How ChatGPT Decides Which Brands to Recommend

ChatGPT picks brands the way a careful but hurried analyst would: it starts with what it already believes about your category, checks that against whatever it can quickly read on the web, and writes down the names that both sources agree on, in roughly the order they tend to appear. There is no ranking table. There is a probability distribution over brand names, shaped by how consistently, how recently and how widely each brand has been described as a good answer to that kind of question. This guide explains each part of that process and which parts you can influence.

Two sources feed every recommendation

When a buyer types “best project management tool for a creative agency with 25 people”, two things can supply the brand names in the answer.

The first is model knowledge: the compressed statistical memory of everything the model read during training, up to its training cutoff. If your brand appeared in thousands of documents describing it as a project management tool for agencies, the model has a strong association and will reach for your name without looking anything up.

The second is retrieval, sometimes called retrieval augmented generation. For many commercial prompts, ChatGPT runs a web search, pulls a handful of pages, and uses them as grounding for the answer. The brand names on those pages have a direct route into the response, which is why ChatGPT can recommend a product launched after its training cutoff.

Most answers are a blend. The model knows roughly which names belong, retrieval confirms or updates that list with current pages, and the final text reconciles the two. The practical consequences of this split, and how to tell which one is producing your result, are the subject of ChatGPT Search vs model knowledge.

Step one: the model decides what kind of question this is

Before any brand is considered, the model interprets the prompt. “Project management tool for a creative agency” triggers a different mental category than “work management platform for enterprises”, even though several brands could fit both. The words a buyer uses define the pool of candidates.

This is where many brands lose without knowing it. If your marketing calls you an “operations intelligence platform” and buyers ask for “inventory software”, the model has to infer that you belong, and inference is where you drop out. Brands that are described, by themselves and by others, in the vocabulary buyers actually use are matched directly. The vocabulary question is covered in depth in prompt clusters: how buyers actually phrase the question.

Suppose a cluster of 36 buyer prompts for a category. In practice it is common to see a brand named in most answers to the prompts that use its own preferred terminology and almost absent from the prompts that use the buyer’s terminology, which is where the volume is.

Step two: candidate brands are drawn from consensus

Once the category is clear, the model needs names. Here its logic is closer to “what do most credible sources say” than “who is best”. A model trained to avoid confident errors will prefer the brand that appears on eight comparison pages, three review platforms and two industry publications over a brand that appears on one excellent review, because the first is the safer answer.

The 2026 research on how AI engines cite the web supports this. The University of Toronto audit found that earned media accounted for 57% of GPT-4o citations, and that overlap between AI-cited sources and Google’s top ten was only 4.0% for GPT-4o. The model is reading a wide, mostly third-party slice of the web and is not simply echoing the Google results page.

Consensus has three components worth separating:

• Breadth: the number of independent sources that place you in the category.
• Agreement: whether those sources describe you the same way (same positioning, same strengths, same customer type).
• Credibility: whether the sources are the kind the model treats as reliable for this category (trade press, established review platforms, expert communities), rather than press releases and paid directories.

A brand can have breadth without agreement (mentioned everywhere, described inconsistently) and fail to be recommended, because the model cannot settle on what it is.

Step three: recency filters the candidates

Retrieval heavily prefers fresh pages. The same research roundup reports that cited content had a median age of 62 days for Claude versus 130 days for Google in consumer electronics. A model answering “best X in 2026” will reach for pages that say 2026 in them.

This creates a quiet disadvantage for established brands with mature coverage. Your definitive reviews may be three years old. A newer competitor with a run of recent articles can appear in grounded answers above you despite a weaker overall reputation. The fix is not to churn out content on your own site; it is to make sure current third-party pages describe you accurately and that the pages you control are visibly maintained.

Step four: entity clarity decides whether the signals add up

All the evidence in the world does nothing if the model cannot connect it to one brand. An entity, in this context, is the model’s internal notion of a single thing with stable attributes: name, category, what it does, who it serves, where it operates, when it started. Entity consistency is how reliably every source agrees on those attributes.

Common ways brands fracture their own entity:

• Operating under two names (the legal name and the product name) that sources use interchangeably without ever connecting them.
• Describing the product differently on the website, in the app store listing, on review platforms and in press boilerplate.
• Sharing a name with an unrelated company, product or common noun and doing nothing to disambiguate.
• Rebranding without updating the long tail of directory entries, old articles and social profiles.

