How Does ChatGPT Decide Which Companies to Recommend?

Is the mechanism published anywhere?

No. How a model weighs one company against another when composing a recommendation has not been documented by the companies that build these systems, and there is no public specification an outside party can check a claim against. Every confident account of the selection rule circulating in marketing material is a reconstruction, and a reconstruction of an undocumented system is a guess delivered in a firmer voice than the evidence supports.

Saying so plainly matters more here than anywhere else on this site. An article promising to explain an internal mechanism, and then explaining it, would be selling the exact thing this firm argues against. What follows is the observable half. Observable is smaller than the promise in the title, and it is the part that converts into work.

What is publicly known about how an answer is assembled?

Two situations exist and behave differently, and telling them apart is the first useful thing a reader can do. An assistant that browses the web to answer a question draws on pages it fetched during that request, and often shows the sources it used. A model answering without browsing draws on what it absorbed during training, which is a different situation with a different failure mode and no citation list to read.

Google's documentation is the clearest public description of the retrieval half in any comparable product: its AI-generated search answers are drafted from a set of pages a conventional search retrieved first, a step the documentation calls grounding. Whether any other provider builds its answers the same way is not knowable from outside, and no claim about that is made here.

What can actually be observed from outside?

Four readings, all of them recordable by anyone willing to do the work. Whether the company is named for a given question. Which companies are named instead. How the company is described when it does appear. And which sources the answer cited, wherever citations are shown.

Repetition is what turns those readings into evidence. Answers vary between runs and change as providers update the systems behind them, so one screenshot is an anecdote about one moment. The same questions, asked in the same order on a schedule, with every run kept in full, produce a series. A series is the only thing a change can be seen in.

What are the observations good for?

Citation lists repay attention first, because a cited page is a specific thing somebody can go and act on. Where the answers on a subject are consistently built from one publication, a directory entry and a competitor's own comparison page, those are three specific pages to address, and two of them sit outside the company's control. A share-of-voice reading says a company is losing. A citation list says what it is losing to.

Falkview holds a stated expectation here, open to being wrong and written down so a later run can score it. A company several independent sources describe consistently is easier for any answer-composing system to name than a company only its own domain describes. The reasoning is that reconciling several agreeing accounts is a smaller problem than trusting a single self-description. No provider has published a weighting that confirms this, and this firm has not measured a client population large enough to claim it as a finding.

What does the firm's own record say?

One measurement here belongs to Falkview rather than to anyone's theory. On the day this site launched, a run recorded the company named in none of its own category questions while other companies were named in them. What Our Own Baseline Said on Day One publishes that in full, including the expectation that assistant answers would keep omitting the firm for longer than search does. The reasoning recorded there is that a domain with no independent coverage gives a system a single account to work from.

A prediction on the record can be scored later. A prediction described after the fact cannot, which is why the baseline was published on day zero rather than summarized in a case study a year on.

What should a company do with an undocumented mechanism?

Work on the inputs, and measure the output. The inputs are the ones any outside party can reach. Pages stating plainly what the company does and who for. An identity that resolves to the company rather than to older organizations with similar names. And independent sources with something accurate to say. None of those is a lever on a model. Each of them changes the material an answer would be assembled from.

Measurement is what keeps the work honest, because it records movement without claiming a cause. An answer that changes between two runs may reflect work done on the site, or a change the provider made that week, and separating those from outside is rarely possible. A record that keeps both, and labels which is an observation and which is an attribution, survives a year of scrutiny.

Where can a reader take this further?

How to Get Your Company Mentioned in ChatGPT covers the practical half of what remains within reach once the mechanism is set aside. What Is AI Search Optimization? sets out the full sequence that work runs in. Why AI Assistants Ignore Most Small Companies covers why a company with real customers and a real reputation can still be absent from every answer.

Anyone comparing proposals has one test available here. A firm that explains the selection rule confidently is describing something no provider has published, and that is worth knowing before the second meeting.