How ChatGPT decides which businesses to recommend

Two different machines behind one answer

When someone asks ChatGPT to recommend a plumber in Brisbane or a commercial lawyer in Austin, the reply comes out of one of two processes, and they work nothing alike.

The first is recall. The model answers from what it absorbed during training: a compressed impression of an enormous amount of text, with no live lookup, no sources and usually no links. The second is grounding. The assistant runs one or more web searches, reads some of what comes back, and writes an answer over the top of it, with citations attached.

Which one you get is not something you control, and usually not something the person asking notices. The two can look almost identical on screen. They have entirely different causes, and therefore entirely different fixes.

What a recalled answer is made of

Recall is slow-moving. The model's impression of your sector formed from text published before its training data was collected, and nothing you publish this week changes it. If a model names three joinery firms in Nottinghamshire without searching, it is naming the firms that were written about most often and most consistently in the material it saw.

Two things follow from that. Recalled answers favour businesses with a long, broad, consistent written footprint across the open web: trade registers, local press, association listings, supplier case studies, forum threads where someone asked for a recommendation and several people answered. And recall is where the errors live. A model reconstructing a plausible business from fragments can produce a trading name that is slightly wrong, a service it never offered, or a firm that closed two years ago.

There is no quick lever here. Recall responds to being genuinely well documented over a long period, which is a slower project than most owners want it to be.

What a grounded answer is made of

Grounding is the part you can influence this month. The assistant turns the question into one or more searches, retrieves a set of pages, and summarises them. The answer is therefore constrained by the retrieval step: a business that does not appear in the pages retrieved cannot be named, however good its website is.

That reframes the problem usefully. You are not trying to persuade a model. You are trying to be present, and clearly described, in the documents the model is likely to pull for a question like "who should I use for X in Y". Very often those documents are not your site at all — they are a directory, a review platform, a trade body's member list, or a local roundup article. Our guide on what AI assistants actually read goes into which source types keep turning up.

It also explains a result that frustrates people: a business can rank perfectly well on Google for its main service and still never be named, because the pages the assistant retrieved were third-party listings it was absent from.

Why the same question twice gives different names

Ask the identical question an hour apart and you may get a different shortlist. Three separate mechanisms cause this, and they stack.

• The model samples. It does not look up one correct answer; it generates text probabilistically. Where two businesses are close in its estimation, which one is mentioned first can flip.
• It may or may not search. The same question can be answered from recall one time and from a live search the next, producing two answers with completely different source material.
• The web moved. A grounded answer is a snapshot. A directory reshuffling its order, a new review, a competitor publishing a comparison page — any of these change the inputs.

On top of that, the consumer app personalises. Location, memory and earlier conversations all feed in, so two people in the same town can get materially different lists. Why answers differ between people and between runs covers this in detail, and it is the reason a single check tells you almost nothing. The pattern across many questions, repeated over time, is the only reliable read.

The two routes, on one question

Take an online gift retailer and the question "where can I get a personalised engraved hip flask". Answered from recall, the model produces the handful of retailers whose names recur most in the text it was trained on — typically large marketplaces and a few long-established specialists, with no links and occasionally a shop that no longer operates. Nothing a small retailer did this quarter affects that answer.

Answered with a live search, the same question produces something quite different. The assistant retrieves a set of pages, which for a question like this usually includes a marketplace category page, one or two gift roundups and perhaps a comparison article, then writes a shortlist from them with citations. A small retailer absent from those pages is absent from the answer, regardless of how good its own product pages are.

Same question, same assistant, two different competitive fields. When you look at an answer and cannot tell which route produced it, the presence or absence of citations is the quickest tell.

What becomes a citation, and what does not

Citations only exist on grounded answers. A recalled answer has nothing to cite, which is worth remembering before you conclude that a competitor "got the citation" — on that run, nobody did.

When citations do appear, they are the pages the assistant read, not an endorsement and not a ranking. A few patterns are consistent enough to plan around. Pages that make a plain, checkable statement get cited more than pages that gesture at one. Pages that answer the specific question asked get cited more than homepages. And aggregator pages — "best X in Y" style roundups, directories, review sites — get cited heavily for discovery questions, because one page answers the whole query.

The practical reading of that: if your own page is never cited but a roundup you are missing from is cited every time, the work is getting onto the roundup, not rewriting your homepage again.

GPTBot, OAI-SearchBot and what blocking costs you

OpenAI publishes documentation for its crawlers, and the distinction between them matters because they are frequently blocked by accident, in a single sweeping rule.

• GPTBot crawls content that may be used to train OpenAI's foundation models. OpenAI states that blocking it signals your content should not be used for that training, and that it does not govern search inclusion.
• OAI-SearchBot is the one that surfaces sites in ChatGPT's search features. OpenAI's documentation says sites opted out of it will not be shown in ChatGPT search answers, though they can still appear as navigational links.
• ChatGPT-User fetches a page when a user's own question requires it. Because those visits are user-initiated, OpenAI notes that robots.txt rules may not apply to them.

So the costs are different. Blocking GPTBot is a defensible position on training, and some publishers take it deliberately. Blocking OAI-SearchBot removes you from the grounded answers that are the fastest-moving part of this whole picture. Most businesses that have blocked both did not decide to — a security plugin, a CDN bot rule or an inherited robots.txt did it for them. It is worth checking whether you are blocking the AI crawlers before concluding anything else about your visibility, because a firewall rule will quietly override every other improvement you make.

What to do with all this

The mechanism points at a short list of work, in order of how fast it pays back. Confirm the crawlers can reach you. Make sure the pages you want quoted contain sentences a model could lift verbatim and be correct — what you do, where, for whom, at roughly what cost. Get your trading name consistent across every third-party source that mentions you, so several listings read as one business rather than three weak ones. Then work on being present in the aggregator pages that keep getting cited for your category.

The tactical version of that sequence, with the specific checks, is in our guide on how to get recommended by ChatGPT. This article is the explanation underneath it: if you understand which of the two machines produced an answer, you know which lever applies and roughly how long it will take.

Before changing anything, get a baseline. You cannot tell whether you are improving, or whether a bad week is just variance, without a starting number — a free visibility check takes a few minutes and tells you whether you are being named at all today, and who is being named instead.