How AI assistants choose which businesses to recommend

An assistant recommends the companies that are named clearly and repeatedly in the small set of sources it is reading for that exact question. It is not scoring your product. It is reading a few pages, looking for names it can describe in one line, and writing a short answer out of what those pages agree on.

What is actually happening underneath

Two different things can produce the answer, and they behave differently.

What the model already holds. The model was trained on a large amount of text. If your company was written about in that text, it may be able to name you with no searching at all. If it was not, it cannot invent you.

What it fetches while answering. Many assistants now run a search first, read a handful of pages, and answer from those. This is the path a smaller or newer business can actually influence, because it happens today rather than at training time.

In both paths, the question wording narrows the field before anything else happens. "Best billing software" and "best GST billing software for a small trading firm in India" pull different sources and produce different names. The qualifiers in the question are doing most of the filtering.

Then the assistant has to compress. It usually names three to five companies, and it prefers names it can attach a short, confident description to. A company with no clear one line description is the easiest one to leave out.

The six filters a name passes through

It helps to think of it as a sequence rather than a score. A company that fails an early filter never reaches the later ones, which is why product quality so often turns out to be irrelevant to the outcome.

One, the question is read literally. Every qualifier is treated as a requirement. City, country, company size, budget, industry, regulation. A question about a free tool for one person will not return a company that sells annual enterprise contracts, and that is the correct behaviour.

Two, a source set is assembled. Either from memory or from a search. This is the step where most companies are lost, because a name that appears nowhere in the fetched pages cannot appear in the answer built from them.

Three, candidate names are pulled out of those pages. Names that appear in more than one source are safer, because the assistant is effectively looking for agreement.

Four, each candidate is checked against the question. Does this page actually say the company serves this segment, this place, this price band. A page that says only that the company is a leading provider of solutions fails this check quietly.

Five, a one line description is written for each survivor. If there is no clear sentence available anywhere about what you do and who for, this step is expensive, and the cheap alternative is to name somebody else.

Six, the list is cut to length. Three to five names, usually. Order tends to follow the ordering and emphasis of the sources rather than any independent ranking.

A worked example

This comes from our own work rather than from a scenario. clawlaw.in is our sister company, and in July 2026 we took a blind reading of its category: 78 buyer questions, with no brand named in any of them.

ChatGPT, on 27 July 2026, then confirmed it had run no live web search for any of those 78 questions. So the whole reading described what the model already held about the category rather than what it could find that week.

The ordering it produced is the useful part. Counted by how many of the questions each company was named on, ProVakil came first on 28, CLAW second on 21 and Legistify third on 18. CLAW is the brand of clawlaw.in. That order is a measure of how much had been written and absorbed about each name, not a measure of the products. Nothing was tested.

Two things follow for any business. On a question where no search happens, your place is decided by text that already exists, and a page you publish this month is not in it. And the six filters above are still the right order to work in, because a company absent from the sources cannot be named by either path.

What it means for a business

You are competing to be the easy thing to say, not the best thing on the market. That has three practical consequences.

• Other people's pages matter as much as yours. A name that appears on your own site and nowhere else has one source behind it.
• Vague copy is a real cost. If the only sentence available about you is that you offer solutions tailored to your needs, the assistant has nothing to repeat.
• Category fit beats quality. If the question is about free tools and you sell to large firms, you will not be named, and that is the correct answer.

It also changes who owns the work. Filling in a directory profile, correcting an old listing and writing a plain one line description are not marketing campaigns. They are small, unglamorous tasks that decide whether you are a candidate at all.

What this explanation does not cover

Nobody outside these companies can see the ranking logic, and none of them publish it. Everything above is drawn from what is observable: which sources get cited, how answers change when the question changes, and what appears or disappears when a page or a listing changes. That is enough to act on and it is not the same as knowing the mechanism.

Three things this page cannot tell you. It cannot tell you why a specific answer named a specific company on a specific day, because the same question can answer differently within the hour. It cannot promise that work on the six filters will get you named, and anyone promising a position in an AI answer is promising something they do not control. And it will go out of date, because these products change how they retrieve and cite without announcing it.

Common questions

Can I pay to be recommended?

You cannot buy the recommendation itself. What money can buy is presence: a paid listing or a sponsored slot on a directory can get you onto a page the assistant may read, which makes you eligible. Whether it then names you is not for sale, and a vendor who implies otherwise is selling a guarantee they do not hold.

Does the assistant know whether my product is any good?

No. It has never used your product or your competitor's. It is reporting what its sources say, including their opinions. This is why a well described average product is named more often than an excellent one nobody has written about.

Why did the answer change when I asked the same thing again?

Two reasons. These models produce different wordings on the same input, and the search step may fetch a different set of pages. So a single answer is one draw, not a measurement. Watch the pattern across several readings instead.

Does my website even matter, if other people's pages decide it?

Yes, in a narrower role than most people expect. Your pages are where a careful assistant checks your facts: what you cover, who you serve, what you charge. They rarely win you the mention on their own, and they very often lose it by being vague.

Do I need to be the biggest company in the category?

No, but you do need to be visible in the sources for the question as asked. Narrow questions are where smaller companies get named, because the big general pages do not answer them well and fewer vendors have written for them.

How is this different from ordinary SEO?

Most of the technical work carries over, because a page a crawler cannot read cannot be quoted either. What changes is the target. Ranking puts you on a list somebody chooses from. Being named puts you inside the answer they were given, and there is no position eleven.

What to do first

Write down ten questions your buyers actually ask, in their words. Ask each one of two assistants without naming your company, because naming yourself ruins the reading. Record two things for every answer: who was named, and which pages were cited.

Then do what the worked example did. Open every cited page and mark yourself present, absent, or present but wrong. The absent and wrong rows are your first week of work, and they are usually cheaper than writing anything new.

We build a tool that runs this at scale, so treat that as our interest declared. The manual version above costs nothing and tells you the same thing about ten questions.