We ran 404 buyer prompts for 27 small online stores. Here's what AI got wrong.

ChatGPT rarely recommends small online stores: in our test of 404 buyer questions for 27 small US and Canadian stores, ChatGPT named the store in 17% of its answers and Google AI Overviews in 18%. When ChatGPT did name a store, 63% of the product links it attached to that store went somewhere other than the store's own website.

The stores sell their own products: supplements, granola and cereal, bike bags, tents, camp lights, walking poles, activewear. We ran each store's questions once on both engines between September 25 and October 1, 2026, then read the answers. Not the summaries our reports write about them; the answers themselves, with every link pulled out and sorted by where it actually goes.

Most advice about getting recommended by AI is general: get listed on Amazon, get mentioned on Reddit, add schema. This post is the opposite. It is what two AI engines actually said about 27 real small brands in one week, with the counts, the method and the caveats. The stores aren't named; they didn't ask to be in a blog post.

The numbers

Every row shows its denominator. "Named" means the store's brand name appears in the text of the answer.

The sections below go through each row, with examples. The method and what we excluded are further down, and they matter: 27 stores is a small sample, and every one of these numbers comes from a single run.

1. Small stores were named in about one answer in six

Each store had 14 to 16 buyer questions written around what it actually sells, in its market, and none of them contained the store's name. Things like "best touring panniers for a long-distance bike tour" for a bike-bag maker or "halal omega-3 for adults who don't eat much fish" for a halal supplement brand. These are questions the store should have a fair shot at, not generic category searches.

ChatGPT named the store in 69 of 404 answers. Google AI Overviews named it in 70 of the 393 questions where Google showed an AI Overview. Spread across stores, the middle store was named in 3 of its 15 questions by either engine. The best was a halal supplement brand at 10 of 14, followed by a Canadian bike-bag maker at 9 of 15. Two granola brands were never named by either engine in any of their 15 questions.

None of this is surprising if you have read why products don't show up in ChatGPT. What the rest of the data adds is that "not named" is only the first problem, and often not the most expensive one.

2. ChatGPT and Google AI Overviews rarely agreed

For 393 questions we had an answer from both engines. In 103 of them at least one engine named the store. Both engines named it in only 34. ChatGPT alone named it in 33, and Google AI Overviews alone in 36. The overlap is about a third, and it is close to even in both directions, so neither engine is simply a stricter version of the other.

At the store level the split is starker. Nine of the 27 stores were named zero times on one engine. A cycling-tool brand was named in 6 Google AI Overviews and 3 ChatGPT answers; a tri-bike maker in 4 ChatGPT answers and no AI Overviews; a bike-bag maker in 3 AI Overviews and no ChatGPT answers. A store checking only ChatGPT would have concluded the bag maker was invisible, and a store checking only Google would have said the same about the bike maker.

The reason is that the two engines draw on different sources. Google's answers lean on pages that already rank in Google Search, which we cover in how Google AI Overviews work. ChatGPT's come from its own web search and the retailer and review pages it finds there. For a store owner the practical point is simpler: check each engine separately. One engine's result tells you little about the other's.

3. When ChatGPT named a store, it usually linked somewhere else

This is the finding we didn't expect at this scale. ChatGPT often formats a recommendation as a product link above the product's name, plus a smaller source link underneath. They are two different things: the product link is where a shopper clicks to buy, and the source link is the page ChatGPT read. We counted them separately, only for the store's own entry in the answer, and we sorted every destination by hand.

Only 30 of 82 product links (37%) went to the store's own site, against 37 of 76 source links (49%). That gap is the trap: ChatGPT often read the store's own page and then sent the shopper somewhere else to buy. 44 of the 69 ChatGPT answers that named a store carried at least one product link to another site, and in 23 of them no link of either kind pointed at the store's own site.

Some examples, with the stores anonymized. Each is a single answer; the counts above are the pattern they illustrate:

• A US camp-light brand was ChatGPT's top pick for a telescoping-lantern question. 4 of the 5 product links in that answer went to an outdoor shop in Chile, priced in pesos, and an Australian outdoor store, priced in Australian dollars. Every source link underneath went to the brand's own homepage.
• A Canadian bike-bag maker was recommended in two ChatGPT answers about commuter panniers, each time with a URL on a .com domain that belongs to an unrelated company with the same name. A third answer linked an "-usa" variant of the name instead of the store's own Canadian site.
• A hyaluronic acid supplement brand got ChatGPT's top spot for dry eyes, with the product link pointing at iHerb's Brazilian site, in Portuguese. In the same answer, a second product of theirs linked to a Walmart category page instead of a product listing.
• A halal supplement brand's adult gummies were linked to a marketplace listing in Oman. A Canadian sports-nutrition brand's waffles linked to its own Swiss storefront, and its recovery mix to a shop in Sweden.

In all, 12 product links in 10 answers, for 7 of the 27 stores, sent a North American shopper to a store in another country. 6 more, for 4 stores, went to an unrelated or look-alike domain or a generic category page instead of the product.

Amazon, the usual worry, barely showed up: 1 ChatGPT product link out of 82, and 2 of the 30 store links in Google's answers. The leaks were smaller and harder to spot: independent resellers, foreign storefronts, and domains that only look right. Why ChatGPT picks these pages, and how to fix each case, is in why ChatGPT links your product to another store.

Google AI Overviews behaved differently. They named the store as often as ChatGPT did but attached a merchant link to it far less: 26 of their 70 answers naming a store had one. When they did, it was usually right, with 23 of 30 links going to the store's own site.

4. AI cited pages that recommend the store, then named other brands

For every third-party page an answer cited, our page check records whether that page names the store. Matching those pages against the 658 answers that didn't name the store, we found 31 answers, across 15 of the 27 stores, where AI used a page that names the store and left the store out of its answer.

