How ChatGPT decides which brands to recommend

Ask ChatGPT "what's the best CRM for a five-person sales team" and it doesn't consult a ranked database of CRMs and return the top result. There is no leaderboard. It's generating an answer, token by token, based on patterns learned from training data and — increasingly — live web sources it retrieves at answer time. Understanding that process, roughly, tells you what to actually go do about it.

It's prediction, not retrieval from a ranked list

The core thing to unlearn: ChatGPT (and Claude, Gemini, Perplexity) doesn't have a stored opinion of "the best CRM" sitting in a database, the way Google has a ranked index of pages. When you ask a recommendation question, the model is predicting the most plausible next words given everything it has seen — in training, and for models with browsing/retrieval (ChatGPT with browsing, Perplexity always, Gemini with Search grounding), in what it fetches live for that specific question.

That means two brands with identical product quality can get very different answers, because the model isn't scoring quality — it's reproducing whatever pattern of association it has learned between "CRM for small sales teams" and specific brand names. If that association doesn't exist anywhere it learned from, the brand doesn't come up. Not because it's bad. Because it's absent from the pattern.

Where the pattern comes from

Three inputs shape what a model says when you ask for a recommendation:

1. Training data. Everything the model learned from up to its knowledge cutoff — articles, review sites, forums, comparison posts, company sites, social discussion. A brand that's been written about a lot, in the specific context of "X for Y," gets baked into the model's learned associations. A brand nobody wrote about in that framing doesn't.

2. Live retrieval. Models with browsing don't just rely on frozen training data — they can search the live web for the specific question and read a handful of pages before answering. This is why fresh content matters even for a model trained months ago: if it retrieves at answer time, this week's comparison post can outrank last year's silence.

3. The question itself. How you ask shapes what gets surfaced. "What is [brand]" pulls from the model's direct knowledge of that brand alone. "Best CRM for a small sales team" forces the model to assemble a shortlist from whatever it associates with that category — a completely different, and much harder, retrieval task. Most companies only ever test the first kind of question on themselves, and it tells you almost nothing about the second.

What actually earns a mention

Given that, three things measurably move whether a brand shows up in a recommendation answer:

• Third-party evidence, not self-description. A model has learned to discount "we're the best" copy on a company's own homepage — everyone says that, so it's low-signal. What it weighs more is independent sources: review platforms, comparison articles, "best of" roundups, forum threads where someone recommends a brand unprompted. That's evidence a real preference exists, not just a claim.
• Content shaped like the question. If a buyer asks "X vs Y" or "alternatives to X," a model answers best when it has read content that directly frames that comparison. A brand with a real, honest comparison page (or one written about them) gives the model a ready-made answer to draw from. A brand with only a generic homepage gives it nothing shaped like the question being asked.
• Consistent, extractable positioning. If ten sources all describe a company as "the CRM for five-person teams that hate data entry," that phrase becomes a stable fact the model can repeat with confidence. If the description varies wildly across sources — or exists nowhere at all — the model has no clean sentence to reach for, and usually just leaves the brand out rather than guess.

None of this is about tricking the model. It's closer to what SEO always rewarded — genuine relevance, clearly stated, backed by outside proof — just applied to a different surface with a different evidence bar.

Why this is hard to self-diagnose

You can't reliably infer any of this by asking ChatGPT about your own brand by name — that tests recall, not recommendation. And a single manual check tests one engine, one question, one moment; answers vary across engines and across runs of the same question, so one anecdote can mislead you either direction. To actually know where you stand, you need the buyer-framed questions, put to more than one engine, read for the actual sentence returned — not just whether a name appears somewhere in it.

That's the exact gap our free scan closes: 25 real buyer questions, run against ChatGPT, Claude, Gemini, and Perplexity, with the literal answer each one gives — so you can see whether you're named, who's named instead, and which of these three levers (evidence, shaped content, consistent positioning) is actually your problem. Enter your domain, no signup, about a minute.