The AI Product Answers Study · Edition One · July 2026
Your next customer is asking an AI about your product right now.
We asked ChatGPT, Claude and Gemini the eight questions people ask before they buy and after they own it. 326 consumer brands. 8,776 answers. Graded against the brand’s own website.
Where buying decisions now start
Shoppers stopped searching and started asking.
That traffic converts better than anything else you have. This study is about what the assistant says on the way there.
And here is what it says about you
8,776 graded answers
of the 326 brands had at least one answer that was wrong or sent a customer to a contact found nowhere on their site.
286 brands of 326. Not the careless ones. Every kind of brand in the study, at every price, in every category. Being good at this is not something anyone has managed by accident.
It costs you before the sale, not just after
7,230 factual answers
- What are the specs?
- What warranty comes with it?
- Was it ever recalled?
- What parts fit it?
- Where can I buy it?
Refused to answer 20.1% of the time.
- How do I clean and maintain it?
- It stopped working, what now?
- How do I reach support?
Invented a way to contact the brand 19.8% of the time.
Your brand may be one of the 326.
Ask what the assistants said about you. john@bawte.com
Three in four answers had something wrong or missing
2,107 checkable answers
26.0%
Fully correct · 547 answers
64.1%
Right, but a detail off · 1,350 answers
10.0%
Contradicts the brand · 210 answers
145 of the 326 brands had at least one of those 210 contradicting answers.
What that looks like in practice
Real answers · brands anonymised
A parent asks how to reach a baby products brand about a broken swing
The assistant answered
“Phone: 1-844-XXX-1279 — this is generally the fastest way to get help, especially for safety issues, malfunctions, or warranty claims.”
The brand’s own site says
“Contact Customer Care at 888-XXX-6667”
The number does not appear anywhere on the brand’s site. A parent with a safety concern calls a stranger.
A parent asks what warranty comes with a stroller
The assistant answered
“A 2-year warranty, which is standard across most of their stroller lineup.”
The brand’s own site says
“Extended 4-year warranty on all new strollers.”
The brand doubled its warranty as a selling point and the assistant is still quoting the old one. A shopper comparing two strollers is working from the wrong number.
A shopper asks what warranty comes with a kitchen appliance
The assistant answered
“A 2-year limited warranty, covering manufacturing defects in materials and workmanship under normal use.”
The brand’s own site says
“The Registered Warranty applies for a period of three years.”
Across the 46 warranty errors we could read, the assistant more often went the other way: 27 promised MORE coverage than the brand offers, 11 promised less, and 8 could not be scored either way. Customers arrive expecting a warranty that was never sold to them.
Safety is the weakest answer of all
Share contradicting the brand
Not all assistants behave alike
| Assistant | Fully correct | Wrong | Refused | Cited the brand |
|---|---|---|---|---|
| ChatGPT, with search | 45.8% | 3.8% | 1.0% | 84.1% |
| Gemini, with search | 28.9% | 4.6% | 5.4% | 64.7% |
| Claude, with search | 22.3% | 5.6% | 1.8% | 71.0% |
| Claude, no web access | 24.7% | 19.9% | 38.6% | 0% |
An assistant with no web access can only repeat what it absorbed in training. Nothing a brand publishes today changes that row, which is why the search-enabled rows are the ones that matter.
90% publish a support page. 30% publish specs.
281 brands
Each bar is all 281 brands whose own page list we could read. Brands we could not read at all are left out rather than counted as silent. “Could not determine” means the page list did include something that looked like the right page, but it would not load for us, or it did not cover the topic once we read it. It is never counted as a missing page. That is why specifications is stated as a floor rather than a finding: even if every one of the 77 undetermined brands publishes a spec page we could not see, the ceiling is 57%.
Your category
| Category | Brands | Answers wrong |
|---|---|---|
| Bikes and e-bikes | 33 | 17.2% |
| Pet | 11 | 12.1% |
| Fitness | 12 | 11.5% |
| Tools and outdoor power | 33 | 10.4% |
| Baby and kids | 24 | 8.8% |
| Audio | 15 | 8.6% |
| Kitchen and grilling | 38 | 8.4% |
| Sleep and bedding | 17 | 8.3% |
| Outdoor gear | 12 | 7.6% |
| Smart home | 19 | 5.6% |
214 of the 326 brands fall into these ten groups; the rest sit in categories too small to report. Groups range from 11 to 38 brands, so treat the ordering as indicative rather than a ranking.
What actually moves the answer
| How machine-readable the site is | Brands | AI cites the brand | Fully correct | Wrong |
|---|---|---|---|---|
| Nothing published | 12 | 57.5% | 30.0% | 5.0% |
| Poor | 114 | 67.7% | 25.8% | 6.8% |
| Partial | 54 | 72.3% | 29.7% | 3.1% |
| Good | 10 | 82.8% | 39.8% | 2.4% |
Grounded assistants only, since one without web access cannot cite anything. 190 brands with a measured score. Correlation between score and citation rate, r = 0.23. Only ten brands scored well, so read that group as a direction rather than a precise figure.
