Three different outcomes get called "visibility"

When people say a brand is visible in ChatGPT they can mean three things, and they move for different reasons. Mentioned: the brand's name appears somewhere in the answer. Cited: one of the brand's URLs is attached as a source. Recommended: the answer actively suggests the brand for the question asked. A brand can be cited without being recommended, and recommended without a single link. Reporting them as one number hides exactly the gap you need to see, so track them as separate columns from the start.

Why it starts with crawler access

Nothing downstream matters if ChatGPT cannot reach your pages. OpenAI documents three agents with different jobs. GPTBot crawls content that may be used to train models; blocking it is a defensible position on training use and does not remove you from ChatGPT search results. OAI-SearchBot builds the index that surfaces websites in ChatGPT's search features; blocking it removes you from those results, and it is the agent sites most often block by accident while meaning to opt out of training. ChatGPT-User fetches a page because a person asked for it, and it may not follow robots.txt at all. One wildcard rule aimed at OpenAI collapses three decisions into one. Name each agent explicitly, then verify by fetching your own pages as each agent rather than by reading robots.txt — a CDN or WAF rule can return 403 to a named bot while the file says "Allow". The agent names and IP ranges are published at OpenAI's bots documentation.

How to measure it so the number survives being checked

  1. Fix a question set. Twenty to forty buyer questions in the words a buyer would use, with no brand names in them. A question that names your brand will be answered with your brand; that is an echo, not visibility.
  2. Run every question more than once. Answers vary between identical runs, so a single sample proves nothing. The unit is a rate over the set, not one observation.
  3. Detect every brand with the same rule. Your brand and your competitors must be matched by the same code path and the same name normalisation, or the result flatters whoever wrote the detector.
  4. Count citations only from the answer body. Sources that were retrieved but never attached in the reply are not citations.
  5. Keep the raw answers. Without the answer text, engine and date behind each number, you cannot tell a detection bug from a real change — and you will have both.

The same discipline is laid out in more depth in brand share in AI answers and in how to track brand mentions in ChatGPT.

What people get wrong

  • Treating it like a rank. There is no position ten inside an answer. Tools that report a "ChatGPT rank" are usually reporting the order brands happened to appear in one sampled answer, which is not stable enough to trend.
  • Rendering unmeasured as zero. A run that failed or a week with no collection is not 0% visibility. Rendered as zero it turns a flat trend into a fake decline.
  • Checking once and acting on it. Variance is the default. Two consecutive runs moving in the same direction is a finding; one run is noise.

If you want a first sample before building any of this, the free checks at tools.insightwonder.com start with the access questions above.