How do AI assistants decide which brands to recommend?
By Ben Frampton · · 5 min read
Short answer: an AI assistant recommends a brand when two things line up. The model already associates the brand with the question (from its training data), and the pages it retrieves while answering mention the brand favourably. You can influence the second directly and the first slowly, by being described clearly and consistently in the places the web talks about your category.
The rest of this post explains both halves, and why the same question can get a different list of brands each time you ask it.
Where does an AI assistant's answer come from?
Every answer an assistant gives about your market draws on some mix of two sources.
| Source | What it is | How fast it changes | What you can do |
|---|---|---|---|
| Training data | Public text the model learned from before its cut-off date | Months to years, when a new model ships | Be discussed often and consistently across the web |
| Live retrieval | Pages the assistant finds by searching while it answers | Days to weeks | Rank for the questions buyers ask, and be mentioned on the pages that rank |
Which source dominates depends on the assistant and the question. Perplexity and Google's AI Overviews and AI Mode search for nearly every answer. ChatGPT, Gemini, Copilot and Claude search when the question seems to need fresh or specific information, and otherwise answer from what the model already knows.
How does training data shape a recommendation?
A language model has no list of approved brands. It has a statistical picture of which names tend to appear near which ideas. If "affordable", "small teams" and your brand name often appear together in reviews, forum threads, comparison articles and your own pages, the model learns that association.
That has three consequences:
- Volume matters, but so does consistency. A brand mentioned a thousand times in conflicting ways gives the model a blurry picture. A brand described the same way in a few hundred places gives it a sharp one.
- The cut-off date matters. A model trained before your rebrand or new product line will describe the old version until a newer model replaces it.
- You cannot edit the training data directly. You can only change what the web says about you and wait for the next model.
How does live retrieval shape a recommendation?
When an assistant searches, it usually:
- Rewrites your question into one or more search queries. Google calls this "query fan-out" for AI Mode: one question becomes several related searches.
- Fetches the top results for those queries.
- Reads those pages and writes an answer, often citing some of them.
The brands that appear in the answer are mostly the brands that appear on the retrieved pages. If the top results for "best CRM for estate agents" are three listicles and a Reddit thread, the brands named in those four pages are the strongest candidates for the answer.
This is why classic search ranking still matters for AI visibility. Retrieval rank (whether your page, or a page that mentions you, shows up for the queries the assistant runs) is the best-supported predictor of being cited.
Why does the same question get different brands each time?
Assistants generate text by sampling: at each step they pick among several likely next words, not always the single most likely one. Add different search results from one day to the next, different phrasings from different users and different models across products, and the answer moves around a lot.
In practice this means:
- One answer is an anecdote, not a measurement.
- A brand can appear in 30% of answers to a question and still be missing from the one you happened to see.
- The useful number is how often you are mentioned across many answers, with a margin of error. We explain how to read that in why one AI answer tells you almost nothing.
What makes a brand more likely to be recommended?
From what can be observed across many answers, brands that get recommended tend to share these traits:
- A clear, repeated positioning. The site, review profiles and press all say the same thing about who the product is for.
- Pages that answer buyer questions directly. Not a feature list, but "Which X is best for Y?" answered honestly near the top of a page. See how to write pages AI assistants can quote.
- Presence on the sources assistants cite. Review sites, independent comparisons, community threads and trade press in the category.
- Accessible pages. Crawlers for search and AI can fetch and read the content. See which AI crawlers to allow in robots.txt.
None of these is a trick. They are what makes a brand easy to understand and easy to vouch for, whether the reader is a person or a model.
Frequently asked questions
Can I pay to be recommended by ChatGPT or Gemini?
Not in the organic answer. Some assistants have started testing clearly labelled ads beside or below answers, but those are separate from the recommendation itself. The organic answer comes from training data and retrieved pages.
Do AI assistants favour big brands?
They favour brands that are discussed a lot, which often means big brands. But for specific questions ("best accounting software for freelance photographers") a smaller brand that is clearly associated with that niche can appear ahead of larger generalists.
How long does it take to change what AI assistants say about my brand?
Changes that work through live retrieval (new pages ranking, new mentions on cited sites) can show up within weeks. Changes that depend on training data take until the next model version, which can be many months.
Is getting recommended by AI the same as ranking on Google?
No, but they are linked. Ranking helps you get retrieved, and being retrieved helps you get cited. The difference is that the assistant then decides which brands to name, and that decision also depends on how the pages describe you. Our introduction to GEO and how it differs from SEO covers this in more detail.