How AI assistants choose which brands to recommend

Every assistant has two ways of producing an answer. The first is training memory, sometimes called parametric knowledge. During training, the model absorbed a snapshot of the web, and it can answer from that snapshot without fetching anything. This mode is fast and cheap for the vendor, so engines use it whenever they judge a question does not need fresh information. The drawback is that the snapshot is months or years old, and the model reconstructs facts statistically rather than looking them up, so details drift. In our July 2026 test of 30 brands, 27 were described with at least one materially wrong claim when the AI answered without live web search (full study).

The second mode is live retrieval. The assistant runs one or more web searches, reads the top results, and writes its answer from what it just read. Recommendations in this mode are largely determined by which pages the underlying search index returns and how clearly those pages state their claims. Freshness improves, but a new dependency appears: if the pages that rank for your buying queries never mention you, the assistant has no material to recommend you with.

Most commercial questions ("best CRM for a small charity", "alternatives to Xero") now trigger retrieval on most engines, but not reliably, and users cannot always tell which mode produced the answer in front of them.

Whichever mode is active, the selection logic is similar: the model favours brands that many independent sources agree on. In training memory, that agreement was learned statistically; a brand mentioned consistently across thousands of pages becomes the high-probability answer for its category. In retrieval, the model cross-references the handful of pages it fetched and leans towards names that appear in several of them with compatible descriptions.

This is why a single excellent article rarely changes anything on its own, and why contradictory information actively hurts. If your pricing page says one thing, a review site says another and an old directory listing says a third, the model either hedges or picks a competitor whose story is consistent. Entity consistency, meaning the same name, category and core claims everywhere you are mentioned, is worth more than any individual placement.

Your website is one voice with an obvious motive. A model that treated vendor copy as authoritative would end up recommending whoever wrote the boldest homepage, so engines appear to discount it. The retrieval layer discounts it structurally too: the pages that rank for "best X" queries are comparison listicles, review platforms, forum threads and industry publications, not vendor homepages. When an assistant reads its retrieved set, most of what it learns about you was written by someone else.

Reddit deserves a specific mention. Threads are candid and specific, and they contain the sort of comparative judgement ("we switched from A to B because...") that models find genuinely useful. Several engines also have licensing or crawling arrangements that surface Reddit heavily. That candour cuts both ways: a well-upvoted complaint can follow you into answers for a long time.

Your own site still matters, just later in the process. Once you are shortlisted, assistants pull specifics such as pricing, integrations and limits from your pages. If that content is unclear or locked behind JavaScript, the specifics get guessed instead.

A caveat first: no vendor publishes how sources are weighted, so everything below is inference from documentation and repeated testing, not confirmed mechanics.

ChatGPT appears to answer from memory more readily than its rivals on broad questions, switching to search for fresh or niche topics. When it does search, its retrieved set has historically drawn on Bing's index, so visibility there seems to matter.

Perplexity is retrieval-first by design and cites almost everything, which makes it the most transparent engine to test against: if you are absent from its answers, you are usually absent from the pages it retrieved.

Gemini sits on Google's search infrastructure, and the overlap with conventional Google rankings looks strong in practice, though Google has not confirmed how directly the two connect.

Claude searches the web when a question needs current information and tends to hedge more when answering from memory, which reduces confident errors but can also mean it declines to name brands at all.

The practical upshot is that the same question can produce four different shortlists. Testing each engine separately is the only way to know where you actually stand.

This work has a name: generative engine optimisation, or GEO (what is GEO). The order below reflects effort against likely impact.

1. Establish a baseline. Run your brand and your main buying questions through our free AI visibility checker. It currently tests Claude live with web grounding, with ChatGPT, Perplexity, Gemini and Copilot support rolling out, and it shows you what an assistant actually says about you today. Re-run it manually after each round of changes; there is no scheduled monitoring, so the cadence is yours to set.

2. Fix contradictions about your brand wherever they exist: old directory entries, outdated review-site profiles, and stale press mentions you can get corrected. Consistency is the cheapest win available.

3. Get onto the third-party pages that already rank for your buying queries. Search your own "best X for Y" phrases, list the top ten results, and pursue inclusion in the ones that accept submissions or briefings. This step does more for retrieval-mode visibility than anything on your own domain.

4. Build genuine review and community presence. Encourage real customers to review you on the platforms your buyers cite, and participate honestly where your category is discussed. Astroturfed posts tend to get spotted and quoted back as a negative, so do not fake it.

5. Make your own site quotable: plain statements of what you do and for whom, visible pricing, server-rendered text, and structured data. This is about being extractable once you are shortlisted, not about ranking.

6. Re-test and iterate. Model updates and index changes shift answers without warning, so treat visibility as something you re-measure after every meaningful change rather than something you fix once.