ChatGPT recommends the businesses that the pages it reads already recommend. When it searches the web to answer “which X should I buy” or “who does Y near me”, it mostly cites independent reviews, publishers, forums and video — in our study of 8,441 AI answers to 414 buying questions, the brands' own websites were only 0.6% of everything cited. So the work is: know which questions your buyers ask, measure whether you are in those answers, get onto the pages the assistants already cite, make your own pages easy to quote, and re-measure.
Written from our own measurements, not from opinion: the figures come from our published study and methods note. ChatGPT here means its web-search answers; we measure through OpenAI's API, which is not identical to the consumer app.
Asking ChatGPT “what is your brand?” tells you almost nothing: anyone who types your name has already found you. The answers that send customers are to questions without a brand in them — “best ergonomic chair under £300”, “family hotel with a pool in South Tyrol”, “induction sealer for small batches”. Three good sources for those questions, in order:
Get this wrong and you get a precise, worthless number: a shop measured against the wrong questions scores zero with a tight interval (in the study, 14 of 79 brands turned out to be measured on a category that did not match what they sell, and were excluded).
One answer is a mood, not a measurement. Assistants re-draw their sources every time they are asked: asked the same question twice, minutes apart, the same engine's cited sources overlapped by only 61.3% on average across 9,996 answers (re-run noise note). So ask each question more than once, on more than one assistant, and read the result as a rate with an interval, not a yes or no. In our study, 63% of brands were cited in none of the answers to their category's questions. A zero is common and it is a starting line, not a verdict.
Across the study, the most-cited sources were youtube.com, independent.co.uk, reddit.com, theguardian.com, goodhousekeeping.com, forbes.com, which.co.uk, expertreviews.co.uk — video, national publishers, consumer-review sites and forums; ChatGPT alone cited 5.5 sites per answer on average. Your category will have its own list. For each question you lose, look at which pages are cited instead and ask which of them could mention you: a review you could be included in, a comparison that leaves you out, a forum thread where an honest answer from you is welcome, a directory or marketplace listing you do not have yet.
Assistants also name businesses without linking to them — in the study, a brand was named in the text 3.6 times as often as it was linked as a source. Being talked about on cited pages is how you get named; being the best answer on your own page is how you get linked.
These are patterns we see on cited pages, not guarantees: whether a change moved anything is what step 6 is for.
ChatGPT's search fetches pages with OAI-SearchBot; GPTBot collects pages for training. Blocking GPTBot is a reasonable choice and does not take you out of today's answers; blocking OAI-SearchBot does. Firewalls and bot protection can block both without anyone deciding to. Check yours with the AI crawler checker and the AI-readability check.
After you change something, ask the same questions again and compare. Small samples move on their own: at 108 answers, a score starting at 2% needs to rise by about 7.4% points before a two-sided Fisher exact test at 5% can call it a change. Anything smaller is within noise, however good it looks on a chart. That is why we re-measure every two weeks or every month rather than daily, and why every report says which moves cleared the test.
Other guides: Reading an AI visibility score · how we measure · who AI assistants cite