How ChatGPT decides which companies to recommend, and what is still unknown.

The short answer.

When ChatGPT searches the web to answer “which companies should I consider for…”, it rewrites the question into one or more searches, reads a set of pages that come back, and names the companies those pages mention most clearly and most often. Ranked list articles, directories and reviews feature heavily in what it reads. The exact weighting is not published, and the answer changes from one run to the next.

So the practical question is not “how do we rank in ChatGPT?” It is “when ChatGPT reads the pages that answer our buyer’s question, are we on them, and do they describe us correctly?” The rest of this guide explains each step, so you can see where your company drops out.

Two ways ChatGPT can answer.

ChatGPT can answer from what its model learned in training, or it can search the web first. The two behave very differently for a company, and OpenAI’s own crawler documentation shows the split. OpenAI runs separate crawlers: GPTBot collects content that “may be used in training our generative AI foundation models”, while OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features”. A third agent, ChatGPT-User, visits a page when a user’s question needs it, and OpenAI says it “is not used to determine whether content may appear in Search”.

That matters for two reasons. First, an answer given without a search reflects the web as it was when the model was trained, so recent changes to your site will not show. Second, if your site blocks OAI-SearchBot in robots.txt, OpenAI says it “will not be shown in ChatGPT search answers”, apart from navigational links. OpenAI also notes that a robots.txt change takes about 24 hours to be picked up.

This week, open your robots.txt (yourdomain.com/robots.txt) and search it for “OAI-SearchBot”. If a blanket rule blocks unknown bots, ask your developer whether it catches this one. It is the cheapest check in this guide.

Your buyer’s question is not the one it searches.

When ChatGPT does search, it does not paste the buyer’s words into a search box. It rewrites them. Reporting on OpenAI’s April 2025 update, TechCrunch quoted OpenAI’s documentation: ChatGPT “will rewrite that prompt into a search query”, and that query “may also leverage relevant information from memories”. The example in that report turned “restaurants near me that I’d like” into “good vegan restaurants, San Francisco” for a user whose memory said they were vegan and lived there. Memory can be switched off in settings.

A single question also tends to become several searches. Google describes the same technique for its AI features as “query fan-out”: “issuing multiple related searches across subtopics and data sources”. For ChatGPT, a Grow & Convert study of 100 buying-intent prompts (last updated 30 March 2026) found that 83 triggered a web search, each generating two to four fan-out queries, and that ChatGPT answered without searching 17% of the time. The authors note they could only see the searches visible in their browser.

What that means for a B2B company

Your buyer asks a long, specific question: industry, country, size, problem. ChatGPT turns it into shorter searches, such as the category plus “agencies”, the category plus the country, or the category plus “reviews”. You are competing for the pages that answer those shorter searches, not for the buyer’s exact sentence. Write down five questions a real buyer would ask, then guess the two or three plain searches each would become. Those are the pages to check first.

Which pages get read, and which get cited.

A common assumption is that ChatGPT simply reads Google’s top ten. The evidence says otherwise. In the same Grow & Convert study, only 27% of the sources ChatGPT cited ranked on Google for the fan-out queries, and 23% on Bing. Allowing for overlap, about 40% appeared anywhere in the first ten pages of either engine. Good Google rankings help you get read, but they are not the whole story.

There is also a gap between what is read and what is shown. OpenAI’s developer documentation for its web search tool, which describes the same kind of search for developers, separates the full list of “URLs the model consulted” from the inline citations, and states that “the number of sources is often greater than the number of citations”. A page can shape the answer without ever appearing as a link under it. That is why counting citations alone undercounts your influence, and why being absent from the citations does not prove you were not read.

Our earlier guide on checking your AI search visibility separates mentions, citations and recommendations properly. For this week, the useful habit is simpler: when you test a question, open the sources panel and save every URL, not just the ones linked in the text.

Why lists and mentions carry so much weight.