23 of those 31 cited pages were reviews, articles or videos rather than shops, and in 22 of them the answer recommended other brands instead. Google AI Overviews did this 15 times and ChatGPT 8. Some examples, each from a single answer:

• Asked for the highest-rated low-sugar granola, Google's AI Overview cited a food magazine's roundup that picks the store's granola as its best paleo option, and recommended KIND.
• A commuter-backpack answer cited a gear site's tested roundup that names a small US bag maker as its American-made pick, then recommended Ortlieb and Thule.
• ChatGPT's answer on sports bras for active women cited a women's magazine guide that includes the store's bra, and listed five large brands.
• Two of Google's answers about bike torque tools cited a cycling magazine's group test that rates the store's wrench, then recommended other toolmakers.

This is the most fixable finding in the study. The model already has a source that recommends you; it just isn't carrying your name into the answer. We go through why that happens and what to change in cited but not named, and how to spot it for your own store in our guide to tracking AI citations.

5. Prices held up, where we could check them

We expected wrong prices to be a headline. They weren't. AI wrote a price into its answer next to a link to the store's own product page 10 times. For 6 of those our page check could read the live product page, and all 6 quotes matched a price on it, including a range ("around $21–$23") that brackets the real $22.99.

We check the quoted price against every variant on the page, not just the first price we find. That matters: one quote that looked $18 too high in an earlier report turned out to be the correct price of a larger bag size on the same page.

The other 4 quotes we couldn't check at all, because the store's host refused our page check that week. We counted those as unknown, not as matches. Six checked prices is too few to say AI gets prices right in general. It is enough to say we didn't find a wrong one, and that a price mismatch should be verified against every variant before anyone acts on it.

How we ran it, and what we excluded

The stores. 27 small brands that sell their own products online, 24 in the US and 3 in Canada, all in English: 15 in food and supplements, 12 in outdoor gear, cycling and activewear. We picked them as outreach prospects for SearchMention, so they aren't a random sample of online stores, and we chose them before running anything.

The questions. 404 in total, 14 to 16 per store, generated from each store's catalog and market and reviewed by hand before running. None contains the store's name. The mix: 142 discovery ("best X for Y"), 113 use case, 78 where-to-buy, 41 attribute, 28 price and 2 comparison questions.

The engines and dates. ChatGPT and Google AI Overviews, localized to each store's country, one run per store between September 25 and October 1, 2026. Google returned an AI Overview for 393 of the 404 questions; ChatGPT answered all 404. That is 797 answers.

What counts as named. The brand name appears in the answer text. Our software's own matcher flagged 14 more AI Overviews and 1 ChatGPT answer as naming the store, but we couldn't confirm the name in the saved answer text (in some, only the store's page or a product card points to it). We left those out of the named count. Counting them would lift the AI Overview figure from 18% to 21%.

Links. We pulled every link out of each answer and kept the ones attached to the store's own entry, then sorted each destination by hand: the store's site, reseller, marketplace, foreign storefront, review, or unrelated. "Another country" means the site serves a different country than the store's (by domain or currency). We don't call a link dead unless our page check for that run recorded it, and none of the examples above rely on that.

Cited but not named. Based on our page checks of third-party sources, which record whether a page names the store. The report keeps up to 10 such sites per store, so 31 is a floor. We excluded answers our matcher had flagged as naming the store, to avoid counting the same answer as both.

Prices. Only quotes next to a link to the store's own page, only where the page check succeeded. Runs from September 30 onward had page checks refused (HTTP 429) on several store hosts, so those quotes count as not checked. The two earliest runs predate saved price quotes.

One run per question. Each question was asked once per engine. AI answers change from run to run, so a single answer is a snapshot, not a measurement: the same question tomorrow could name the store, or not. That is why we report patterns across stores (how many stores were never named, where links went across 82 of them) rather than leaning on any one question, and why each example above is labeled as a single answer. A second run of the same questions would tell us how stable these shares are; we haven't done it yet.

Other studies. We're not the first to audit AI answers for online stores. Digital Commerce Partners, an ecommerce SEO agency, published a report on 25 ecommerce brands across three AI platforms and 450 buyer-intent queries, and runs each query twice to measure drift. What this study adds is where the links go (own store, reseller, marketplace or a storefront abroad, with product links and source links counted separately), quoted prices checked against every variant on the live page, and Google AI Overviews counted as its own engine.

Caveats. 27 stores and one run per question is a small sample, so treat every percentage as a rough size rather than a benchmark. We read the raw answer behind each example above; none comes from a report summary.

What a small store can do with this

Check where the link goes, not only whether you're named. Open ChatGPT, ask three questions a buyer would ask, and click the product link under your name. In this sample, a named store's product link went to its own site only about a third of the time. If it lands on a reseller abroad, that reseller's page is what ChatGPT trusts for your product, and making your own product page the clearer, more complete source is the fix.

Check both engines. A third of these stores were invisible on one engine and present on the other. A clean result on one tells you nothing about the other.

Find the reviews that already name you. Look at the sources cited in answers that don't name you. If a roundup that picks your product is among them, you have a page the model already trusts. Ask that publisher to link your product page by name, and make sure your own page states the same thing the review praises, in plain sentences.

Verify before you panic about prices. We found no wrong prices in the quotes we could check. If a tool flags one, compare it with every size and variant on your page first.

Frequently asked questions

Next steps

1. Run a free AI readiness scan to see what keeps your product pages out of answers.
2. Ask ChatGPT and Google three buyer questions each, and click every link next to your name.
3. Read our earlier first-party study, ChatGPT sent us 199 visitors, Google sent 9, for why getting named is worth the work.