As holiday shopping ramps up, a new test offers a glimpse of how AI assistants can turn a shopper’s question into a short list of products. In this case, the assistants kept returning to a small group of running-shoe brands. That’s worth a look if you sell online, but it isn’t a forecast of what agents will recommend across the holiday season.
The test, published by Recommended by Agents, covers one category, one price range, and a limited set of questions. Treat it as an early benchmark: useful for asking better questions about your product information, not proof that AI assistants have settled on a fixed roster of winners.
Five agents, 40 questions, one narrow category
Recommended by Agents asked five assistants the same 40 questions about U.S. running shoes priced under $150. The questions covered different shopper needs, including shoes for wide feet, trail running, and daily training. Each assistant answered through its developer API, producing 200 answers in total. The results were collected on September 28, 2026.
The assistants were ChatGPT, Gemini, Claude, Meta’s Muse, and Perplexity. Across their answers, seven brands accounted for 83% of recommendations. Each answer named about four brands on average. That gives you a sense of how short an assistant’s shortlist can be: a lot of other products won’t make it into the answer at all.
Brooks led the results, but there’s a small discrepancy worth keeping in view. Shopifreaks’ October 22 coverage put Brooks in 67% of answers. The published results on Recommended by Agents now list 68%. Until the publisher clarifies the difference, it’s best to describe Brooks as appearing in roughly two-thirds of answers rather than treating either figure as exact.
There’s also useful context for evaluating the report: Recommended by Agents offers services to help online stores prepare for AI agents. That doesn’t make the results invalid, but it’s relevant background when you’re weighing a study from a company that works in the area it measures.
A useful snapshot, not a market-wide verdict
The most important limit is simple: each question was asked once per assistant. That’s 200 answers, but no repeated runs to show how much the recommendations might change if the same questions were asked again. Recommended by Agents says small gaps between brands may be noise. That’s a sensible warning when you’re looking at a single round of results.
The answers also varied by assistant. Gemini recommended HOKA in 65% of its answers, while ChatGPT did so in 13%. Puma appeared in 35% of ChatGPT’s answers and none of Perplexity’s. Those differences show why one assistant’s shortlist can’t stand in for every assistant’s, even when they’re answering the same questions.
And this was a test of recommendations, not a record of completed purchases. It doesn’t tell us which shoes shoppers bought, which recommendations led to sales, or how the results compare with people searching on a retailer’s site. It also can’t establish how assistants will respond in other categories, to other prompts, on consumer apps rather than APIs, or during the holiday season as a whole.
The test’s price cap matters, too. A product that doesn’t fit the stated budget may be left out for that reason, even if the brand is well known. Change the price, category, or shopper’s needs and the shortlist may change with it. A single snapshot can show what happened under those conditions, not what will happen everywhere.
What sellers can take from it
Use the results as a reason to review whether a shopper, or an assistant answering for one, can quickly find the facts about your products. Check that price, availability, sizes, materials, key features, and descriptions are clear and consistent wherever they appear. If a product page says one thing and a feed or marketplace listing says another, you’ve created a question the shopper shouldn’t have to solve.
That’s practical housekeeping, not a claim that the test identified why any brand was recommended. It didn’t test whether a particular product attribute, page format, or source caused an assistant to name one brand over another. So don’t treat the results as a recipe for getting recommended. They’re a prompt to check the basics and find gaps you can fix.
For marketplace sellers, that means reviewing the information shoppers and assistants can access on your product detail pages, including titles, bullets, descriptions, and catalog attributes. For brands selling through their own stores, review the product page and the information shared with shopping platforms. Clear facts help humans compare products, too. That’s a worthwhile outcome even when an assistant’s recommendation remains unpredictable.
The broader question is what information different agents can find and read. Parker-Lambert’s guide to which parts of the web AI agents read explores why product details across your own pages and other sources may matter. The running-shoe test doesn’t settle that question, but it gives sellers a concrete example of how quickly recommendations can narrow.
What to watch next
The next useful benchmark would repeat questions across multiple runs, test more categories, and show whether the same patterns hold over time. Results from different assistants are worth tracking separately, too. A brand that appears often in one assistant’s answers may be much less visible in another’s.
Until broader tests arrive, don’t overhaul your strategy based on a single category and one round of answers. Use the report to ask whether your product details are complete, easy to compare, and consistent across the places an assistant might consult. Keep watching for repeated findings before drawing bigger conclusions. The shortlist may be short; the evidence behind a big business decision shouldn’t be.
Tell Parker-Lambert about your brand and ask about a custom quote for your project.