Short answer
You do not need to buy tools or services first. Prepare 5 English questions, ask each 3 times in ChatGPT and Gemini (30 answers), and tally three numbers: how often category questions mention you, how often brand questions get you right, and whether your domain appears in cited URLs. Nobody can guarantee AI will recommend you; the test only shows what AI says about you today.
The short version
If you export B2B products with a long buyer decision cycle, a half-day test is worth it. Test first, then decide whether to invest further, not the other way round.
The test is concrete: when a buyer asks AI in English "who are reliable suppliers of X", which companies does it list? When someone asks about your company by name, is the answer correct? Three possible results lead to three next steps, shown in the table in step 5.
One correction first. As we understand it, AI produces an answer when someone asks. What a company can study is how AI describes it, and whether it is mentioned, when your category comes up. A promise to "make AI recommend you" is one we believe cannot be kept, so skip anyone who offers it.
Why now
Exports are up in the first three quarters. Will more buyers be looking for suppliers? Xinhua reported on 14 October 2026 that goods trade in the first three quarters reached 39.94 trillion yuan, up 18.8% year on year, with exports at 23.2 trillion yuan, up 16.4%; September alone was 5.16 trillion yuan, up 27.5% (Xinhua, in Chinese). By 19 October, phase one of the 140th Canton Fair had drawn 189,000 overseas buyers from 214 countries and regions, up 13.3% on the previous edition (Xinhua, in Chinese). Phase two runs 23 to 27 October, and the offline fair continues to 4 November.
These figures cover trade overall and fair attendance. None of them says how many of those buyers use AI to find suppliers, and we do not cite any statistic on that, nor surveys run by service providers. Our view: with more buyers attending, more will check and compare after the fair, and AI may be one of the channels they use. So test your own category with your own data instead of trusting an industry average.
What it means for you: more enquiries around the fair are a chance to collect first-hand answers on how buyers found you.What to do: start with the zero-cost first step below.
Step 1 (zero cost): ask your buyers whether they use AI
Add a "How did you find us?" field to your website enquiry form, with an option "AI assistant (ChatGPT, Gemini, etc.)". When sales staff reply to a new enquiry, they can ask the same thing. In our view, a month or two of records says more about your buyers than any industry article.
If you have no form, add one line to your first email reply. Record two columns only: date and the buyer's own words.
Step 2: write 5 questions
Three category questions that do not name your company, in a buyer's voice. Two brand questions that use your English company name. Replace the brackets with your own details.
Category question templates
Who are reliable manufacturers of [your product] in China? List 8 with their websites.
I need a supplier for [custom/OEM version of your product] from China. Which companies would you recommend? Give me 6 names.
Who are reliable [product] manufacturers in China for export to [your main market]? Give me 8 names with websites.
The first two follow the valve questions we asked Gemini on 16 October 2026; the third is adapted from our mould question.
Brand question templates
What do you know about [your company English name]? What do they make and where are they based?
Is [your company English name] a reliable supplier of [your product]? What certifications do they hold?
Use the English name most common on your website and platforms, not an abbreviation.
Step 3: ask 30 times and record facts only
Ask in both ChatGPT and Gemini. Start a new chat for every ask and repeat each question 3 times: 5 questions × 2 assistants × 3 runs = 30 answers.
Use a new chat because earlier content can influence the next answer. For every run, note the assistant, the model, whether you were signed in and whether web search was on. Our view: sign-in and web search can both change answers, so do not mix results from different conditions.
Log sheet (copy into a spreadsheet, one row per answer):
| Date | Assistant and model | Signed in / web search | Type | Question text | Run | Mentioned me? | Facts about me | My domain in cited URLs? |
|---|---|---|---|---|---|---|---|---|
| [example] | [example] | [example] | category/brand | [example] | 1/2/3 | yes/no | right/wrong/missing | yes/no |
Fix the rule for "mentioned" in advance, or it will loosen as you read: your company name or domain must appear in the answer text. A generic reference to the industry does not count.
Check "facts" against your own website, item by item: city, main products, certification names, export markets, founding year. For wrong items, write down what was wrong.
We have not timed this method. Every one of the 30 answers has to be read and logged, so set aside half a day.
Step 4: tally three numbers
- How many times category questions mention you, out of 18 (3 questions × 2 assistants × 3 runs).
- How many times brand questions get your facts right, out of 12 (2 questions × 2 assistants × 3 runs). "Right" means none of the items you checked is wrong.
- How many times your domain appears in the cited URLs.
