How AI shopping assistants actually choose which products to recommend

Picture a shopper typing this into ChatGPT: "what's a good fragrance-free candle for a small apartment, under $30?" A few of these turn into a specific recommendation — sometimes with a link. The question every store owner should ask is: what decided which products got named?

It isn't magic, and it isn't luck. Here's the chain of events, and where you can actually influence it.

Step 1: the assistant turns a sentence into intent

The first thing the model does is read the shopper's natural-language question and pull out the constraints: a candle, fragrance-free, suited to a small space, under $30, probably a gift or for personal calm. None of those are keywords in the old SEO sense — they're attributes and use-cases.

This is the part most listings are blind to. A title like "The Aurora" describes none of those constraints. A title like "Hand-poured soy candle — fragrance-free, 8 oz, gift for a new home" describes almost all of them.

Step 2: it looks for products whose data matches

The assistant then tries to match that intent against product information it can access — from its training, from connected shopping sources, and increasingly from live retrieval of pages and feeds. What it's matching against is your structured-ish product data: titles, descriptions, attributes, and any machine-readable schema you've published.

Two things matter enormously here:

• Specificity. "Smells nice" matches nothing. "Unscented, clean-burning soy wax" matches a fragrance-free query directly.
• Coverage. If a shopper's likely question mentions materials, size, who it's for, or how it's used, those words need to exist somewhere in your listing. The assistant can't infer a fact you never stated.

Step 3: it decides what's safe to recommend

Even among matching products, an assistant won't surface something it can't describe confidently. Models are tuned to avoid making claims they can't support. So a listing that clearly states real, checkable facts is safer to recommend than one that's vague or padded with marketing adjectives. Clear beats clever.

This is also why a short, honest FAQ is so powerful: it pre-answers the exact follow-up questions a shopper would ask ("is it scented?", "how big is it?", "is it a good gift?") in a form the assistant can lift directly.

Step 4: it writes an answer, not a list of links

Here's the biggest shift from classic search. The assistant doesn't return ten blue links for the shopper to scroll — it writes one answer and names a handful of options. If you're not in that answer, there's no second page to be on. You were either legible enough to be included, or you weren't.

What this means for your listings

You can't control the model. You can control whether your data gives it something to match and something safe to repeat. In practice that's five things:

1. A title that says what the product is and who it's for — not just a brand name.
2. A description that leads with facts: material, size, two or three real use-cases.
3. The keywords shoppers actually say — the materials, occasions and audiences they'd mention to an AI.
4. A short FAQ answering the obvious follow-ups, in plain language.
5. Structured data (Product + FAQ schema) so the facts are machine-readable.

None of this guarantees you'll be recommended — no one can promise that, and you should be wary of anyone who does. What it does is make your products eligible to be recommended, which is the entire game.

See where you stand

The fastest way to find out whether your listings give an assistant anything to work with is to score them. Essio runs a free AI-visibility audit on your products — no card needed — and shows, per product, exactly which of the five it's missing. If you want the deeper version of the mechanics above, read how Essio works.