How we measure what AI shopping agents see

Every reading comes live from shop.dev, the product catalog AI shopping agents search when a shopper asks them for something. We see what an agent sees: the products it finds, the summaries and ratings it reads, and the order they come back in.

We measure what agents receive, not what a particular assistant finally says. That's what a brand can change, and it's the same for every agent that searches the catalog.

We search signed out. Signed-out results are the same for everyone, so they're the baseline every shopper starts from, and a fair way to compare one day with the next.

When a shopper is signed in with Shop, the catalog ranks results by their preferences and shopping history. Two shoppers asking the same question can see different products.

That ranking can only work with what your listings say. If your titles, descriptions, sizes, materials and ratings are thin, the agent can't tell that your product is the right one for a shopper who should want it. That's why every report checks how complete your product data is, and why it counts for a fifth of your Agent score.

We read your catalog and how other sellers describe the same kinds of products, and an AI model proposes what a shopper who doesn't know you would type: topic searches. We add brand searches, where shoppers ask for you by name. Every proposed search is then run against the catalog and checked: it has to return products like yours and use words other sellers use, so you're never measured on a search you'd win by definition.

You remove the ones that don't fit and add the ones your customers use before anything runs.

Once your first report runs, the searches stay the same, so a rank today can be compared with a rank last month.

The Agent score runs from 0 to 100. Being found in topic searches counts for 45%, your ratings and reviews for 20%, how complete your product data is for 20%, and being found when shoppers ask for you by name for 15%.

Higher places count for much more than lower ones, because an agent shows a shopper only a few options.

Agents change their picks from day to day. Each tracked brand gets a new reading every day, a trend line, and a list of what changed. Small day-to-day wobbles are averaged out; a big move is searched again before it's reported, so a single odd result doesn't read as a change.

A reading where the catalog doesn't answer is thrown away, never recorded as a drop.

Every report ends with a ranked list of fixes. For each search you're losing, we take the product of yours that should be winning it and check what its listing tells agents. An AI model then writes the edit: a clearer title, a first sentence for the description, and the details shoppers ask about. It works only from what your listing says and leaves anything it can't know, like a measurement, in [brackets] for you to fill in.

Each fix also names what we measured and, where we can fairly estimate it, how many points it's worth. We can't promise how agents will rank you; we can show you where you stand and what's holding you back.

Your report is about your brand alone. You see your rank among sellers and your share of attention, but never who the other sellers are, and no other brand ever sees yours.