How do AI assistants decide which products to recommend?

The recommendations are organic, which is unusual and temporary

ChatGPT product results are currently not ad-driven. Placements are organic and unsponsored, ranked on relevance to what the shopper asked. That is a genuinely different market structure from paid search, where the top of the page has been purchasable for two decades.

It is worth treating this as a window rather than a permanent state. Every previous discovery surface that started organic ended up with an advertising layer, and there is no reason to assume this one is exempt. The practical implication is that authority accumulated now is cheaper than authority bought later.

The four signals

These are the inputs AI search systems weigh when deciding which brands to surface. They are listed in roughly the order a merchant can act on them.

The uncomfortable finding

Here is the part that most guidance on this topic leaves out, and it changes what a sensible plan looks like.

Brand-owned sites account for roughly 5–10% of the sources AI search draws on. Affiliate content, publishers and user-generated content make up over 65% in key categories. So a merchant who perfects their own domain — flawless schema, server-rendered everything, an answer hub, an llms.txt — has addressed the minority of what determines whether they get recommended.

This does not make the on-site work pointless. It is the part you fully control, it is the part that has to be right before third-party authority helps you, and it is what an assistant verifies against once a third-party source has surfaced you. But treating it as sufficient is the mistake, and any vendor telling you otherwise is describing the half of the problem they can sell you.

We are aware this argues against the obvious reading of a page like this one. It is still the accurate answer.

What that implies for where effort goes

Given the 5–10% versus 65% split, a defensible allocation looks quite different from a conventional SEO plan.

• Fix the on-site layer first, because it is a prerequisite rather than a lever. Structured product data with real values, server-rendered, is table stakes — a third-party mention that leads an assistant to an unparseable page wastes the mention.
• Then invest disproportionately in being reviewed, covered and discussed. Product reviews on third-party sites, coverage in publications your category actually reads, and genuine presence in the communities where your buyers compare options.
• Make your facts easy to restate consistently. When your positioning, pricing model and category are stated the same way everywhere, an assistant has a confident fact to repeat. Inconsistency is resolved by omission far more often than by picking a side.
• Do not fabricate the third-party layer. Planted reviews and paid placements dressed as editorial are detectable, and the downside is a brand an assistant learns to distrust — a much worse position than being unknown.

The demand this is competing for

For scale: eMarketer forecasts 31% of Americans will use generative AI for search in 2026, and AI search is reported to influence over 40% of product discovery. Traffic arriving from AI assistants has been measured converting at 5.53% against 3.7% for organic search.

Those are early-market numbers against a small base and deserve the usual scepticism about percentage growth on small denominators. The direction, though, has been consistent across independent measurements for over a year now, which is more than can be said for most channel forecasts.