What Is AI Visibility, and How Is It Measured?

AI visibility is not one thing. It is three separate questions you ask of each AI engine, for the questions your buyers actually ask.

Presence and citation are the parts most tools report. Accuracy is the part most brands actually complain about: the engine names them but quotes an old price or the wrong product.

Treat the three as separate readings. A brand can be mentioned often and described wrongly. Another can be cited rarely but quoted exactly. One number that blends the three hides which problem you have.

Every AI answer passes through up to four stages. Knowing which stage you are looking at explains most of what a visibility check shows.

1. Parametric. What the model knows with no web access, frozen at a training cutoff. OpenAI's model docs, checked in August 2026, list a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for ChatGPT's Instant path. So "what ChatGPT knows" is not one value.
2. Access. Whether a crawler may fetch and store your page at all. This is set by your robots.txt, your CDN or firewall, and your rendering.
3. Retrieval. Whether your page enters the model's context for a given question. The engine decides this, often after rewriting the user's prompt into its own search queries.
4. Generation. Which retrieved pages the model uses and cites.

The key fact: retrieval is optional and per query. As of October 2026, ChatGPT searches at the model's discretion, Claude.ai search is a user toggle, and Google AI Overviews often do not trigger at all. Some answers you see were built from live pages. Others came straight from training data, where nothing you publish this quarter can reach.

Because retrieval fires per query, any AI visibility score mixes retrieval-on answers with retrieval-off answers. You usually cannot tell which is which unless the answer shows citations.

Published estimates of how often ChatGPT searches do not agree. Semrush's US clickstream data put it at 34.5% of queries in February 2026, Profound reported about 18% of conversations, and cloro reported 80.5% globally on a 100-prompt set. They measure different things on different populations. The honest reading is that nobody knows the mix for your brand.

Two more problems make a blended cross-engine score shaky:

• Engines are different products. Google states that AI Overviews and AI Mode "may use different models and techniques" and can return different links. Gemini the app is a third surface.
• Citations are not the same unit. Anthropic's API cites short text spans. Perplexity cites whole sources by number. Adding them together has no shared denominator.

This is also why AI visibility tools disagree about the same brand. They test different prompts, different models and different retrieval settings. Many also call API models that differ from what the consumer app serves, which is worth asking about before you trust a trend line (see whether tools test the same models people use).

The move: read each engine on its own, and ask any tool whether its answers had retrieval on or off.

Presence and citation

You ask each engine a fixed set of buyer questions and record whether your brand appears and whether your URL is cited. Run them logged out with memory off, or the results won't repeat. Record the date, because engines change models often.

Know the limit: your prompts are a proxy. OpenAI says ChatGPT search rewrites a prompt into one or more targeted queries, and Google may fan a query out across subtopics. Nobody outside the engine sees those queries, which is covered in whether tools can see real AI prompts.

First-party reports see more than any prompt sample. In OtterlyAI's 2,000-blog experiment, Otterly's prompt tracker logged 3,934 citations for one test site, while Microsoft's own AI Performance report credited the same site with 42,817 (the time windows may not match). Google's own tools are thinner: as of our August 2026 review, Search Console's AI Overviews and AI Mode report shows impressions only, with no clicks, click-through rate or position.

Accuracy

Accuracy needs a list of your true facts: prices, plan names, founding year, location, ownership. You compare each answer against it and flag what's wrong or unverifiable.

Score it as a rate, not a pile of penalties. Established factuality metrics divide errors by a count of facts or answers, so the same error rate scores the same at 20 answers or 400. The hallucination rate and FActScore's factual precision both work this way. Count only answers that actually came back. A failed request is missing evidence, not a clean answer.

The evidence is clear about where your effort pays off. A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) found that already-retrieved content can change how it is cited, but no technique shows a stable, cross-platform effect on being discovered in the first place.

In your control (access)

• Rendering. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Content that only appears after JavaScript runs may be invisible to them.
• Per-bot robots.txt. OpenAI's bots page says each setting is independent. Blocking GPTBot stops training use. OAI-SearchBot controls ChatGPT search.
• Network reachability. A firewall or CDN can return a 403 to a crawler that robots.txt allows. OpenAI's help centre says eligibility needs both.
• Index and snippet eligibility. Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out.
• Stated facts. A specific price or one-line definition on a fetchable page gives the engine a sentence it can quote.

