Blog
We asked an AI assistant 264 buying questions. It never named 4 in 5 small stores.
August 18, 2026
We asked an AI assistant 264 real buying questions about 88 small independent online stores. It never named 69 of them. Not once. Roughly four in five small stores were named in zero of the three questions we asked about them, in the channel that is starting to answer the questions shoppers used to type into Google.
By Zachary Stevens · 18 August 2026 · Download the data
Disclosure, before anything else
I run Velantia, which sells a tool that measures the thing this study measures. I have a commercial interest in this finding being interesting. That is a reason to check the work, not to take my word for it. So the whole dataset, the sampling method, the exact statistics and the code that computes them are published, including the parts that came out dull and the parts where the instrument is weaker than I would like.
If you find an error, write to help@velantia.app. Corrections go at the top of this page with the date, not in a footnote.
What we did
We picked ten product categories that small independent makers dominate: handmade soap, hand-dyed yarn, pottery mugs, enamel pins, dog bandanas, soy candles, hot sauce, dog treats, leather wallets and indoor plant pots. Then we sourced 88 real Shopify stores from ordinary purchase-intent searches.
For each store, we wrote three category-level questions a shopper would ask when they don’t know any brand yet. The questions never mention the store. Then we asked Perplexity and counted, in code rather than by eye, whether the store’s name appeared in the answer.
264 questions. 88 stores. One assistant. One day.
What came back
69 of the 88 stores were never named in any of their three questions. Only 13% of questions named the store being asked about.
Three categories returned nothing at all:
| Category | Questions that named the store |
|---|---|
| Soy candles | 0 of 30 |
| Dog treats | 0 of 30 |
| Hot sauce | 0 of 21 |
| Indoor plant pots | 1 of 12 |
| Handmade soap | 4 of 30 |
| Leather wallets | 4 of 30 |
| Hand-dyed yarn | 6 of 30 |
| Pottery mugs | 6 of 30 |
| Dog bandanas | 8 of 30 |
| Enamel pins | 6 of 21 |
In soy candles, dog treats and hot sauce, not one store we sampled was mentioned in a single answer.
Ranking on Google is not the same as being named by AI
This is the finding we did not expect to be so clean.
27 stores, nearly a third of the sample, appear on Google’s first page for their category and were never named by the assistant.
These are businesses whose search visibility works. They did the SEO. They rank. And in the channel that increasingly answers the same question, they do not come up.
The same names absorb the recommendations
Across every answer we collected, 969 recommendation slots were counted. The ten most-named brands took 27.6% of them.
Most frequently named: Bellroy, Malabrigo, East Fork, Madelinetosh, Deneen Pottery, Miss Babs, Saddleback Leather, Shinola, Brooklyn Candle Studio, Diptyque.
None of these are our customers. It is simply what the assistant said, on 18 August 2026.
Would you get the same answer tomorrow?
AI assistants are non-deterministic: ask twice, get different wording. So we re-ran 20 of the stores five more times, changing nothing.
| Sweep | Named in zero of three |
|---|---|
| 1 | 16 of 20 |
| 2 | 16 of 20 |
| 3 | 16 of 20 |
| 4 | 16 of 20 |
| 5 | 16 of 20 |
Identical every time. Not one store changed sides.
And the answers were genuinely different each run. Across those 100 runs, no store received the same answer text twice, and the questions varied between sweeps too. The assistant re-answered from scratch, named brands in a different order, and reached the same verdict about who was worth mentioning.
The wording churns. The outcome doesn’t.
An SEO audit would not have told these stores anything
We fixed six on-page signals in advance and compared the stores that were named against the ones that weren’t:
| Signal | Named (19) | Never named (69) | Difference |
|---|---|---|---|
| Schema markup | 89% | 91% | none (p=0.81) |
| Meta description | 100% | 96% | none |
| H1 present | 84% | 86% | none |
| Homepage word count | 2,032 | 1,862 | none |
| FAQ page | 74% | 55% | leans, not significant (p=0.14) |
| About page | 95% | 86% | leans, not significant (p=0.28) |
Schema markup was slightly more common among the stores that were never named. The classic on-page SEO checklist (schema, meta description, H1, word count) shows no relationship at all with whether an assistant names you.
This is the most useful thing in the study, and it is not what we expected to write.
It means you cannot infer AI visibility from an SEO audit. A store can pass every technical check, rank on Google’s first page, and still be absent from the answer a shopper actually reads. Twenty-seven of our stores were in exactly that position. If markup predicted mentions, nobody would need to measure AI visibility separately. You would just run an SEO tool and read across.
The two signals that do lean the right way, FAQ and About page, are the two that give an assistant something concrete to quote about you. At 19 stores in the named group, neither gap is statistically significant, so treat that as a hypothesis rather than a finding. We are not going to dress up p=0.14 as evidence.
What the data does support, by elimination: being named is earned off your site, in the roundups, directories and third-party reviews an assistant reads, not in your markup. That is a slower and less satisfying answer than “add schema”, and it is the one the numbers point at.
What this doesn’t show
- One assistant. Perplexity, on one day. Not ChatGPT, not Gemini, not Google’s AI Overviews. Nothing here describes those.
- 88 stores, not a census. The 95% interval around “four in five” is roughly 70% to 87%. That’s why we don’t quote a decimal.
- Three questions per store. The aggregate is the measurement. Any single store’s result is close to noise, which is why we publish no store’s individual outcome.
- A search-sourced sample can only reach stores that rank somewhere. A store invisible everywhere can’t turn up in search results and so can’t be sampled. The real figure for small stores is therefore worse than ours, not better.
- Correlation, not cause, for the on-page comparison, since 19 stores is a small group.
The data
- Per-store results (anonymised)
- Brands named instead
- Most common problems found
- Run counts and corrections
Store identities are not published. We are not going to name a small business alongside the finding that nobody recommends it. Brands that were named appear under their real names, because that is a positive and checkable fact about what an assistant said.
One correction is in the data and worth stating plainly: our first pass at matching brand names got 14 verdicts wrong in both directions, mostly by demanding a store’s full page title where the assistant had used its short name. We recounted from the stored answers and published both numbers. Before the correction the headline read 72 of 88; after it, 69.
Licence and citation
This study, its data files and its methodology are released under Creative Commons Attribution 4.0 International (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/. Use the numbers, redraw the charts, re-analyse the CSVs, disagree with us in public. Just cite it.
Velantia (2026). We asked an AI assistant 264 buying questions. It never named 4 in 5 small stores. https://velantia.app/blog/ai-assistants-ignore-small-stores/. Data and methodology licensed CC BY 4.0.
Plain-text attribution, for a chart caption or a footnote:
Source: Velantia, “We asked an AI assistant 264 buying questions. It never named 4 in 5 small stores” (2026). CC BY 4.0. https://velantia.app/blog/ai-assistants-ignore-small-stores/
The four CSVs linked above are already downloadable. No email, no form. I will answer questions about the method, including hostile ones, and I will re-run any slice of this you want to check. help@velantia.app
Velantia checks whether AI assistants and Google can find your store. Get the demo.