specific question being asked.
3. Extraction - a passage on the page can be lifted out and used as answer text.
4. Corroboration - the same claim appears somewhere the model already reads.

Most published advice conflates all four, which is why so much of it cannot be falsified. Split apart, each one becomes testable, and most of them become boring. That is the point. This guide is seven chapters, one per problem, plus one on the things that do not work.

The definition, precisely

AI search visibility is whether a brand or page is retrieved, mentioned and linked inside an assistant's generated answer. It is not a ranking position, and treating it as one produces the wrong programme.

A ranking is a list. An answer is a paragraph with a handful of sources hanging off it, chosen by a model after a retrieval step you cannot observe. Google describes its own AI features as sometimes using a "query fan-out" technique - issuing multiple related searches across subtopics and data sources to build one response.[^google-fan-out] The user asked one question. The system asked several. Your page competes not for the query the human typed but for one of the sub-queries the system generated from it, and you will never see that list.

That single mechanism explains most of the confusion in this category. It is why a page can rank first and go uncited, why a page nobody finds in Google turns up in an AI Overview, and why "what is our rank" stopped being the question.

What the research actually shows

The most-cited academic work here is GEO: Generative Engine Optimization (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande), published at KDD 2024. It matters because it ran an experiment rather than an argument: nine content modifications applied to real source pages, scored on two visibility metrics, over GEO-bench - 10,000 queries drawn from nine sources and tagged by domain, difficulty and intent.[^geo-bench]

The headline result is narrow and useful.

Read those two rows together. The thing classical SEO taught a generation of marketers to do - put the query terms in the copy more often - is the thing the study found does not transfer. What transfers is what a competent analyst does anyway: quote a source, give a number, say where the number came from.

There is a second finding that gets quoted less and matters more if you are not already the category leader. The paper breaks results out by the source's existing rank in the underlying search engine, and the gains concentrate at the bottom: Cite Sources improved visibility by 115.1% for fifth-ranked sources and reduced it by 30.3% for first-ranked ones.[^geo-lower-ranked] The authors state it plainly - the method is especially helpful for lower-ranked websites. If you already own the top result, restructuring the page is not your lever. If you do not, it is the cheapest one available.

Two caveats, since this guide's entire argument is that sourcing beats assertion. The study is from 2024, and generative engines have changed since. And its metrics measure prominence inside an answer, not clicks, revenue, or whether a buyer remembers you. Anyone quoting "GEO increases visibility by 40%" as a business outcome has quoted a benchmark as though it were a P&L.

Access is not one decision

The operators do not run one crawler each. They run several, with different purposes, different robots.txt tokens, and different documented consequences when you block them.

This table is why "should we block AI crawlers?" is the wrong question. The real questions are three: do we want to be in the answer (allow the search agent), do we want to be in the training set (a different agent), and do we want a person who explicitly asks an assistant about us to be able to fetch our page (a third agent again). Anthropic's documentation states the trade-off for each bot individually, including that disabling its search agent may reduce a site's visibility in search results. A single User-agent: * disallow answers all three questions at once, and rarely the way the person who wrote it intended.

Chapter one works through the robots.txt patterns operator by operator. Our crawler log feature exists because the only way to know which of these agents actually visits you is to look at the requests, and the robots.txt generator exists because writing the rules by hand is where the mistakes happen.

Where the sentences come from

Here is the finding that reorders most AI visibility programmes.

AirOps analysed more than 500 commercial-intent queries across GPT-5, Claude Sonnet 4.5 and Perplexity Sonar, capturing 21,311 brand mentions, and reported that 85% of brand mentions came from external domains and 13.2% from the brand's own domain.[^airops-third-party-share] That is a vendor-published figure, not an operator-published one, and it deserves the usual discount - but the methodology is stated, the sample is large, and the direction matches what anyone who has read a hundred AI answers about a software category already suspects. The model is mostly reading review sites, comparison articles, forum threads and documentation. It is mostly not reading your homepage.

The operational consequence is unglamorous. If roughly one mention in seven comes from your own domain, a programme that only ever edits your own domain is working on one-seventh of the surface. The other six-sevenths is a different job: being listed, being compared, being reviewed, and being described accurately in places you do not control. Chapter four is about that job, and about which parts of it are legitimate and which are the 2026 version of buying links.

What Google says, in Google's words

Google publishes a site-owner guide for AI features that is more useful than most of the commentary written about it, mainly because it rules things out.

