The AI search visibility funnel: eight places a company drops out of AI answers

The engine has to be able to fetch your pages, find them when it searches, trust them, describe you correctly and choose you over alternatives; the answer has to send the reader on; your page has to state the answer in text a reader and an engine can use; and the visitor has to convert. Each stage has a different fix, and fixing the wrong one changes nothing. This page names the eight stages, the symptom each produces, how to check it and what to change.

Why is "we don't show up in ChatGPT" not a diagnosis?

"We don't show up in ChatGPT" describes an outcome, and at least eight different failures produce that same outcome. It is the AI search version of "sales are down": true, worth knowing, and no guide to what to do on Monday.

A SaaS company that blocks OpenAI's search crawler by accident, a company whose pages rank well but never get read for the question buyers ask, and a company the engine knows well but describes as the thing it sold two years ago all report the same symptom. The first needs a one-line change to a file. The second needs different pages. The third needs other websites to say something new about it. Treating all three with "more content" helps none of them.

Does every AI answer go through these stages?

Not every answer does, because an assistant can also answer from what its model learned in training, and then the first three stages do not apply to that answer. What counts for a memory answer is what the web said about you before the model's training data was collected, and nothing you change today reaches a model that is already trained.

The vendors document the two paths as separate crawlers. OpenAI says GPTBot collects content that may be used to train its foundation models, while OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features", and that "each setting is independent of the others". Anthropic says restricting ClaudeBot signals that a site's "future materials" should be excluded from training, and documents Claude-SearchBot separately for search. So a block on a training crawler affects future models, and a block on a search crawler affects answers that search today.

This matters for diagnosis. If an engine describes your company as it was two years ago and cites no sources, you are probably reading a memory answer, and stage 3 is the place to work. If the answer cites pages, it searched, and every stage below applies. The eight stages are the path a searched answer takes, followed by what happens after the reader sees it.

What are the eight stages of the AI search visibility funnel?

The eight stages run from whether an engine can fetch your site to whether the visitor it sends you becomes a customer.

Stage 6 is a property of your pages and affects stages 1, 2 and 4. We list it near the end because it is the stage most companies start with, and it is rarely the first thing that is broken.

Stage 0: can AI search engines fetch your pages at all?

Every engine that searches the web needs to be allowed to fetch your pages, and each engine sets its own rules. On Google, AI Overviews and AI Mode have no separate crawler or file: Google says a page must be "indexed and eligible to be shown in Google Search with a snippet", and its July 2026 guide adds that "a site must be included in Search generative AI features in Search Console". Inclusion is the default. Search Console's "Search generative AI" setting lets a site owner exclude a site's links and content from AI Overviews, AI Mode and generative AI features in Discover, and Google notes that this setting does not affect AI training, which is controlled by Google-Extended.

The other engines name their own crawlers. OpenAI documents OAI-SearchBot for ChatGPT search, Anthropic documents Claude-SearchBot, and Perplexity documents PerplexityBot, which it says "is designed to surface and link websites in search results". A robots.txt rule or a firewall setting written years ago for a different purpose can block one of them without anyone noticing. Content that exists only in images, behind a login, or in JavaScript a crawler does not run is also missing at this stage; Google's own advice is to make sure "important content is available in textual form".

How to check: in Google Search Console, open Settings and confirm the Search generative AI setting is on include. Read your robots.txt for each search crawler by name, and ask your hosting or CDN provider whether automated bot protection is returning errors to new bots.

What to change: allow the search crawlers you want, and decide about training crawlers separately. OpenAI says a robots.txt change can take about 24 hours to show in its search results. The full crawler list is in the indexability guide <!-- link-later -->.

Stage 1: when the engine searches, is your page among what it reads?