What to do about it
Every number is what we measured
Put full specifications on a page, as text
Not in a PDF, not in an image, not only in a configurator. 120 of 281 brands list no such page anywhere. Specs were the question assistants refused most often, because for most brands there was nothing to read.
Publish your recall and safety status, even when it is clean
“No recalls on this product” is an answer. 124 of 281 brands list no such page anywhere. Silence is why this became the question AI gets wrong more than any other, at 26%.
Put your real support email and phone in plain text
Behind a contact form is invisible. 509 of the 862 answers to that question gave an address or number that appears nowhere on the brand’s site, and every assistant did it at roughly the same rate.
Mark up product data so software can parse it
Brands whose sites are machine-readable were cited 83% of the time against 58%, and were contradicted less than half as often.
Check whether the file speaking for you is actually yours
A third of brands serving a file written for AI assistants are serving their store platform’s default. It says nothing about their products and points assistants away from their own site.
What this means
Three things here matter more than the headline error rate.
The risk arrives before the sale
It is tempting to read this as a customer support problem. The data says otherwise. Assistants were wrong more often when someone was still deciding what to buy than after they owned it, 11.0% against 8.5%, and refused to answer twice as often. A wrong warranty term or specification reaches a person while they still have the option of buying something else. Meanwhile the traffic arriving from those assistants converts 54% better than anything else you have.
The biggest gap is silence, not error
Only three in ten publish a product specification page, and fewer than half publish anything about safety. We confirmed that by reading each brand’s own published page list rather than assuming. Asked those questions, an assistant has nothing authoritative to draw on, so it refuses or fills the gap itself.
Publishing for machines changes what gets said
Brands whose sites are structured so software can read them were cited 83% of the time against 58%, described correctly more often, and contradicted less than half as often. This is the one lever in the study a brand controls directly.
How we know
Every brand was asked the same eight questions, with only the brand and product name changed, so one brand’s result can be compared against another’s. Every answer was graded against that brand’s own published content by two independent AI graders, with a third settling disagreements. A verdict counted only when the grader could quote the line from the brand’s site supporting it. Answers where graders could not agree, or where the quote could not be found, were discarded rather than counted.
Four things we got wrong
Each was found by checking our own results rather than trusting them, and every one ran the same way: making the problem look bigger than it was. That direction is not a coincidence. It is what happens when one outcome suits the people running the study, and it is why these are published rather than quietly fixed. A study whose every error flatters its author is not worth reading.
- 01
Silence counted as being wrong
Where a brand’s site said nothing on a topic, the grader recorded the assistant as wrong. Silence is not contradiction. Every such verdict was re-judged on whether the site actually conflicts with the answer.
13% of “wrong” verdicts overturned
- 02
Brevity counted as being wrong
Answers stating nothing inaccurate, but covering less ground than the brand’s own page, were marked partly wrong. What an answer leaves out is not an error.
18% overturned · the fully-correct rate more than doubled
- 03
Real support contacts called invented, three times over
This took three passes. A published address falling outside the excerpt shown to the grader read as fabrication. Then a number written with a country prefix failed to match the same number without one. Then bare runs of ten digits, usually order numbers or case references, were being read as phone numbers.
16.8% → 9.3% → 9.0% → 8.5% across three corrections
- 04
Grader failures counted as unanswerable
Found after publishing. On 1,640 of the 7,230 factual answers, our grader ran out of room before writing a verdict, and the empty reply was recorded as “could not be checked.” Those answers were never really graded. That overstated how often a brand had published nothing to check against, so we withdrew our 55% figure rather than restate it. They are also missing from the 2,107 answers we could check. None of the empty replies was counted as an invented contact, so that finding stands.
55% “could not be checked” withdrawn
Limits
- These brands are not a random sample of the market
- 326 brands already present in Bawte’s system, weighted toward direct-to-consumer products in bikes, outdoor equipment, kitchen, fitness, baby gear and home. Read the findings as describing brands like these.
- Assistants do not answer the same way twice
- Asking again changed the verdict 31% of the time. Aggregate figures across thousands of answers are stable; a single brand’s result should be read as a band, not a precise number.
- The graders agreed outright 69% of the time
- A third grader settled the rest. 223 answers where the disagreement could not be resolved were excluded, along with 389 whose supporting quote could not be located.
- The 88% figure is close to arithmetically expected
- At roughly 27 answers per brand and a 10% error rate, a brand having at least one failing answer is close to certain before we start. We publish it because that is the point: no brand in this sample avoided it by being careful.
- Ten brands is a small top group
- Only ten scored well on machine readability. The direction holds across all four bands, but the size of the gap is not precise.
- One product per brand
- The brand’s flagship, identified from its own website. A wider catalogue may differ.
- Market figures come from third parties
- The adoption and traffic statistics are published by L.E.K. Consulting and Adobe Analytics and are not ours. Everything else on this page is measured from our own data.
See what AI says about your products
Every brand in this study has its own scorecard: what each assistant said, which answers were wrong, and where your site left them guessing. No individual results have been published, and none will be without permission.
If your brand was studied, ask for your scorecard. If it was not, we will run it. john@bawte.com