How AI decides which products to recommend

When a shopper asks ChatGPT or Perplexity "what's the best running belt for long trail runs," the assistant comes back with a short, confident list of three or four products. No ten blue links, no scrolling, just names and a reason for each. To anyone watching their own brand not get named, it can feel like there's a secret ranking somewhere, and if you could just find the lever, you'd climb it.

There isn't a lever. There isn't a single ranking. The most useful mental model is that AI assistants are doing two things at once: judging your reputation (what the wider web seems to think of you) and your readability (how clearly your product can be understood and matched to a question). Get both right and you become an easy, safe thing for an assistant to name. Let me walk through the signals that actually feed into that.

It's a blend of signals, not a score

The first thing to unlearn is the idea of a single number you can optimize. With classic Google SEO there's at least the comforting fiction of a position: you're number four, you want to be number two. AI product recommendations don't work like that. The assistant is assembling an answer on the fly, weighing several loosely-related things, and the same question asked twice can come back slightly differently.

So instead of chasing one metric, it helps to understand the handful of inputs that consistently seem to matter. None of these has a published weight, and anyone who quotes you a precise percentage is guessing. But the categories themselves are real and observable.

Reputation: how much the web talks about you

This is the big one, and it's the part most store owners underestimate. AI models learn an enormous amount from text written by other people: product reviews, Reddit threads, forum answers, "best of" roundups, blog comparisons, newsletter mentions. If lots of independent sources describe your product, and describe it positively, the model has been "told" many times that you're a real, credible option in your category.

Two things matter here, and they pull in the same direction:

• How often you're mentioned. A brand that appears in twenty independent writeups is far easier for an assistant to recommend than one that appears nowhere outside its own site.
• How positively. Sentiment matters. Glowing reviews and "I switched to this and never looked back" threads do more than a bare directory listing.

Notice what's missing from that list: your own marketing copy. The model knows you wrote your product page, so it weighs it lightly. The web's collective, independent opinion of you is the real input. This is why reviews quietly shape what AI recommends and why getting onto a credible third-party list is some of the highest-leverage work you can do. When a respected site publishes "best trail running belts" and you're on it, you've handed the assistant exactly the kind of source it trusts.

AI recommends what it has heard about a lot, from places it trusts. Your reputation across the web is, in a very real sense, the thing being ranked.

Source trust isn't equal

Not every mention counts the same. A thoughtful review on a well-known publication or an active discussion among real users carries more weight than a thin page that exists only to host affiliate links. Models lean toward sources that seem established and genuine. You can't control how much any one source is trusted, but you can aim your effort at the credible ones rather than chasing volume on low-quality directories.

Readability: how clearly AI can understand your product

Reputation gets you into the conversation. Readability decides whether the assistant can confidently match you to a specific question. Even a well-regarded product gets skipped if the model can't tell what it is, who it's for, or where it ships.

Clear, structured product data

The assistant has to understand what you sell before it can recommend it. Vague descriptions, missing specs, and buried details all make its job harder. Plain language, obvious categories, real specs, and structured product data all help an assistant parse your page and feel sure enough to name you. If your page makes a human squint to figure out what the thing actually does, the model squints too.

Relevance to the exact question

Shoppers ask narrow questions. "Best running belt" and "best running belt that holds a large phone and doesn't bounce" are different queries, and the second one rewards the product whose data actually mentions phone size and a no-bounce fit. The closer your information maps to the specific thing being asked, the more likely you are to be the answer. Generic pages win generic questions, which barely exist anymore.

Availability

An assistant trying to be helpful is reluctant to recommend something a shopper can't actually buy. If a product is out of stock, or clearly unavailable in the shopper's country, that's a strike against naming it. Keeping your popular items in stock and your availability clear isn't just good operations, it's a recommendation signal. We dug into the stock side of this in how out-of-stock products affect AI visibility.

Letting the crawlers in

All of the above assumes AI systems can read your site in the first place. A surprising number of stores quietly block AI bots, sometimes from a line a developer added months ago, sometimes a plugin default. If the crawlers can't reach you, none of your reputation or readability work can be seen. It's worth checking that you're not accidentally blocking AI crawlers before anything else.

Why two stores with the same product get different answers

Put the pieces together and the unfairness starts to make sense. Take two shops selling a near-identical product. One has dozens of genuine reviews, a couple of "best of" placements, crisp specs, and reliable stock. The other has a prettier website and almost no independent footprint. The first gets named again and again; the second is invisible. The product quality might be identical. The recommendability is not.

That's also why you can't pay your way in. There's no ad slot inside an assistant's answer and no submission form that bumps you up. What you can do is improve the raw material: earn more honest mentions, make your data unmistakable, and stay in stock. It's slower than buying a placement, but it compounds, and it's real.

What to actually do with this

Start by seeing where you stand today rather than guessing. Ask the assistants the questions your customers would ask and note who gets named. The competitors that keep appearing are showing you the reputation and readability you're missing. From there the work splits cleanly along the two themes: earn more independent mentions and reviews to build reputation, and tighten your product data, schema, and availability to build readability.

If checking by hand across several assistants and a pile of questions sounds tedious, it is, and the answers shift from run to run. You can run a free AI visibility audit and we'll look across ChatGPT, Perplexity, Gemini, and Google AI for you, including which competitors are getting named instead. Either way, the takeaway is the same: there's no single algorithm to game. There's a reputation to earn and a product to make legible. Do both, patiently, and you stop being the brand the assistant has never heard of.