An answer engine runs one or more searches for the buyer's question, reads a set of results, and writes from those. Google calls the multiple searches "query fan-out" and says its AI features rely on its core ranking systems to retrieve pages from its index. Being indexed makes you eligible; being retrieved for the specific question is a separate event, and a strong Google ranking does not guarantee it. In our own measurement of Google: "In our study of 150 non-branded Google searches (US, July 2026), about 57% of the websites cited in AI Overviews did not appear in that same search's top 10 organic results (95% interval 54% to 60%)." An independent study found a similar share (Kirsten et al., arXiv 2510.11560, a preprint). The figure describes which sites AI Overviews cited, not the effect of moving up a position.

AI Overviews also draw on many sites for one answer. On commercial software searches we measured on July 19, 2026: "On 148 commercial SaaS searches such as 'best CRM for small business', Google showed an AI Overview on 143 of them, and each one cited a median of 9 different websites." A median of nine sources per answer leaves room for pages that do not hold the top position, and it also means your page usually shares the answer with several others.

How to check: ask the engine the questions your buyers ask, several times, and record which pages it cites. If your pages never appear for questions you rank well for, this is your stage.

What to change: pages that answer the sub-questions a buyer's question breaks into (integrations, pricing, compliance, alternatives), not only the head term. More in the retrievability guide.

Stage 2: does the engine trust your site as a source?

Once an engine has read several sources for a question, what it relies on depends on how it judges quality, and each engine judges differently. For Google, the guide says its AI features depend on the same core ranking systems that "focus on high-quality content" and on the systems that block spam, and it warns that seeking inauthentic mentions "isn't as helpful as it might seem". What we see in AI Overviews is that some platforms are cited far more than others: "YouTube is one of the few social platforms that regularly shows up as a source; posts on X almost never do"

How to check: list the domains cited across your buyers' questions. If third-party sites are cited even where your page is the original source of the fact, the gap is trust, and more pages of your own will not close it.

What to change: get the facts about you stated on the independent sites the engine already cites for your category: review platforms, documentation, analyst and comparison pages, video. More in the source reputation guide.

Stage 3: does the engine describe your company correctly?

An engine that mentions you can still describe an out-of-date version of your company, merge you with a company with a similar name, or put you in the wrong category. This is the most common failure after a rebrand, a new product or a move upmarket, because the web still holds years of pages about the old version of you, and your own site is only one voice among them. It is also where memory answers do the most damage, because a model trained before the change cannot know about it.

How to check: ask "What does [company] do?", "What category is [company] in, and who are its competitors?" and "Is [company] any good?" on each engine, three times each. Compare the description with your current positioning, sentence by sentence, and note whether the answer cited sources or came from memory.

What to change: make the current description consistent everywhere the engine reads about you, starting with the independent profiles it cites, keep the facts and dates on your own pages current, and report factual errors through each engine's feedback button. More in when AI describes your company wrongly.

Stage 4: when a buyer asks for options, does the engine pick you?

Being known and being recommended are measured by different questions. "What does [company] do?" tests whether the engine knows you. "Best [category] for [use case]" tests whether it puts you on a shortlist, and a company can pass the first and fail the second for months. A common cause is vocabulary: the buyer asks for "auth for startups" and your site only says "identity orchestration platform".

How to check: run the comparison and "best X for Y" questions your buyers ask, in their words rather than yours, and count how often you are listed across repeats and engines.

What to change: say plainly who you are best for and who you are not for, in the buyer's category words, on your own pages and in the comparisons others publish. More in being recommended over competitors.

Stage 5: does the answer send the visit, or replace it?

A citation inside an AI answer is often a brand impression rather than a visit. Pew Research Center followed 900 US adults through 68,879 Google searches in March 2025: people who saw an AI summary clicked a traditional result on 8% of visits, against 15% for those who did not, and clicked a link inside the summary itself on 1% of visits to pages that showed one. The same study's 18% figure is the share of searches in that panel's browsing that showed an AI summary; it is sometimes repeated as the share of all US Google searches, which the study did not measure.

How to check: compare AI referral sessions in your analytics with the number of questions on which you are cited. For Google, Search Console's Generative AI performance report shows how often links to your site were shown in AI Overviews and AI Mode, by page, country and device, so you can see impressions that never became visits.

