I build Geomint, a tool that asks Google’s AI the questions a shopper would ask and checks whether a given store comes up in the answer. Between September and early October 2026 I ran it on 42 independent Shopify stores. Most were stores I picked to contact about the tool. A few belonged to merchants I’d seen posting on the Shopify Community forum.
This is what came back, including the parts that made my own tool look bad. I’ve written down exactly what I left out and why, because a study you can’t check is just an ad.
What I tested
For each store, the tool reads the public product catalog that every Shopify store publishes, works out the main things it sells, and writes 12 questions. Ten are about the category, without the store’s name. Two ask about the store directly. These are real examples from the run, quoted exactly:
- “What are the best men’s knitwear brands in 2026?”
- “Compare top brands of trousers”
- “Top-rated t-shirt for everyday use”
- “Recommend a good wet cat food for someone on a budget”
- “I’m looking for dog food that ships fast — any recommendations?”
- “Is [store] a good brand for [category]?”
Each question went once to Gemini 2.5 Flash through Google’s API. Gemini is the model family behind Google’s AI Overviews, but the API is not the same as a live AI Overview. More on that under limits.
The stores were mostly apparel, plus coffee, pet food, jewellery, home textiles and a few others, based in the UK, the US and Europe. It’s not a random sample of Shopify. It leans towards small brands that make or design their own products, which are exactly the stores this question matters most to.
What I left out, and why
- Five whole audits. The tool misread what those stores sell and asked about the wrong products, like asking a water-bottle brand about “canvas”. Their answers say nothing about the stores, so they’re out. The question generator has been rewritten since.
- Answers that were cut off. About a quarter of Gemini’s answers stopped mid-sentence. The cause was a bug in my tool: the model spent part of its word allowance on internal reasoning before it started writing. It’s fixed now. I only counted answers that finished, which left 319 of the 400 category answers and 47 of the 80 brand questions. A cut-off answer might have named the store further down, so read the numbers as close rather than exact.
- Four false “named” results. My tool originally counted five category answers as naming the store. When I read them, four had named a brand the store sells, such as Chanel for a perfume retailer, rather than the store itself. A shopper told to “try Chanel” isn’t being sent to that shop, so I counted those as misses.
1. If the shopper doesn’t say your name, you’re almost never in the answer
Across 319 complete answers to category questions, one named the store it was about. It was an activewear brand, mentioned fourth in a list after Nike, Lululemon and Vuori, for a question about interval shorts.
The other 41 stores didn’t come up once. These weren’t dead shops. They were trading stores, with real products and real customers. None of that showed up in the answers.
2. The spots went to the same big names, whatever the category
When an answer named brands, it named three or four of them. Across all the answers, 309 different brands came up, but a handful dominated:
| Named | Times | For how many of the 42 stores |
|---|---|---|
| Nordstrom | 63 | 28 |
| Uniqlo | 56 | 18 |
| Target | 45 | 28 |
| Everlane | 45 | 14 |
| Amazon | 42 | 29 |
| H&M | 39 | 22 |
| Old Navy | 33 | 20 |
| Zara | 22 | 13 |
| ASOS | 22 | 13 |
| Nike | 21 | 9 |
Nordstrom came up in the answers for 28 of the 42 stores, across linen dresses, bags, jeans and jewellery. Amazon came up for 29. More than half of all answers (170 of 319) sent the shopper to a big retailer or marketplace rather than to a brand. For a small brand, the competition in an AI answer isn’t only the brand next to you. It’s the department store.
3. Asked about the store by name, Gemini often didn’t know it
The two brand questions should be the easy ones. A shopper who has already heard of you asks whether you’re any good. In 23 of 47 complete answers, Gemini said it didn’t recognise the store or couldn’t speak to its quality, then recommended someone else. Two examples, with the store’s name removed:
“I’m not familiar with a brand called [store] for jeans, so I can’t speak to its quality. However, if you’re looking for reliable and stylish options, I highly recommend checking out Levi’s for their classic fit and durability.”
“[Store] isn’t a brand I’m familiar with for herbs and spice blends. For high-quality options, I’d highly recommend checking out Penzey’s Spices or McCormick Gourmet.”
In the other 24, Gemini knew the store, sometimes in real detail: materials, where things are made, what the brand is known for. Even then, many of those answers went on to suggest a competitor in the next sentence.
One caveat belongs right here. My tool wrote the store’s name the way it appears in the web address, so a brand called something like “Blue Fox Knitwear” was asked about as “bluefoxknitwear”. A real shopper would type the proper name, and Gemini may know some of these brands better than this suggests. I’m fixing that in the tool too.
What this means if you run a Shopify store
Technical fixes make you readable, but they don’t decide who gets named. Clean product data, complete schema and clear category names help the model understand what you sell. Nothing in these answers suggests that a missing tag was what kept 41 stores out. The brands that got named are the ones the model has read about in lots of other places.
So the work that moves this is mostly off your site. Look at who gets named in your category and where they get written about: “best [category]” roundups, review sites, gift guides, Reddit threads, press. Getting into two or three of the sources that already cover your category is slow, unglamorous outreach, and it’s what changes what the model knows about you.
Start with the brand question. Ask Google’s AI or Gemini “Is [your brand] a good brand for [what you sell]?” If it doesn’t know you by name, category questions are a long way off, and the first job is getting written about anywhere credible. If it does know you, the next job is getting into the roundups that answer the category question. The diagnosis guide walks through both checks.
Measure it, don’t guess. Write down the answers now, do the work, and ask the same questions again in a month or two. That before and after is the only honest way to know whether anything changed.
Limits of this data
- The API isn’t a live AI Overview. The API answers from what the model learned in training. A live AI Overview also searches the web as it answers and cites sources, so a store with strong coverage could appear there more often, especially for brand questions. Treat this as a measure of what the model knows about you, not a screenshot of Google.
- One run per question. AI answers vary from one ask to the next. A store that came up zero times in ten questions might come up once in the next ten.
- Small and skewed sample. 42 stores, chosen by me, mostly apparel. A different mix of categories could give different numbers.
- Clumsy wording. Some early questions read awkwardly (“Recommend a good dresses for someone on a budget”). The question templates have been rewritten since.
- Store names written as web addresses in the brand questions, as above.
I’ll rerun the study with the fixed tool and more stores, and update this page with the new numbers when I do.
Check your own store
You can do the rough version by hand in five minutes: ask Gemini, or Google with AI Mode, the brand question above and three “best [your category]” questions, and write down who gets named. The free Geomint check does the same with 12 questions built from your own catalog, and shows every answer word for word along with who got named instead of you.