Not in your control

• What the model learned before its cutoff.
• Whether the engine searches for a given question, and which queries it runs.
• Where your page lands in the model's context, and which pages it cites.

Brand prominence shows a strong association with citation across studies, but it is not a lever you can pull this quarter. When engines cite other sites instead of yours, being named on those sites is the route into the answer.

Work from the gates outward. A page the crawler cannot reach will never be cited, however good it is.

1. Check access first. Confirm robots.txt lets OAI-SearchBot, PerplexityBot, Claude-SearchBot and Googlebot in, then confirm your firewall actually serves them a 200. An AI crawler checker requests a page as each crawler and shows which ones get turned away.
2. Check rendering. Compare the raw HTML with the rendered page. Move prices, product names and key claims into the HTML the server sends.
3. Check index status per engine. Look at Google Search Console and Bing Webmaster Tools separately. In Otterly's experiment, Google deindexed both test sites while Bing kept them, and Copilot kept citing them.
4. Write down your true facts. One list, one source of truth. Fix contradictions across your own pricing, about and footer pages.
5. Take a baseline per engine. The free checker on this site runs up to three buyer questions and scores the answers. It is a snapshot, not tracking.
6. Re-measure on a schedule. Models and defaults change. Set a re-check cadence and log the date with every reading.

The AI SEO checklist covers each step in more detail.

• Trusting one blended score. In Otterly's test, Copilot produced 2,718 of one site's 3,934 tracked citations while four other engines barely cited it. A blended number would have hidden that.
• Reporting a rank. No engine exposes position inside an AI answer. Google's AI report has no position field, so an "AI rank" has nothing to be checked against.
• Counting citations as influence. OpenAI's docs say the sources a model consults often outnumber the citations it shows. Cited and used are not the same.
• Allowing GPTBot to get into ChatGPT search. GPTBot governs training only. Search uses OAI-SearchBot.
• Adding schema or an llms.txt to get cited. Google says structured data isn't required for generative AI search and that Search ignores llms.txt. Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. Existing markup does no harm, but it is not a lever for AI answers.
• Comparing scores across tools or formula versions. Different prompts, models and retrieval settings produce different numbers. Compare a tool only with itself, over time.

Is AI visibility the same as AI SEO? They overlap. AI SEO usually means the work you do to appear in AI answers. AI visibility is the outcome you measure: presence, citation and accuracy, engine by engine. Much classic SEO hygiene carries over, because Google says AI Overviews and AI Mode need no special optimization beyond being indexed and snippet-eligible. Other engines add their own crawlers and indexes, so access has to be checked per engine.

Can I see where I rank in ChatGPT or Google AI Overviews? No. No engine publishes a position inside its AI answers. Google's Search Console report for AI Overviews and AI Mode showed impressions only when we reviewed it in August 2026, and Google says third-party tools have no access to its internal ranking data. Any tool showing an AI rank is reporting its own model, not a measurement. Ask for presence and citation counts with the retrieval state disclosed instead.

Why did my AI visibility change when nothing on my site changed? Engines change models, defaults and retrieval behaviour without notice. A new model can carry a different knowledge cutoff, and retrieval fires per query, so the mix of live and training-data answers shifts. Infrastructure can change too: from September 15, 2026, Cloudflare began blocking training and agent crawlers by default on ad-displaying pages for some customers. Re-check access before assuming the content is the cause.

Does fixing rendering or robots.txt guarantee citations? No. Access fixes remove blocks. They make your page eligible to be fetched and retrieved, but the engine still decides whether to search, which pages to pull and which to cite. Nobody can promise a citation. What these fixes do guarantee is that a missing citation isn't caused by a crawler that could never read your page in the first place.

How many engines should I measure? Measure every engine your buyers use, each on its own. At minimum that usually means ChatGPT, Google AI Overviews, Gemini, Perplexity and Claude, plus Microsoft Copilot if Bing matters in your market. Google AI Overviews, AI Mode and the Gemini app are separate products that can return different answers, so do not treat one as a stand-in for the others.