Ask a model to recommend companies and it needs pages that compare companies. Your own service page says what you do; it cannot say you are one of several firms worth considering. Pages that do say that are ranked list articles, directories, review sites and comparison posts, so those are what a shortlist question pulls in.

The largest study we could verify on this is by Ahrefs (4 December 2025). Across 750 top-of-funnel prompts for software, products and agency recommendations, “best X” blog lists were the most common page type ChatGPT cited, at 43.8% of all page types. In 250 “best X software”-style results, 169 (67.6%) included a list where the publisher ranked itself first, and 35% of cited lists sat on low-authority domains. The author is careful: “I hesitate to give a definitive conclusion.”

A second Ahrefs study (26 May 2025) looked at 75,000 brands in Google’s AI Overviews and compared signals with how often each brand was mentioned.

Being talked about on other sites tracked AI mentions about three times as closely as raw backlinks. Ahrefs adds the right warning: “correlation ≠ causation”. Read together, the two studies point the same way. An engine answering a shortlist question leans on what other pages say about you, not only on what your page says about itself.

The trap in that finding

Self-published lists clearly get read, which tempts every firm to publish one with itself at the top. Engines are not blind to this. In our own test, Perplexity’s answer warned that many of the lists it drew on were written by the agencies they ranked. Honest comparison content that a buyer would find useful is worth writing. A list that exists only to rank yourself first is a weak foundation and a reputational risk.

Why the answer changes every time you ask.

If you asked ChatGPT about your category once and were not named, you have one data point, not a verdict. Research by SparkToro with Gumshoe.ai, reported by Search Engine Journal on 30 January 2026, ran 2,961 prompts through ChatGPT, Claude and Google’s AI search features with hundreds of volunteers over November and December 2025. It found less than a one-in-100 chance that ChatGPT or Google’s AI would return the same list of brands across repeated runs of the same prompt. The researchers concluded that how often a brand appears across many runs is more consistent than its position in any single list.

Engines also disagree with each other. Google states that AI Mode and AI Overviews “may use different models and techniques, so the set of responses and links they show will vary”, and those are two features from the same company. Add memory, approximate location, the model version and whether a search ran at all, and two buyers asking the same question can see different shortlists.

What to do with this: never judge your position from one answer. Fix five to ten buyer questions, run each several times per engine in a clean session (a temporary chat with memory off), record the date and every company named, and track the share of runs that name you. Our guide to AEO and GEO visibility sets out a fuller protocol.

What happened when we tested ourselves.

On 6 November 2026 we asked three engines which agencies a US B2B company should use to improve its visibility in AI search. We work in exactly this field, so it was a fair test of whether we practise what we sell. We ran each question once, so these are snapshots, not rates.

We were not named anywhere. Perplexity drew on ten sources and, as mentioned above, warned that many were agency-written rankings. Only one agency appeared in all three answers, which matches the research: each engine reads a different slice of the web.

Then we asked ChatGPT about us by name. It found us, described our ten service areas correctly, and called us “worth shortlisting” for SEO and AI search visibility, suggesting a buyer start with a small, clearly scoped audit. It also judged our evidence for AI search weaker than our evidence for SEO, quoted an older figure for a client result that we had since corrected on our site, and gave a founding year that conflicted with older directory listings.

The lesson for any B2B firm: being known and being shortlisted are different problems. ChatGPT could describe us when asked directly, so our facts were reachable. We were missing from the category answer because the pages it read for that question did not include us. Those need different fixes, and the order of work is in our guide on why brands go missing from AI shortlists.

What is genuinely not known.

Anyone who tells you exactly how ChatGPT picks companies is guessing past the evidence. OpenAI has not published how it weights sources, and every study above was run from outside, by firms that sell SEO or AI-visibility tools. That does not make them wrong. It does mean their numbers describe what was observed in a sample, not how the system is built.