• No new file formats. "You don't need to create new machine readable files,

AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add."[^google-no-special-markup]

• The eligibility bar is the ordinary one. A page must be indexed and eligible

to be shown in Search with a snippet; there are no additional technical requirements.[^google-snippet-eligibility]

• Measurement is folded in, not broken out. Sites appearing in AI features are

counted inside overall search traffic in Search Console, in the Web search type.[^google-search-console-reporting]

The first bullet is the one to keep. An entire cottage industry sells llms.txt files as an AI-visibility measure. No major operator currently documents that file as a retrieval input, and Google explicitly says no such file is needed. It costs ten minutes to publish and there is no harm in having one; there is real harm in a plan whose first three items are all of that kind. Chapter seven catalogues the rest, including the mistakes in our own history.

Measuring it without lying to yourself

Everything above is untestable unless you can answer "are we cited more than we were last month?" honestly - and that is harder than it sounds, because the system is non-deterministic. Ask the same question twice and you can get two different answers with two different sources. That is not noise to be eliminated. It is a property of the thing being measured.

So every number needs three companions to mean anything: the prompt set it was measured over, the sampling depth - how many times each prompt was asked - and the cadence. A citation rate with no sample size is not a measurement, it is a mood. Attensira publishes its own rule for this: a rate carries the sample size behind it, a null value means "never measured" rather than "measured zero", and a per-platform metric records whether that platform was queried at all, because reporting an untracked platform as 0% invents a failure that did not happen.[^attensira-reporting-numbers] You do not need our product to adopt the discipline. You need it written down before the first number is reported, because after that it becomes negotiable.

Chapter six covers the honest version: what to sample, what a defensible sample size looks like, why Search Console answers only the Google part of the question, and which changes you should not expect to be able to detect at all.

How long any of this takes

No operator publishes a propagation time, and any specific number you see is someone quoting their own sample. What is documented is the dependency. Google says crawling can take anywhere from several days to several months, depending on how often its systems determine a page needs refreshing.[^google-recrawl-latency] Nothing you publish can be retrieved before it is fetched, and nothing you fix off-domain counts until the page carrying the fix is itself crawled.

Plan in quarters, not sprints. Measure weekly so you can see drift, but judge on a horizon long enough that a single model update does not read as a result.

The four conditions, in operational terms

Split apart, each condition has a different owner, a different symptom when it fails, and a different instrument. This table is the guide in one screen.

The reason this decomposition earns its space: in practice the first and last conditions have no owner at all in most companies. Access sits with engineering, who were never told it was a marketing input. Corroboration sits nowhere - the directory entries and review profiles that describe you were created by whoever happened to sign up, years ago, and have not been read since. The middle two conditions get all the attention because they are the two a content team can act on without asking anyone.

If you do one thing after reading this pillar, assign the first and the last.

What actually changes versus classic SEO

The overlap is larger than the vendors selling a new discipline imply, and the differences are specific rather than philosophical.

The unit of work changes from keyword to question. A keyword list ranked by volume is a poor input when the system decomposes a question into sub-questions you cannot see. The useful artefact is a prompt set: the twenty questions a real buyer asks, in their words, including the ones where you are not named.

The unit of measurement changes from position to rate. There is no position. There is "cited in 6 of 20 runs", which is a poll, and it needs a sample size attached every single time.

The surface expands past your domain. Classic SEO is played almost entirely on pages you control, with links as the off-domain component. Here, the pages describing you are themselves the competitor set, and correcting a stale directory entry can matter more than publishing a new page.

The winning content edit changes. Keyword coverage and persuasive voice were the SEO staples. The measured winners in generative answers are sourced statistics, quotations and citations - substance, not framing.

The technical failure modes move. Not sitemaps and canonicals, but per-agent robots.txt tokens, CDN bot rules, and whether the answer survives without JavaScript.

What does not change: a site that cannot be crawled, indexed and rendered will not be cited, and Google states plainly that its ordinary SEO fundamentals still apply to its AI features. Anyone selling GEO as a replacement for technical hygiene is selling the second storey of a building with no ground floor.

A sequence that does not waste the first quarter

Order matters, because most of these steps are only measurable once the earlier ones are done.

1. Week one - access. Read your logs per AI user agent. Check robots.txt and

the CDN bot rules. Fetch a key page with JavaScript disabled. Fix what is broken. This is the only step that can silently invalidate everything after it.

