How ChatGPT picks which companies to name

ChatGPT names the companies that appear repeatedly in the small set of sources behind that answer, that clearly fit the question as asked, and that it can describe in a single confident line. It is not ranking products on quality. It has no way to test them.

What ChatGPT does differently

The output shape drives a lot of the behaviour. The answer is usually three to five names in prose, each with one short description. That is a hard limit on how many companies can appear, so the question is not whether you are good enough but whether you are one of the few it can describe quickly.

Order is worth understanding too. It tends to follow the ordering and the emphasis of the sources it read rather than any independent judgement. If a comparison article put your competitor first, that ordering often survives into the answer.

It also answers the question as asked, quite literally. Qualifiers narrow the field before any company is considered. Change the city or the company size in the question and the list changes, because the sources it reaches for change.

And it sometimes does not search at all, in which case the sources are whatever the model holds. The same question can therefore produce a source driven answer one day and a memory driven answer the next. The selection logic itself is not published, so treat everything here as inference from what can be observed.

What that means for whether you are named

Four practical things follow.

Agreement across sources beats a single strong page. A name on four independent pages is safer than a name on one excellent page, including your own.

A missing one line description is a common reason to be cut. If nothing anywhere says plainly what you do and who for, the assistant has to guess, and guessing is more expensive than naming somebody else.

Category fit is checked before quality. If the question asks for a free tool for freelancers and you sell to large firms, not being named is the correct answer and no content fixes it. Work on the questions you should win.

The ordering of third party pages matters more than your own page. Getting moved up inside an existing comparison article can change answers faster than publishing a new page of your own.

What to change

• Write your one line description once: what you do, who for, where. Put the same sentence on your home page, your profiles and your listings. Consistency is what makes it repeatable.
• Get named on more independent pages, even briefly. Breadth of mention does more than depth on your own site.
• Put yourself in the right categories on the sites the assistants read, including the secondary ones you genuinely fit.
• Answer the qualified questions specifically. A page about your city or your customer size can win a question your general page cannot.
• Where a third party page describes you wrongly or places you oddly, write to the author with the correct facts. This is the cheapest lever on the list.

A worked example

The most awkward result in our own programme came on 17 September 2026, in a Claude run over the category of clawlaw.in, our sister company.

Its pages ranked in the assistant's raw results and shaped what it wrote. It still landed as a name inside a list rather than as a linked recommendation. The reason given was that its own comparison pages read as vendor advocacy, so they were usable as background and not citable as a source.

That is the difference between being read and being recommended, and it turns on what your page reads like rather than on whether it was found. A vendor comparison page can be good enough to inform an answer and still not be good enough to be the thing the answer points at. We recorded this on Claude rather than on ChatGPT, so read it as the mechanism and not as a ChatGPT specific finding.

The lesson matches the list above. Your own page is read to confirm facts, and the shortlist is built out of pages that read as independent. Breadth across other people's pages is what moves you from a name inside a list to the name an answer stands behind, and correcting how a third party describes you is the cheapest lever on that list.

What this does not explain

Four honest gaps.

The internal logic is not public. No weighting, no scoring formula, no ordering rule is documented. What is written above is drawn from observable behaviour and can go out of date.

Why a particular answer named a particular company. Single answers vary between runs. Only patterns across readings support a conclusion.

What happens in uncited answers. There is no source list to inspect, so the honest statement is that those draw on training data you cannot audit.

Whether the displayed citations are the full set considered. They are what the app chose to show. Nobody outside can confirm more than that.

And no amount of this work is a guarantee. It makes you a candidate, which is the only honest product in this field.

Common questions

Does being mentioned more often on my own site help?

Very little on its own. Your site is one source, and repeating your name inside it does not create agreement across sources. Breadth across independent pages is what changes the shortlist.

Why does it name a competitor that is clearly worse?

Because it is not judging products. It is reporting what its sources say, and a well documented average product beats an excellent undocumented one every time in this process.

Can I influence the order I appear in?

Indirectly, by influencing the ordering of the third party pages that get read, and by being unambiguously the fit for the question as asked. There is no setting and no promise available here.

Does it matter that the answer describes me in one line?

It is the whole of your presence in that answer, so yes. Decide that sentence yourself, publish it everywhere in the same words, and correct it where a third party has written a different one.

Should I worry about being named fifth rather than first?

Track both, because people act on the first names they read. Closing the gap between named anywhere and named in the first three is usually cheaper than getting named in new questions.

How many independent mentions do I need?

There is no threshold worth quoting and any figure offered as one is invented. The useful target is relative: more of the repeatedly cited pages for your questions than the competitor you lose deals to.

What to do first

Log three things from each answer, not one. Whether you were named. Where in the list you appeared, because being fifth and being first are different outcomes. And the exact sentence used about you.

That sentence is the most useful line in the whole exercise. Compare it against the description you publish. If it does not match, the assistant is reading somebody else's version of you, and you now know what to correct and where. Fix the source that carries the wrong version before you write anything new.