What to change: decide which questions are worth a citation without a click (brand questions often are) and which need a visit, and give the reader a reason to click on those. More in zero-click AI answers <!-- link-later -->.

Stage 6: is the answer on your page stated plainly, in text an engine can fetch?

Stage 6 asks whether a reader who lands on your page, or an engine that fetches it, can find the answer to the buyer's question without guessing. A page that answers the question in its third paragraph, under a heading written for internal reasons, or only in a diagram, makes both work harder.

What this does not mean, for Google Search at least, is cutting pages into fragments or rewriting them in a special style for AI. Google's guide says there is "no requirement to break your content into tiny pieces", that its systems "are able to understand the nuance of multiple topics on a page and show the relevant piece to users", and that you do not need to write in a specific way just for generative AI search. Those statements are about Google. Other engines publish much less about how they pick what to quote, so we treat stage 6 as writing for the reader first and check each engine separately instead of assuming one engine behaves like another.

How to check: for each important question, find the sentence on your page that answers it, and confirm it is in the page text, not only in an image or a script. If a colleague cannot find it in ten seconds, a buyer will not either.

What to change: put the question in a heading where it helps the reader, answer it in the first sentence beneath, and state the fact plainly. More in pages written for machines versus for people.

Stage 7: do visitors from AI answers convert?

Visitors who arrive from an AI answer have usually read a summary of you already, so they land with a specific question and less patience for a generic home page. Treat them as a distinct segment.

How to check: segment visits whose referrer is an AI assistant, such as chatgpt.com or perplexity.ai, and compare their conversion path with search traffic.

What to change: send them to the page that answers the question they arrived with, and put the next step on it. More in converting visitors who arrive from AI answers.

What does not move any stage on Google?

Google's July 2026 guide lists things site owners can ignore for Google Search, including its AI features: llms.txt and other "special" files or markup, "chunking" content, rewriting content just for AI systems, seeking inauthentic mentions, and overfocusing on structured data. It says Google Search ignores llms.txt, so creating one "will neither harm nor help" a site's visibility there, and that "there's no special schema.org markup you need to add". Several popular AI visibility guides list schema markup as a stage of its own; for Google, that is not a stage.

These are Google's statements about Google Search. They say nothing about ChatGPT, Perplexity, Claude or Copilot, and we do not extend them. For those engines, check their own documentation in the engine guides.

How do I find out why AI assistants leave my startup out of their answers?

Start from the symptom, then check stage 0 first because it is the cheapest to fix and blocks everything after it.

After that, the order depends on your situation:

• Before a funding round or an enterprise deal: stages 3 and 2. Investors and procurement teams can ask an assistant about you before the call, and what it says about your credibility depends on what independent sites say, not on your own pages.
• After a rebrand or a new product: stage 3. Expect the old description to persist until other sites describe the new one, and longer in answers written from memory.
• Launching into a crowded category: stages 1 and 4, measured with the comparison questions your buyers actually ask.
• Traffic from AI is flat while citations rise: stages 5 and 7.

How do we measure what an engine tells a buyer about us?

We measure it with the same buyer questions, asked of more than one engine, three times each, with every answer's verdict and cited sources recorded. A single answer proves nothing, because answers change between identical asks. We describe the method in the AI due-diligence check.

How do you measure a funnel when answers change every time you ask?

You measure each stage with repeated questions and read the spread, never a single answer. AI answers vary with the wording of the question, between runs of the same question, and between engines, so one screenshot is an anecdote. The rules we follow in our research measurements, including the measurement of ourselves:

1. Ask every question at least three times, and repeat on another day before you believe a change.
2. Ask from a logged-out session or a temporary chat, so the assistant's memory of you is not in the answer.
3. Measure each engine separately and never average engines together; they search, rank and cite differently.
4. Use the buyer's wording, including the category words buyers use rather than the ones on your website.
5. Log, for every answer: whether you were named, how you were described, which competitors were listed, which sources were cited, the engine and the date. Sources tell you which stage failed.
6. Report a change only when it is larger than the run-to-run variation you measured before you changed anything.

