We Asked ChatGPT to Shop 3 Industrial Websites. It Passed the First-Visit Test.

Everybody asks whether AI will get their company wrong. We wanted to ask a different question: when a buyer's first visit happens inside an AI chat instead of on your website, does the AI do a competent job with what your site gives it? So we ran the visit. Three real industrial supplier websites, five buyer questions each, one fresh ChatGPT chat per site — and we graded the AI the way we'd grade a new hire on their first day.

It passed. Which turned out to be the less interesting half of the story. The interesting half is why.

The test

The three sites: a fractional-horsepower gearmotor manufacturer (Site A), a Siemens-focused automation distributor and integrator (Site B), and an employee-owned multi-division industrial integrator (Site C). Real companies, anonymized here — they aren't our clients and we didn't ask to feature them; the point is the pattern, not the names. The five questions were the same everywhere: what do you do and who is it for; a specific plant problem (a conveyor drive tripping on humid days); how to get a quote or a human; lead times and pricing; and proof of past work in food and beverage plants. For each answer we recorded whether it cited the company's site, whether it hedged or refused, whether anything contradicted the site's own pages, and whether it named competitors.

What it got right

The identity questions were clean: all three company descriptions matched the sites' own homepages — what they sell, who buys, where they operate. The contact questions were better than most websites' own contact pages: real toll-free numbers, regional office lines, quote forms, and the distributor-finder, all matching what the sites publish (one minor phone-number mismatch flagged but unverified). Asked about lead times and pricing, ChatGPT browsed the live sites and came back with specifics — same-day shipping cutoffs, per-product lead times, actual listed prices — each appropriately qualified with “as listed” and “confirm at quote.”

The problem question was the surprise. Given a conveyor drive tripping on humid days, it didn't just describe the company — it triaged the problem: moisture-cause checklists, fault-code-first diagnosis, lockout/tagout safety warnings, and then the engagement route (who to call, what to have ready, even a drafted email). It behaved like a competent applications engineer doing intake. Zero competitors were named as alternatives in any of the 15 answers. Brand names surfaced twice — three drive manufacturers as examples inside a diagnostic follow-up question, and four more in a pre-call checklist of drive brands the buyer might already own — but never as cited facts or suggested vendors.

The honest hedge is the feature

Here's the finding worth your attention. Question five asked for proof — named food-and-beverage projects — and wherever the sites hadn't published any, ChatGPT said so instead of inventing it. For Site A: the food-equipment application page and washdown product lines were real and cited, but named plants were “not established by evidence reviewed.” For Site B: it explicitly separated product expertise (IP69K controls, F&B-named applications) from named-project proof, flagged a case study as not food-processing evidence, and listed what it could not verify. For Site C, which had the strongest proof, it still flagged an unverified heritage claim rather than repeating it.

Compare that with the WashU 2026 audit finding — up to about 1 in 9 AI Overview claims unsupported by their sources. Unsupported claims are the failure mode everyone fears, and this test suggests the failure isn't automatic: given company sites with clear, bounded facts, the model hedged at the boundary instead of crossing it. The hedge is the feature. But notice the dependency: it could only hedge honestly because the sites gave it real material to work from. An AI can't be carefully accurate about a company it can't read.

Why it worked: the quotable surface

Step back and the pattern is almost boring. These three sites publish facts: prices on the page, lead times by product, phone numbers in the footer, named customers in case studies, application pages organized around buyer problems. ChatGPT's good answers were just those facts, rearranged. There is no clever prompting here and no AI magic — the entire performance is downstream of publishing.

Which is also why this result should worry most industrial companies rather than comfort them. Our September 2026 grading of 207 industrial websites found 79 with no site search at all and 106 burying answers three or four levels deep in navigation. The three sites in this test are the exception, not the rule. For the median industrial site — brochure copy, taxonomy navigation, proof locked in PDFs — the same test would have much less to quote, and the honest hedge would be doing a lot more of the talking.

The playbook: give the AI something to quote

Publish the facts or say plainly that you don't.Real prices where you can; where you can't, a clear “quote required” beats a missing page. In our test, “quote required” got an accurate answer and a missing page would have gotten a hedge.

Name your customers.The strongest answers in the test came from named case studies — a distillery automation project, a food-ingredient feeder modernization. Anonymous “leading manufacturer” proof is unquotable proof.

Write problem-led pages. The triage answers worked because the sites had application content organized around what buyers describe (“drive tripping,” “washdown environments”), not just product taxonomy. The use-case page playbook.

Run the five questions on your own site, monthly.Fresh chat, the five questions above, and the same four checks: cited your site? hedged or invented? contradicted your pages? named competitors? It's a free AI-visibility audit you can run in an afternoon — the full accuracy-audit version.

The ambassador angle

Notice what ChatGPT added on top of the facts: call scripts, pre-call checklists, drafted outreach emails, safety warnings. Unprompted, it behaved like a pre-sales ambassador working from the company's content — answering the problem first, then routing the buyer to a human with context. That's the job description, with one advantage a site ambassador holds over the chatbot in this test: it answers only from your content, so unlike ChatGPT it can't browse off to a competitor's site mid-answer. And it's a job that shouldn't require the buyer to leave your site for a chatbot: the same ambassador, trained on the same content, can live on your website — answering at 2 AM, and capturing the inquiry instead of just advising it. See the pre-sales ambassador on your own content.

Method

Three real mid-size US industrial supplier websites (fractional-HP gearmotor manufacturer; Siemens-focused automation distributor/integrator; employee-owned multi-division integrator), anonymized as Sites A, B, and C; homepages read ~2 minutes each before testing. Five fixed buyer questions per site (company identity; conveyor-drive-tripping triage; quote/human contact; lead times and pricing; food-and-beverage proof), each in a fresh ChatGPT chat (signed-in Plus session; exact model version and memory/personalization state not recorded — answers vary across models and settings), October 19, 2026, PDT; read-only throughout, nothing submitted. Each answer scored on four checks: cited the company's own domain, hedged vs invented at proof boundaries, contradicted the site's own pages, named competitors. Results: 15/15 cited the company domain; 0 competitors named as alternatives (brand names surfaced twice — diagnostic examples and a pre-call checklist — never as suggested vendors); 0 contradictions with site pages (1 minor phone-number discrepancy flagged, unverified); explicit honest hedging on unproven claims at every proof boundary. Grader figures (79/207 no site search; 106/207 navigation overload) from the Nimbiq Website Grader, September 2026. WashU audit figures via our correction playbook. All figures attributed inline with years.