Buyers now ask ChatGPT for suppliers. Will AI recommend your factory? Three spreadsheets to find out
We tested this on one of our own export factories. We asked AI assistants 40 questions a buyer would actually type, none containing the company name. The factory was mentioned in 43% of the answers. Its website was cited
We tested this on one of our own export factories. We asked AI assistants 40 questions a buyer would actually type, none containing the company name. The factory was mentioned in 43% of the answers. Its website was cited in only 16%.
Being mentioned and being cited are different things. A mention means a buyer read your name. A citation means the AI sent them to your site.
If you run a small B2B business or build sites for one, here is the method I use. It is three spreadsheets and seven days, for one product and one target market.
Why bother
Before this check, the same kind of factory told us "no inquiries". When we traced it, everything worked, but 14 days of search data showed only 59 clicks to the site, and 0 to 1 real person reaching the contact page. Another of our export sites showed just 14 impressions in 28 days. Nobody was arriving, so there was nothing to convert.
Sheet 1: the question list
The mistake: asking AI about your own company name. Buyers do not know it. They ask "where do I find a reliable OEM factory for X" or "how do I check a factory really has capacity".
Fields: id, question text, type (find supplier, compare, certification, price, MOQ, lead time), the page on your site that should answer it, whether that page exists.
Write 20 for one product and one market. AI can draft candidates from your product pages and past inquiries. You delete the ones nobody asks.
Sheet 2: the citation check
The mistake: asking once, seeing your name, and calling it a win.
Ask the same fixed set in ChatGPT, DeepSeek, Perplexity and any other assistant your buyers use, with web search on. Fields: date, question id, assistant, web search on or off, mentioned (y/n), our URL cited (y/n/only others), which other pages were cited, accuracy (right / partly / wrong), screenshot.
One caution from Surfer's research (updated 2026-10-23): answers pulled through an API overlapped with what users actually see by only 15.5% to 31.6%. Ask in the real interface and keep screenshots.
AI can classify the pasted answers and compute the rates. You must ask, screenshot, and judge accuracy. Prices, certifications and capacity are the usual errors.
Sheet 3: the evidence list
The mistake: a website that says "professional manufacturer, guaranteed quality" and nothing else. An AI has nothing to quote.
Surfer found 38% of citations in Google AI Overviews come from the first 100 words of a page, so each page should answer its question right at the top. Google's own guide (Search Central, updated 2026-07-10) puts unique, non-commodity content first and says you do not need llms.txt or special markup for Google.
Fields: page URL, question id, does the first 100 words answer it, key fact (spec, certification, capacity, MOQ), evidence file (certificate, test report, photo), status (have / todo), last updated.
For a different export site of ours, we generated about 900 pages with AI and required every spec to carry a source. Over two consecutive 28-day windows, search impressions went from 411 to 1,374 and clicks from 14 to 52. That is not a controlled experiment and I do not claim cause and effect. I treat it as a signal to keep going.
AI can compare page text with your evidence list and mark gaps as TODO. You supply the real documents, check every number, and press publish. Never let it invent a certificate, a customer, a price or an MOQ.
The 7-day test
- Day 1: list 20 questions
- Day 2: ask each assistant, screenshot
- Day 3: mark questions where you were mentioned but not cited
- Day 4: pick the 3 biggest gaps and their pages
- Day 5: add real evidence, AI drafts the first 100 words, gaps stay TODO
- Day 6: human review, update pages, confirm robots.txt and your CDN do not block search crawlers such as OAI-SearchBot and PerplexityBot
- Day 7: re-run the same questions, same wording; track mention rate, citation rate, accuracy rate
Ignore any single "rank". The same question can get a different answer tomorrow. Seven days only tells you whether the process runs.
A prompt you can paste into Codex or Claude Code
I have one product line and one target market. Build three CSV tables:
1. Question list: from the product pages and past inquiries I provide, list 20 non-brand questions a buyer would ask. Fields: question, type, page that should answer it.
2. Citation check: I will paste each assistant's answer. Classify mentioned / our URL cited / other pages cited / accuracy, and compute the rates.
3. Evidence list: compare the page text I provide with my evidence list. Mark every fact without an evidence file as TODO and draft the first 100 words of each page.
Use only the material I provide. Do not invent customers, case studies, certifications, capacity, MOQ, prices, lead times or test data.
Do not publish anything and do not send any email or message. Output drafts only; I will approve every change.
Nobody can guarantee AI will recommend you. What you can do is find out where you stand and fix what is missing.
Numbers are from our own export businesses, 2026-09 to 2026-11. External sources: Google Search Central (2026-07-10), Surfer (2026-10-23). Checked 2026-11-20.
Code and setup scripts for Claude Code and Codex: https://github.com/Tommyapi/tommyapi-quickstart
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.