For Google, add the Search Console Generative AI performance report to this: it is the one first-party count of how often your links appeared in AI Overviews and AI Mode. The detail is in how we measure AI visibility and why most AI visibility numbers mislead.

How long does a fix take to show in answers?

Two timings are documented by the vendors: OpenAI says about 24 hours for a robots.txt change to reach its search results, and Google says a change to the Search generative AI setting generally takes a few days. A positioning fix that depends on other sites rewriting what they say about you can take months, and an answer written from memory will not change until the model is retrained.

Engines also change how they search and what they read without notice, and a stage that is fine today can break after a vendor update. That is the argument for measuring regularly rather than once.

Frequently asked questions

Is AI search visibility the same as SEO? They overlap at stages 0 and 1 and diverge after that. Google's own position is that, for Google Search, optimizing for generative AI search "is optimizing for the search experience, and thus still SEO". Across engines, visibility also depends on whether the engine reads your page, trusts it, describes you correctly and picks you. See AEO versus SEO, measured.

Which AI engines matter most for a SaaS, fintech or other B2B company? The ones your buyers use, which you can measure from AI referral traffic and by asking customers. Measure each separately: ChatGPT, Google AI Overviews and AI Mode, Perplexity, Claude and Copilot search, rank and cite differently. The engine guides cover each one.

How many questions do I need to measure? Enough to cover the questions buyers ask at each stage: what you do, what category you are in, whether you are credible, how you compare, and the "best X for Y" questions in your category. A few dozen well-chosen questions, each asked several times, tell you more than hundreds asked once.

Do I need schema markup or an llms.txt file? Not for Google: Google says neither is needed for its AI features and that Google Search ignores llms.txt. Other engines have not said either helps them. Spend the time on stages 0 to 3 first.

Can a vendor guarantee that ChatGPT will recommend us? No. No one controls what an engine writes. What can be measured is where in the funnel you drop out, and whether a specific change moved that stage by more than normal variation.

Work with us

If you want the diagnosis and the fixes done for you, our managed AEO services work through these stages with your team: Book a call. If you would rather start by seeing where you stand, Lenz, our AI visibility platform, gives you a free analysis of how ChatGPT and Gemini answer your buyers' questions: Get your free AI visibility report.

Sources

• Google Search Central, "Optimizing your website for generative AI features on Google Search", last updated 10 July 2026: developers.google.com
• Google Search Central, "AI features and your website", last updated 10 December 2025: developers.google.com
• Google Search Console Help, "Search generative AI control": support.google.com
• Google Search Console Help, "Generative AI performance report (Search)": support.google.com
• Pew Research Center, Athena Chapekis et al., "Google users are less likely to click on links when an AI summary appears in the results", 22 July 2025: pewresearch.org
• OpenAI, "Overview of OpenAI Crawlers": developers.openai.com/api/docs/bots
• Anthropic, "Does Anthropic crawl data from the web, and how can site owners block the crawler?": support.claude.com
• Perplexity, "Perplexity Crawlers": docs.perplexity.ai
• Kirsten, Grosse Perdekamp, Wu, Upadhyay, Gummadi and Zafar, "Characterizing Web Search in The Age of Generative AI", preprint: arXiv 2510.11560
• Elementera, measurement of Google AI Overviews on non-branded searches and on commercial software searches.

Written byDr. Maziar "Maz" GomrokchiCo-Founder and CEO

AI/ML scientist with 20+ years of research & industry experience. Ex-Google DeepMind, MIT, Harvard, Mila. McGill-Mila PhD.

Reviewed bySoroush MehriCo-Founder and Chief AI Officer

Ex-Microsoft. 10+ years of experience in AI/ML, MSc from Université de Montréal (Mila). US patent holder.