How AI Assistants Decide Which Brands to Recommend
What we know, and what we don't, about how ChatGPT, Gemini, Perplexity and Claude pick the brands they recommend, and what that means for marketers.
When someone asks an AI assistant for "the best project management tool for a small agency", a short list of brands comes back. Marketers understandably want to know how that list is made. The honest answer is that the exact mechanics are proprietary, but enough is public to build a useful mental model.
Two sources of knowledge
A modern assistant draws on two kinds of knowledge when it answers:
- What the model learned in training. Large language models are trained on large amounts of text. Brands that were widely and consistently discussed before the training cutoff are "known" to the model.
- What it retrieves at answer time. With web search or grounding enabled, the assistant runs searches, reads some of the results and uses them to compose the answer, often citing them.
For questions about products and current options, assistants increasingly lean on retrieval, because the answer depends on up-to-date information such as prices and features. That is good news: retrieval is something you can influence through the open web.
What retrieval tends to favour
Without inside knowledge we can only describe patterns that follow from how retrieval works:
- Pages that can be fetched. If a crawler is blocked, or your content only renders with JavaScript, it cannot be used.
- Pages that match the question. A page titled "Best invoicing tools for freelancers (2026)" is a natural match for that question; a generic homepage is not.
- Sources that already rank. Many assistants use a traditional search index for retrieval, so pages that are visible in search are more likely to be read.
- Clear, extractable statements. Answers are assembled from passages. Content with explicit facts, comparisons and lists is easy to summarise accurately.
Why the same question gets different answers
Generation involves sampling, so wording varies run to run. Retrieval results change as the web changes. Engines are updated often. And consumer apps may personalise answers based on a user's history or location. This is why a single screenshot is weak evidence, and why we recommend measuring trends across many prompts and repeated runs. Our guide to answer variability goes deeper.
What brands can do
- Be retrievable. Allow AI search crawlers, server-render important content, maintain a sitemap.
- Be explicit. Say plainly what you do, for whom, at what price, and how you differ from alternatives.
- Be present where answers come from. Find which domains assistants cite in your category, then earn honest inclusion there: reviews from real customers, accurate listings, expert commentary.
- Be consistent. Conflicting facts across the web make it harder for any system to describe you correctly.
- Measure, then iterate. Track a stable prompt set over time and look for persistent changes.
What not to do
Tactics that try to trick models, such as hidden text, fake reviews or spammy content networks, risk your reputation with both search engines and AI providers, and they don't build the kind of genuine presence that lasts. The durable strategy is the same one that has always worked on the web: be useful, be clear and be talked about by others for good reasons.