A study of 9,720 ecommerce stores by AI commerce company Recomaze ran 58,320 product queries through Google Gemini. Six in ten stores were recommended for nothing, and the recommendations that did happen scattered across more than 50,000 brands.
Shoppers are starting to ask AI assistants what to buy, not just how to spell a word or what to cook for dinner. A new study suggests that when they do, most online stores are nowhere in the answer.
The research, from AI commerce company Recomaze, ran six purchase-intent queries against each of 9,720 ecommerce stores through Google Gemini, 58,320 tests in all. In 60% of cases the store was recommended for nothing. Across every test, a store was named just 14% of the time. The rest of the time the assistant pointed the shopper to a competitor, or named no store at all.
The question is becoming less abstract by the month, as Amazon, Google and others push AI assistants deeper into how people find and buy things. When the answer to “what should I buy” is a short list of names rather than a page of links, being one of those names is the new version of ranking on the first page. The study’s argument is that most catalogues were written for human readers and for Google, not for an engine that has to read a product, understand it, and recommend it on a shopper’s behalf.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
The more surprising result is who wins when a store does get named. The recommendations did not concentrate among a handful of giants. They spread across 50,287 distinct brands. The ten most-recommended brands accounted for roughly 4% of all recommendations, the top 100 for about 11%, and the single most-recommended store, Etsy, appeared in just 1.3% of tests. In one case from the scan, a query for cocktail capsules resolved to Bartesian, a focused brand, recommended ahead of Walmart and Target. The pattern, in the study’s reading, is that AI product discovery is a long-tail game, and a smaller store is not automatically shut out by scale.
Category was the strongest predictor of whether a store was seen at all. The hardest were visually driven: in home and living, around 74% of stores were invisible, and in apparel about 67%. The most visible was food and beverage, where roughly 52% were invisible. A text-based engine struggles to judge the look that drives a clothing or homeware purchase, so it falls back on a few large generic retailers, while products with clear, describable attributes are easier for it to match.
The study is direct about its limits. It is a single scan per store rather than an average, Google Gemini only rather than ChatGPT or Perplexity, with queries generated algorithmically and categories assigned by classifying query language, leaving 42% of stores uncategorised. The stores were drawn from BuiltWith, and the queries were written to match each store’s category, the questions a shopper asks when they know what they want but not where to buy it, rather than brand-name searches that would be trivial to win.
“The shop window used to be Google. Now it is whatever the AI decides to recommend, and most brands have no idea whether they are in the answer,” said Delian Coroamă, founder and chief executive of Recomaze. “The stores that get recommended are not the biggest ones. They are the ones whose product information an engine can actually read and trust.”
Recomaze has a commercial stake in the problem it measured: it sells tools that track where a store appears across ChatGPT, Gemini and Perplexity and rewrite catalogue data so the engines can recommend it. But the underlying finding is harder to wave away. A new layer now sits between shoppers and stores, it decides which names get spoken aloud, and on the evidence so far it stays silent about most of them.
Get the most important tech news in your inbox each week.