How Does ChatGPT Choose Which Brands to Recommend?

ChatGPT chooses which brands to recommend using two layers: what the model learned about your category during training, and what it finds when it searches the live web at the moment of the question. Shaping those two layers is the discipline called generative engine optimization (GEO), which you will also hear called answer engine optimization (AEO). Same job under either name: when a buyer asks an AI engine about your category, there is already an answer, and this work is about earning your place in it. In this piece I will walk through the mechanism the way OpenAI documents it, what the published research says shifts recommendations, and the honest list of things nobody controls.

If either term is new to you, the two hub guides cover the foundations: [what generative engine optimization is](/blog/what-is-generative-engine-optimization) and [what answer engine optimization is](/blog/what-is-answer-engine-optimization). This piece goes one level deeper into the machinery.

The two layers: what the model remembers and what it looks up

When ChatGPT names a brand, the name came from one of two places.

The first is training. OpenAI's [crawler documentation](https://platform.openai.com/docs/bots) describes GPTBot as the crawler that gathers web content used to train its foundation models. Everything written about your brand across the open web, by you and by everyone else, feeds the model's baseline sense of who matters in your category. Ask a question with search switched off and this layer is all you get: a compressed memory of the web as it stood at training time.

The second is live retrieval. When a question benefits from current information, ChatGPT runs a web search, reads the results, and builds its answer from the sources it just fetched, citing some of them along the way. This layer is where most practical GEO work concentrates, because the training layer moves at the pace of model retraining while retrieval responds to changes on the web within weeks.

A definition worth pinning down: a brand recommendation in ChatGPT is a synthesis of retrievable evidence. The engine is summarising what the sources it trusts say about your category, so the question underneath every recommendation is whether those sources mention you.

What happens in the seconds after a buyer asks

OpenAI's [help documentation on ChatGPT search](https://help.openai.com/en/articles/9237897-chatgpt-search) describes the retrieval sequence in useful detail. ChatGPT typically rewrites the user's question into one or more targeted search queries and sends those to third-party search providers, with Bing named as one partner. After reviewing the first results it may send additional, more specific queries, then it selects sources and composes the answer. OpenAI's own example: a question about drugs targeting CCR8 for cancer becomes the query "CCR8 immunotherapy drug development 2025", followed by narrower follow-ups.

Two practical consequences fall out of that sequence.

First, your buyer's question is fanned out into several phrasings you never see. Ranking for the exact phrase a buyer used matters less than being retrievable across the cluster of queries the engine generates around it. Content built to answer the deeper pain-point questions in your category, in language the engine can lift cleanly, covers more of that cluster than a single optimised page.

Second, access is binary. The same OpenAI documentation states that sites disallowing OAI-SearchBot in robots.txt will not be shown in ChatGPT search answers, though they can still appear as navigational links. That setting is independent of GPTBot, so a site can stay in search answers while opting out of training, or vice versa. The failure mode I see most often in audits is accidental: sites blocking AI crawlers wholesale, usually through an old bot-management rule nobody has reviewed, which quietly removes them from consideration before any content question even arises. It is the first thing worth checking, because the best content plan on the web does nothing for a site the engines cannot read.

What the research says gets picked

The strongest public evidence on which content changes move AI recommendations is the [GEO paper](https://arxiv.org/abs/2311.09735) from researchers at Princeton, Georgia Tech, IIT Delhi, and the Allen Institute (Aggarwal et al., accepted to KDD 2024). The team tested nine content modifications across a large benchmark of queries and measured how each changed a source's visibility in generative engine answers. The headline finding: adding citations to sources, including quotations, and including statistics boosted visibility in generative engine responses by up to 40% on their benchmark, while classic keyword stuffing did comparatively little. The effect sizes also varied by domain, which argues against one generic playbook.

Read that finding plainly and it says the engines reward evidence density and extractable structure. A page that states a claim, attributes it, and quantifies it gives a retrieval system something safe to quote. This is also why brands with modest traditional SEO can still surface in AI answers: the retrieval layer is weighing whether a passage is useful and corroborated, so a young site with well-evidenced pages can be quoted alongside incumbents with far older domains. We have watched clients gain AI visibility while their classic rankings were still catching up, and the research gives that observation a mechanism.

The consensus layer

Here I am moving from documented mechanism to practitioner observation, so treat this section as what we see across client engagements rather than anything OpenAI publishes.

The engines lean hard on third-party corroboration. Every business describes itself as the best, which makes self-description nearly worthless as a signal, and external mentions become the tiebreaker. When we built AI visibility for BizScout, a business marketplace, the work produced over 450 AI mentions in a year, and ChatGPT now cites them as the go-to source for business buyers. Tides Mental Health, a behavioural health practice, went from page 2 to page 1 with 4.7x organic traffic growth in five months, and AI engines now recommend them for anxiety treatment. In both cases the recommendations followed the accumulation of consistent, retrievable evidence across many sources.

One more observation from the same work: candour performs. AI systems are trying to act as an objective broker for the user, so they favour sources that do the honest comparison work for them, including being clear about who a product does not fit. Naming your competitors and your limits reduces the model's uncertainty about where you belong, and content that does this gets lifted into answers more readily than content that only sells.

Ethan Smith of Graphite covers this same machinery from an independent seat, including how models use retrieval and citations, in his conversation with Lenny Rachitsky:

<YouTube id='iT7kq-R3Gjc' />

What nobody controls

A recommendation mechanism this new deserves an honest ceiling, so here is what no agency, tool, or tactic changes.

**Answers are not deterministic.** The same question on different days can name different brands, because retrieval results shift and models sample. Anyone showing you a single screenshot as proof of anything, favourable or otherwise, is showing you one roll of the dice. Trends across many prompts are the honest unit of measurement.

**Personalisation shapes the query.** OpenAI documents that when Memory is enabled, ChatGPT may use what it knows about the user to rewrite their question into a better search query. Two buyers asking the same thing can trigger different searches and receive different recommendations, and you will never see either rewrite.

**Model updates reshuffle the board.** A retrained model carries a new baseline picture of your category, and nobody outside OpenAI knows the schedule or the effect in advance.

**There is no purchase path.** You cannot buy a place inside the organic answer, and in full transparency, nobody can promise you a citation by a date. New content does not show up in AI answers overnight; every result we have produced took months before it snowballed.

Why moving early still wins

Those caveats might read as reasons to wait. The market data argues the opposite. Organic search traffic is down roughly 30% across the board as AI answers absorb clicks, while AI referrals convert at 3 to 6x the rate of traditional organic visitors. The buyers arriving from AI answers are fewer and far warmer, because the engine already did the comparison work before they clicked.

And the mechanism itself favours early movers. Once a brand sits inside the answers for a category, it accumulates the mentions, links, and coverage that being recommended generates, and that fresh evidence is exactly what the retrieval layer weighs next time. Citations compound in AI visibility the way backlinks compounded in early search, which means the durable route is the patient one: real evidence, honest comparisons, clean technical access, built up over months. A visibility position earned that way keeps paying out long after the work is done, and a steady organic channel of that kind is an asset an acquirer values in a way no ad account ever is.

If you are weighing whether to build this capability in-house or hire it, I have written an [honest decision guide on the agency vs DIY question](/blog/geo-agency-vs-diy) that includes the case against hiring us.

The place to start either way is knowing where you stand. Our free 46-point AI visibility checklist covers seven categories, from crawler access and technical structure through to citation-ready content, and it will show you which layer of the mechanism above is holding you back. [Download the checklist](https://lorebuilders.com) and run your own site through it this week.