1. Week one - the prompt set. Write the twenty questions. Include the

unflattering ones. Do not skip the ones where your brand is not mentioned in the prompt, because those are the ones a real buyer asks first.

1. Weeks two to four - baseline. Sample the prompt set repeatedly across the

platforms you care about. Record mention, citation, recommendation, and every other URL cited. Publish the baseline with its sample sizes before you change anything, so nobody can later relitigate what the starting point was.

1. Month two - the source list. Rank the third-party domains cited across your

prompt set by frequency, read what they say about you, and fix the errors. This is usually the largest single gap and the cheapest to close.

1. Month two - the extraction pass. On the ten pages closest to the questions

you care about: lead with the answer, add sourced numbers, convert the biggest comparison paragraph to a table, put a real date on it.

1. Month three - re-measure and judge. Same prompt set, same depth. Report the

change with the denominator, and report "not distinguishable from variation" when that is the truth - which it often will be, and saying so is what makes the other reports believable.

Nothing in that sequence requires a tool. It requires someone to own steps one and four, which is the part that usually does not happen.

Read the guide in order

Each chapter answers one question and stands alone, but the order is the argument.

1. [How AI assistants find your pages](/guides/ai-search-visibility-guide/how-ai-assistants-find-your-pages) - the crawlers, per operator, and what blocking each one actually costs.
2. [Ranking and being cited are not the same thing](/guides/ai-search-visibility-guide/ranking-versus-being-cited) - retrieval, fan-out, and why rank is a weak predictor of citation.
3. [Passages that survive extraction](/guides/ai-search-visibility-guide/passages-that-survive-extraction) - the page structures that come out of a document intact.
4. [Why most of your citations are not on your domain](/guides/ai-search-visibility-guide/third-party-presence) - the six-sevenths, and the honest way to work on it.
5. [Schema markup and technical access](/guides/ai-search-visibility-guide/schema-and-technical-access) - what structured data does and does not buy you.
6. [Measuring AI visibility honestly](/guides/ai-search-visibility-guide/measuring-ai-visibility) - sampling a non-deterministic system without fooling yourself.
7. [What does not work](/guides/ai-search-visibility-guide/what-does-not-work) - the tactics with no evidence behind them, ours included.

Where the platforms differ, and where they do not

Treating "AI search" as one destination is the second most common planning error after treating it as a ranking. The differences that actually change what you do are few, and they are documented.

Google is one crawler and one control. There is no separate AI Overviews crawler: robots.txt for Googlebot governs Search, and AI features are part of Search. Your lever is snippet control, and restricting your snippet restricts your AI presence by the same action. Anything you do for classic Google SEO is work on your AI Overviews presence too, which is why Google-only programmes rarely need a separate budget line.

OpenAI, Anthropic and Perplexity split the decision three ways. Training, search indexing, and user-initiated fetching are separate agents with separate tokens. This is where the genuinely new technical work lives, because a one-size-fits-all robots.txt makes three decisions with one line and gets at least one of them wrong.

User-initiated fetching behaves differently everywhere. Perplexity says its user fetcher generally ignores robots.txt because a person asked. Anthropic says disabling its user agent stops retrieval in response to a user query. Same intention, different outcomes - which means a policy has to be written per operator, not as a principle.

What does not differ: every one of them needs to be able to fetch the page, read the answer without executing your JavaScript, and find a passage worth quoting. The unglamorous bulk of this work is identical across platforms, and it is the part most programmes skip in favour of the differences.

The practical implication for a small team: do the shared majority once, well, and pick one or two platforms to measure properly rather than sampling five badly. A citation rate on five platforms at a depth that cannot detect change is five numbers you cannot act on.

What we do not know

A guide claiming certainty about a closed, fast-moving system two years old would be lying, so here is the boundary.

We do not know the retrieval mechanics of any major assistant. None publish their ranking function, their index composition, or how a candidate document becomes a citation. We do not know how quickly a change propagates beyond Google's days-to-months statement about recrawling. We do not know whether the effects measured in the 2024 GEO study still hold at the same magnitudes on 2026 models, because no replication at that scale has been published. And we do not know the true third-party share of citations in your category - the 85% figure above is one vendor's sample of commercial software queries, and your industry may sit somewhere else entirely.

What is knowable is what the operators document, what a controlled experiment found, and what your own server logs and your own sampled answers say about your own brand. This guide is built from those three and nothing else. Where a chapter states something we measured, it says how. Where it states something a vendor measured, it names the vendor and the sample.