The question “how do I rank on ChatGPT” contains a wrong assumption, and clearing that up is most of the value of this article.
There is no ranked list. There is no position three. There is a retrieval step, a synthesis step, and a short list of citations that changes between runs of the same prompt.
Once you hold the right model of how it works, the practical actions become obvious, and so do the things that are not worth doing.
How does ChatGPT decide what to cite?
Two mechanisms feed an answer, and they behave very differently.
Live retrieval. When a question needs current information, the system searches the web, retrieves candidate passages, and writes an answer from them with citations. This is the route you can actually influence.
Model knowledge. Where the question does not trigger a search, the answer comes from what the model already holds from training. You cannot edit this, and it changes only when the model is updated.
The retrieval path is documented in OpenAI’s own material on ChatGPT search, and the crawlers involved are listed in OpenAI’s bot documentation. Those are the sources to check when something in this area changes, which it does frequently.
What gets selected within retrieval follows familiar logic. Passages that directly answer the question, from sources that appear credible and are corroborated elsewhere, in content the system can parse cleanly.
What is the difference between training data and live search?
It matters because it determines how quickly anything you do can show up.
A new page on an established site can be retrieved and cited within days of being crawled. The same page has no effect at all on what the model “knows” until a future model update, which is a much longer and less predictable cycle.
This explains a common frustration. A business publishes a strong page, sees it cited when it asks a specific question that triggers a search, then sees a generic answer with no mention when it asks a broader question that does not.
The practical conclusion: optimise for the retrieval path, and treat improvements in the model’s underlying knowledge as a slow secondary effect of being widely mentioned across the web.
What actually increases your chances of being mentioned?
In rough order of leverage for a typical small business:
- Confirm you are not blocked. Check
robots.txtfor GPTBot and OAI-SearchBot. A blanket disallow makes everything below pointless. - Get accurately listed in the roundups that already get cited. Ask the buying questions, note which third-party pages come back, and go and get listed on them properly.
- Publish pages that answer specific questions directly. One page that answers a real question in its first 60 words beats five pages of general category content.
- Make your entity consistent. Same business name, same description, same location, same contact details everywhere. Inconsistency makes you harder to resolve as a single entity.
- Earn mentions in credible independent sources. Trade press, industry directories, partner sites, local press. Unlinked mentions count here.
- Keep content current and visibly dated. Freshness affects both retrieval and how confidently a system will use you.
- Add structured data. Cheap, quick, and removes ambiguity about what your page is and who wrote it.
Why do third-party mentions matter more than your own site?
Because a claim on your own website is a claim, and the same claim on three independent sites is a fact.
Retrieval systems weight corroboration heavily, which is rational: your own marketing copy is the least reliable source of information about you.
That produces an uncomfortable conclusion for anyone whose entire strategy is publishing on their own domain. Beyond a certain point, another blog post on your site adds very little, while getting accurately described on four pages you do not control changes the outcome.
This is also why the businesses winning AI mentions in a category are often not the biggest. They are the ones that appear in the comparison articles, the directories and the community threads that the models pull from.
The listicle problem
Ask any model “best plumber in Leeds” or “best AI receptionist for small business” and watch what it cites. It will rarely be a vendor’s own homepage.
It will be a roundup, a directory listing, a review platform or a forum thread. Those pages are the actual battleground, and you almost certainly do not own any of them.
There are three practical responses:
- Get listed accurately. Find the pages that get cited, check whether you are on them, and correct your entry where you are.
- Earn inclusion where you are absent. Most roundup authors will consider a well-argued case, particularly if you can offer something specific and verifiable.
- Fix what is wrong. Outdated pricing or an old service list on a third-party page will be repeated back to prospects for years.
None of this is glamorous and all of it works. It is closer to digital PR than to SEO, which is why most in-house SEO effort misses it.
What do you do about competitor comparison prompts?
“Is [your brand] better than [competitor]” is a high-intent prompt, and the answer is assembled from whatever the web says about both of you.
If no independent comparison exists, the model works from both sets of marketing copy, which usually produces a bland both-have-merits answer. That is a missed opportunity rather than a disaster.
Two things help. Publish an honest comparison that includes where you are not the right choice, because balanced comparisons get retrieved more readily than sales copy. And make sure your differentiators are stated as specific, checkable facts rather than adjectives.
“Answers unlimited simultaneous calls” is retrievable. “Best-in-class service” is not.
How do you check whether you are being mentioned?
Build a fixed set of 15 to 20 prompts, run each one three to five times in a logged-out session, and record how often you appear and what is said.
The repetition is not optional. The same prompt genuinely produces different answers on different runs, so a single check tells you nothing reliable.
Run it monthly, keep the prompts identical between rounds, and record which competitors get named. The full method, including the four metrics worth tracking, is in how to measure AI search visibility.
What you cannot control, and should stop worrying about
You cannot control the wording of an answer, guarantee a mention, buy placement, or make the model forget something outdated that it learned during training.
What you can do is remove the technical blockers, publish content that is genuinely the best available answer to a specific question, and make sure the independent sources describing your business are accurate.
That is the whole job. Everything else marketed under this heading is either a repackaging of those three things or a promise nobody can keep.