How do I get my business recommended by ChatGPT?

When ChatGPT answers a question about current businesses it is not recalling them from training. It runs a search and reads results, drawing on search infrastructure and on OpenAI's own crawlers, then writes an answer from what came back. Being recommended means being in what came back, and being the clearest thing in it.

OpenAI publishes its crawlers separately and they do different jobs. GPTBot gathers data for training. OAI-SearchBot is the one that builds search results. ChatGPT-User fetches a page when a user's question sends it there. They are separate controls, which matters: a site that blocked GPTBot to stay out of training may still be fully available for search, and a site that blocked everything is invisible to the surface it wanted to appear on.

Be retrievable. Check robots.txt actually permits the search crawlers, that pages render their content in the HTML rather than only after JavaScript runs, and that the sitemap is present and current. A page an assistant cannot read is not in the running regardless of how good it is.

Be answer-shaped. The pages that get quoted state the answer in the first few lines, in a form that survives being lifted out of context. A page that takes four paragraphs to reach its point gives a summariser nothing to summarise.

Be consistent as an entity. Same business name, address and phone everywhere they appear, so the machine reading three sources concludes they describe one company rather than three similar ones.

Optionally, publish llms.txt — a proposed convention for a plain-text map of a site's useful pages. It is a convention rather than a standard and nobody is obliged to honour it, so treat it as cheap and speculative rather than as a lever.

There is no paid placement in an assistant's answer, no directory that guarantees inclusion, and no submission that puts a business in front of a model. The supply of people selling all three is growing quickly.

What can be influenced is whether the material exists, whether it can be retrieved, and whether it is the clearest available answer. That is the whole mechanism. It is unsatisfying because it is slow and cannot be rushed with budget.

On 16 September 2026 we ran our own citation baseline: 68 questions from this bank, put to three assistants — Perplexity, Claude and Gemini — twice each, producing 204 answers. We counted how many cited this site and how many named it in the body text.

Both numbers were zero. Zero of 204 cited, zero of 204 named, on the day these answer pages started shipping. That is the baseline the same battery will be re-run against, and it is published here because a before-figure only means anything if it was recorded before.

Two other things that run showed, which are worth more than the zero. One assistant supplied almost the entire citation field while the other two returned nearly none, so a claim about being cited by AI is meaningless without naming which assistant. And the sites being cited in this subject area are overwhelmingly documentation — the standards bodies, the browser makers, the platform docs — not agencies.