What is AI visibility?

For twenty years the question was simple: can Google find this page, and where does it rank? An answer was a list of links, and the goal was a click. That is no longer the only way people arrive. They ask an assistant, get a paragraph, and act on it — and whether that paragraph came from your site or from a competitor was decided by things most audit tools never look at.

AI visibility is the name for that set of things. It is not a single score or a single setting. It is three separate capabilities, and the reason it is worth treating as its own discipline is that a site can pass any one of them while failing the other two — and each failure looks identical from the outside, which is to say invisible.

The three capabilities

1. Can they reach you?

Every AI operator crawls under its own user-agent token, and every token is governed independently by your robots.txt. This is where most sites lose without knowing it, because a rule written in 2023 to keep scrapers out is still in force and now excludes them from products that did not exist when it was written.

The critical detail is that these tokens are not aliases for one another. OpenAI alone operates three: GPTBot for training, OAI-SearchBot for search indexing, and ChatGPT-User for live retrieval when somebody asks about you by name. Blocking GPTBot — the common choice, and often a reasonable one — leaves the other two entirely unaffected. Anthropic splits the same way with ClaudeBot, Claude-User and Claude-SearchBot.

And a robots.txt group is not additive. A crawler obeys the single most specific group matching its name and ignores every other group in the file. So a site that adds a permissive block for GPTBot has not given GPTBot the site-wide rules plus permission — it has released GPTBot from the site-wide rules altogether. That surprises almost everyone the first time they see it.

2. Can they quote you?

Reachable is not the same as quotable. An assistant composing an answer has to identify where a claim begins and ends, decide it is self-contained enough to lift, and determine who to credit. A page that answers the question perfectly in prose spread across four paragraphs with no heading structure is a page it will summarise loosely, or skip in favour of a competitor whose answer it can bound precisely.

What makes a passage liftable is unglamorous: headings phrased as the question being answered, so the match is explicit rather than inferred. FAQ and how-to markup, which state a question and its answer adjacently in a form that needs no parsing. A named author and a visible publication date, because an assistant asked to cite a source needs something to cite. Semantic landmarks, so the article can be separated from the navigation and the cookie notice.

This is where AI visibility diverges hardest from classical SEO. Ranking rewards a page that is comprehensive. Quotability rewards a page that is decomposable. The two are compatible, but only if you write for the second on purpose.

3. Can an agent operate you?

The newest of the three, and the one with the shortest history. Software now books, buys and fills in forms on a person's behalf. It does not see your site; it reads the accessibility tree — the same structure a screen reader consumes.

Which means a div styled to look like a button, with a click handler and no role, is not a button to an agent. An input with a placeholder but no label has no name it can match against 'email address'. A menu that opens on hover will never open, because there is no pointer. A cookie wall with no programmatic dismissal is where the task ends. None of this affects your ranking, and all of it decides whether an agent asked to book a table at your restaurant succeeds or reports that it could not.

A site can be perfectly crawlable, rank first, be quoted accurately — and still be a dead end for an agent asked to complete a transaction on it.

Why each assistant behaves differently

They are four products from four companies with four sets of documented behaviour. Treating them as one audience is the root of most AI visibility mistakes.

• ChatGPT reaches you three ways — training, search indexing, and live fetch — and you may want to permit some and not others. That is a legitimate position, but it has to be expressed as three decisions, not one.
• Claude splits crawling from retrieval the same way. Because Claude-User fetches at request time, what your page says today is what gets quoted today — a fix propagates immediately rather than waiting for a recrawl.
• Gemini and AI Overviews are governed by Google-Extended, which is the single most misunderstood token in this area. It has no effect on ordinary Google ranking. Allowing Googlebot does not allow it, and disallowing it does not cost you a single blue-link position.
• Perplexity answers with citations by design, so quotability carries more weight there than anywhere else. Clear headings and a visible date are often the difference between being the source it names and being absent from an answer assembled from your competitors.

How to check where you stand

All three capabilities are testable, and none require guesswork. Crawler access is a matter of resolving your robots.txt the way a crawler resolves it — per RFC 9309, honouring group specificity and Allow overrides rather than reading the file top to bottom. Quotability is a matter of inspecting the rendered page for heading structure, markup and attribution. Agent readiness is a matter of examining the accessibility tree for named, reachable controls.

Our AI crawler checker tests each user-agent token individually and tells you what your file actually permits, which is frequently not what reading it suggests. The answer engine and AI agent checkers cover the other two capabilities. All three are free, and none of them need an account.

The short version

1. 1Check which AI tokens your robots.txt permits, one at a time, and confirm each result matches what you intended rather than what you assumed.
2. 2Give every substantive answer on your site a heading that states the question it answers.
3. 3Put a named author and a real date on anything you would want cited.
4. 4Make sure every control on your site is a real control with a readable name.
5. 5Re-check after any robots.txt change. This is the file where a one-line edit has consequences nobody notices for months.