In our audits, entity fragmentation is one of the most frequent root causes behind a brand that “should” be recommended and is not. It is also one of the cheapest things to fix, because most of the work is editing listings you already control. The full approach is in brand entity consistency for large language models.

Step five: the answer is written, and position is assigned

With a pool of candidates, the model writes. In our reading of answers, order is not random. Brands that appear first in the source material, that are described as the default or market leader, or that match the prompt’s constraints most precisely tend to be written first. The model is also writing for a reader, so it often leads with the safest, most widely recognized name and uses later positions for alternatives that fit a niche in the prompt.

Position is worth tracking separately from mention rate. A brand that is named in nine answers out of ten but sits fifth each time is losing to a brand named in six answers out of ten that sits first. Buyers read the first three names and skim the rest. The commercial case for position is laid out in why position 1 to 3 is the whole game.

Why the same prompt produces different answers

Everything above is probabilistic. The model samples its words rather than computing a deterministic ranking, retrieval returns slightly different pages on different runs, and the user’s location, device, login state and conversation history all alter the context. Ask “best CRM for a small law firm” ten times and you might see a brand in seven answers, in positions ranging from one to four.

This variance is why a single screenshot of ChatGPT recommending you (or not) means little. Measurement has to be done across many answers per prompt, with the conditions controlled. Our own method reads about 100 real answers per prompt per market, from genuine mobile connections in the target country, and classifies each one for brand named, position, link present and competitors named. It is described on the page describing our audit method. What you get from that volume is a confidence band, not a decimal, and that is the honest shape of the data.

Being named is not the same as being linked

A recommendation can name your brand without linking to your site, and it can link to a third-party page about you rather than to you. Linked sources, or citations, rotate heavily: a 2026 study of 1,127 cited URLs found only 10.6% persisted across 28 days, and 40 to 60% of cited sources rotate monthly. Brand names in answers are far stickier than the links beneath them, because the name comes from accumulated consensus and the link comes from whatever retrieval happened to pull that day.

Treat the mention as the primary outcome and the citation as a bonus that drives a small, high-intent click stream. The distinction, and how each is earned, is covered in the difference between a mention and a citation.

What ChatGPT does not reward

It is worth being explicit about what does not move the answer, because a great deal of money is wasted on it.

• Keyword repetition on your own site. The model is not a keyword matcher and your site is confirmation, not evidence.
• Volume of backlinks for its own sake. Links help pages get retrieved, but a thousand low-quality links do not create consensus.
• Mentions scattered across irrelevant or low-credibility sites. If the source would not persuade a careful human buyer, it does not persuade the model either, and a pattern of such mentions can read as noise.
• Paid placements that are not labeled as such. Beyond the obvious risks, these rarely appear on the pages the model actually grounds on for commercial questions.
• Ads inside ChatGPT. The recommendation inside the answer is organic. According to third-party reporting, ads appear as a “Sponsored” card beneath the answer, only for logged-in US adults on the Free and Go tiers. They do not change the names the model writes.

What does work is slower and plainer: become the brand that credible, current, independent sources consistently describe as a good answer to the buyer’s question, and make it trivially easy for a model to confirm that on your own pages.

What this means for your plan

Because the decision has distinct steps, you can diagnose where you are losing. If you are absent even from prompts that use your own terminology, the problem is entity or consensus. If you appear for your terminology but not the buyer’s, the problem is vocabulary and category framing. If you appear in model-knowledge answers but vanish when ChatGPT searches, the problem is recency. If you are named but always late in the list, the problem is how sources rank you relative to competitors.

Each diagnosis points at different work. The overview of the full discipline is in the complete guide to ChatGPT brand visibility. If you would rather see where you stand first, find out what ChatGPT currently says about your brand and we will show you the prompt-level picture. If you want the work done for you, The Blue Ocean GPT targets consistent mentions across up to 90% of a prompt cluster with a top-three position, reported so your team can reproduce each result in ChatGPT. How we do it is explained on your call.

What to do next

• Write your buyer’s questions in your buyer’s words, then test which category the model places you in.
• Audit the sources ChatGPT cites for your category and check whether they describe you consistently and currently.
• Fix entity fragmentation across every listing you control before investing in new coverage.