Step 5: choose the next step
The thresholds below are our own, not an industry standard, and have not been validated on a large sample. Treat them as a starting point and adjust after a run or two.
| Your result | What it suggests | Next step |
|---|---|---|
| Mentioned in over half of category answers (at least 10 of 18), facts mostly right | A solid base | Re-test each quarter with the same questions; no special investment needed |
| Rarely mentioned in category answers, but brand answers are right | AI knows you but does not connect you to the category | Worth investigating. Put checkable facts on product pages (standard numbers, certificate numbers, issuing bodies); see the 6-point product page checklist. Also check that company profiles in directories and fair listings match your website |
| Brand answers fail or are wrong | Check access first | Run the access checks below. Until they pass, do not talk about recommendations |
Access checks for the third band (a version operations staff can follow)
- Is robots.txt blocking by mistake? Open "your-domain/robots.txt" in a browser and look for a site-wide block such as Disallow: /. Our view: one possible cause is a test-site setting that was never removed at launch. If you cannot read the rules, paste the whole file into a conversational AI and ask it to explain line by line, then have a technical colleague confirm.
- Is the page indexed by Google? Use the URL Inspection tool in Google Search Console on your home page and one key product page. Searching site:your-domain only gives a rough view, and no result does not prove a page is not indexed; trust Search Console.
- Is the key text in the page source? Open a product page, press Ctrl+U (Command+Option+U on a Mac) to view the page source, then Ctrl+F for a model number or parameter value. If it is missing, the content may be loaded by script after the page opens; if the site uses a front-end framework, ask your site builder to confirm.
- Is company information consistent across platforms? List the company English name, address and main products from your website, Alibaba.com, LinkedIn and fair directories in one table and compare. Our view: the more your company is written differently in different places, the easier it is for AI to mix up details. We have not run a controlled test of this.
Our own test, and what it cannot show
On 16 October 2026 we opened new chats in the Gemini web app (signed in, default model) and asked three supplier questions in English, once each: gate and globe valve manufacturers in China (8 requested), OEM non-standard valves (6 requested), and injection mould makers exporting to the US and Europe (8 requested).
Result: all three answers listed named companies, 22 descriptions covering 19 companies in total. None of the 22 gave a certificate number, issuing body or factory audit report. In the first valve answer, only 2 of 8 suppliers had certification names; in the mould answer all 8 had a certification line, but names only.
The test shows only that AI does answer this kind of question and does list names. It cannot show who is recommended most, and it does not prove that "if your site says this, AI will mention you". One assistant, three questions, one run each: AI answers can differ every time. We have not yet run the full 30-answer method above, so its thresholds remain unvalidated.
Researching and buying services are different things
Self-testing, logging, adding text to your website and aligning company details across platforms can all be done by your own export or marketing staff. It costs time. Our view: consider an outside service only in these three cases:
- Your test falls in the third band, the cause is technical (crawling, rendering) and nobody in-house can fix it;
- You need to track many questions, markets and AI assistants over time and lack the people;
- Nobody in-house owns this work.
Even then, ask the provider to test your own questions in front of you first, and compare how they record results with your own method.
If you would rather not ask by hand, you can start with an AI visibility audit. It asks ChatGPT, Perplexity and Gemini questions as an overseas buyer and checks whether AI can read your website; the method is explained in How does an AI visibility audit work?. To re-test with a fixed question set over time, see our GEO service.
Frequently asked questions
- We are a small company with no dedicated staff. Is the self-test worth it?
- Start with the enquiry source record, which takes almost no time. If no buyer ever mentions AI, you can pause. If you export a high-value category where buyers do their homework, schedule the half-day test.
- Can I test only ChatGPT or only Gemini?
- You can start with one, but do not treat a single assistant as the conclusion. Our own test used Gemini only, which is why the article says "this one".
- Are 3 runs per question enough?
- Three is our starting point, meant to show you that answers to one question vary. There is no statistical basis that 3 is enough. If you appear in some runs and not others, results are unstable; try 5 runs next round.
- If AI does not mention me, is my website the problem?
- Not necessarily. The cause may be your website, or that the other pages AI draws on do not mention you. The test finds what may be missing; it is not a ranking.
- How soon will changes to my website show up?
- There is no fixed timeline and nobody can guarantee a change will appear. Re-test with the same questions after changes, record the date and compare before and after.
Further reading
See how AI describes you today
The free AI visibility audit asks ChatGPT, Perplexity and Gemini questions as an overseas buyer would, and checks your website’s technical foundation.