How do AI chatbots decide which brands to recommend?

This is the mechanism behind every “why does ChatGPT recommend my competitor?” story. Knowing which layer produced an answer tells you where you can act, how fast a change can show up and what is outside your control.

Layer one: what the model remembers

A language model’s training data froze at some cutoff date. If your brand appeared often enough, in consistent contexts (“Acme - accounting software for freelancers”), the model learned you as an entity with a category attached, and it can recommend you even with web search off.

You can’t edit this memory directly, and it lags reality by months or years. New brands start invisible here; renamed brands live under their old name for a long time. What slowly writes you into future training runs is the same public footprint that helps everywhere else: consistent descriptions across many credible pages.

Layer two: what the model retrieves

With web search on, the chatbot pulls a handful of live pages before answering and builds the answer largely from them. Each engine reads a different back-end: Gemini, AI Overviews and AI Mode use Google Search, Perplexity its own crawler, and ChatGPT a mix that includes Bing and its own crawler. Which index sits behind ChatGPT is less obvious than it looks; the evidence is laid out in does ChatGPT use Google or Bing.

This is the layer you can influence within weeks, and it explains most short-term wins: if the retrieved pages for “best X for Y” include a ranking that lists you, you have a real chance of being named; if they don’t, no product quality will put you in that answer.

Layer three: what the model knows about the person asking

The third input is the conversation itself. Earlier messages in the same chat, and in ChatGPT the saved memories and chat history when personalization is on, can tilt the answer: a user who mentioned a tight budget or a specific country will usually get a different shortlist from someone asking cold.

You can’t optimise for this layer, but you should know it exists. It is one reason two people checking “does ChatGPT recommend us?” on their own accounts can see different results, and why a fair test runs in a fresh session without personal context.

How does the type of question change the shortlist?

The same brand can win one phrasing and vanish from the next, because each question pulls different sources. These are tendencies, not rules, but they explain a lot of “we show up here and not there”:

• Category discovery (“best CRM for small agencies”) tends to lean on rankings, round-ups and comparison articles.
• Constrained buying (“CRM under $30 a month with a Polish invoice”) favours pages that state the constraint outright: pricing pages, feature tables, detailed reviews.
• Head-to-head (“HubSpot vs Pipedrive”) draws on comparison pages, review platforms and the vendors’ own pages.
• Problem-first (“how do I stop losing leads in email”) pulls in guides and how-tos, and brands get named only if those guides mention them.

The signals that decide the shortlist

• Presence in retrieved sources: rankings, comparisons, directories and reviews that answer the buyer’s exact question. The single strongest lever.
• Consistency of description: the same name, category and claim across your site and third-party pages. Contradictions dilute the entity.
• Corroboration: a claim repeated by several independent sources beats a claim your own site makes alone.
• Tone of mentions: being described as the recommended option in an expert article usually counts for more than a pile of passing or negative mentions.
• Quotability: pages that answer the question in the first paragraph, with structure the model can lift almost verbatim. In the paper that coined “generative engine optimization” (Aggarwal et al., KDD 2024), adding quotations, statistics and cited sources raised a page’s visibility in generated answers by up to 40%, with large differences between topics.
• Indexability where it counts: Bing and OpenAI’s own crawler for ChatGPT, Google for Gemini, AI Overviews and AI Mode, open access for Perplexity’s crawler.
• Recency: engines lean toward fresh sources; a 2023 ranking fades against one that is kept up to date in 2026.

Why does the list change every time you ask?

Because the model samples its answer instead of looking it up. In a study published by SparkToro in January 2026, 600 volunteers ran 12 recommendation prompts through ChatGPT, Claude and Google’s AI 2,961 times between November and December 2025. The chance of getting the same list of brands twice was below 1 in 100, and the same list in the same order closer to 1 in 1,000. The authors flag their own limits: they are not professional researchers, the work was not peer-reviewed and it was run with an AI-tracking startup.

Their practical conclusion is worth keeping: a single answer, or a “position” in it, tells you little, but how often a brand appears across dozens or hundreds of runs is a reasonable metric. That is the number to watch.

What this means for your brand

Three practical consequences. First, recommendations are earned in your category’s sources, not on your homepage - find which pages engines actually cite for your buyers’ questions and get present there. Second, the retrieval layer moves in weeks and the memory layer in months or years, so judge a change by the right clock. Third, the process is measurable: ask the engines your buyers’ questions on a schedule, track whether you’re mentioned and which sources are cited, and treat every change you ship as an experiment against that baseline.

That measurement loop - question, answer, mention, sources, repeat - is what CiteLyzer automates across ChatGPT, Gemini, Perplexity, AI Overviews and AI Mode. However you run it, run it: the shortlist is being decided daily whether you watch it or not.