Two practical rules follow. Be sceptical of any service that promises placement in ChatGPT, because nobody outside OpenAI controls it. And measure with your own repeated tests, because published averages may not match your category, your country or your buyers’ wording.

What you can influence this week.

You cannot change how ChatGPT ranks. You can change what it finds at each step of the chain. Work through these in order, because each one depends on the one before.

1. Make sure you can be read

Allow OAI-SearchBot. Check your key service and about pages load their text without a script, login or cookie wall. If a page cannot be fetched, nothing else on this list matters.

2. Say plainly what you are, for whom and where

A model extracting firms from a page needs a clear statement: what you do, which buyers you serve, which countries, and dated evidence. Our article on AEO, GEO and SEO explains why the same clear page serves search engines and AI answers alike.

3. Get onto the pages that answer the shorter searches

Take the fan-out searches you guessed in chapter 03 and look at what ranks for them. Which directories, review platforms and comparison articles appear? Are you listed, with current facts? Earning a place there, through a complete profile, genuine client reviews or an editor’s inclusion, is slow and unglamorous. It is also where the evidence points.

4. Keep your facts consistent everywhere

When ChatGPT described us, it mixed figures from our site with older copies elsewhere. Every inconsistency is a chance for the answer to be wrong. List your core facts and check them on every profile you own.

5. Show evidence an engine can quote

Specific, dated results with a named source give a model something firm to repeat. Vague claims give it nothing.

Would ChatGPT shortlist you?

Tick each statement that is true for your company today. Items are weighted by how early they sit in the chain: if you cannot be read, nothing later counts. Your score and your weakest area appear below. Nothing you tick is sent anywhere.

A high score does not mean ChatGPT will name you; nobody can promise that. It means you have removed the reasons it could not. If access or clarity is your weakest area, fix that before spending anything on lists or reviews.

Questions owners ask.

How does ChatGPT decide which companies to recommend?

When it searches, ChatGPT rewrites the question into one or more searches, reads the pages returned, and names the companies those pages mention clearly. Ranked list articles, directories and reviews feature heavily. When it does not search, it answers from training data. OpenAI has not published how it weights individual sources.

Why does ChatGPT recommend my competitors and not us?

Usually because the pages it reads for your category question mention them and not you. Check which list articles, directories and comparison pages rank for the shorter searches behind your buyer’s question. Also check that your site does not block OpenAI’s search crawler, and run the question several times, because one answer is not a verdict.

Can I pay to be recommended by ChatGPT?

We know of no documented way to buy a place in ChatGPT’s organic answers. Treat any service that promises one with caution. What you can do is make your site readable, state your facts clearly, and earn genuine mentions on the independent pages that answer your buyers’ questions.

Does ChatGPT give everyone the same answer?

No. Research by SparkToro and Gumshoe.ai found less than a one-in-100 chance of the same brand list on repeated runs of one prompt. Memory, approximate location, model version and whether a search ran all change the answer. Track how often you appear across many runs, not your position in one.

How long does it take to start appearing in ChatGPT answers?

Nobody has published a reliable timeline. A robots.txt change is picked up in about a day, according to OpenAI. Earning mentions on independent lists and review sites takes months. Measure monthly with a fixed set of questions so you can see change rather than guess at it.

Sources: Prolinkage AI self-test, ChatGPT, Perplexity and Google AI Overview, 6 November 2026 (one run per engine). Client result: FullTilt Team Development, GSC Generative AI features impressions, last 3 months vs previous 3 months, captured 4 Sep 2026. OpenAI crawler documentation (accessed 7 Nov 2026); OpenAI web search tool documentation (accessed 7 Nov 2026); TechCrunch, 18 Apr 2025; Google Search Central, AI features and your website (accessed 7 Nov 2026); Grow & Convert, updated 30 Mar 2026; Ahrefs, 4 Dec 2025 and 26 May 2025; Search Engine Journal on SparkToro and Gumshoe.ai research, 30 Jan 2026.