ChatGPT doesn't cite your prompt, it cites its own searches
When someone types 'where to buy 18mm marine plywood online', ChatGPT doesn't hand that sentence straight to a search engine. It rewrites it into a handful of its own queries, runs them, reads the results, and builds an answer from the pages it trusts. The words it searches for are often not the words the buyer typed.
So the merchant that gets named isn't the biggest yard or the cheapest. It's the one whose page most closely matches the query ChatGPT actually ran. Which means the winning move is to find out what it ran, then make your page the obvious answer to that exact query.
How to see the exact queries ChatGPT runs
You can pull the real queries out of ChatGPT in about a minute, straight from your browser. Do this on desktop Chrome.
First, ask ChatGPT a buying question in your category and let it do a web search. Then right-click the page and choose Inspect to open developer tools.
Look at the address bar. After the '/c/' in the ChatGPT URL there's a long string of characters, that's the conversation ID. Copy it.
In developer tools, open the Network tab, paste your copied ID into its filter box, then hit refresh on the page. Click the matching request that appears, and open its Response tab.
Search that response (Cmd+F inside it) for the word 'queries'. There they are: the actual searches ChatGPT ran to answer, in its own words.
Now you're not guessing. You're looking at the literal language ChatGPT used to go and find a supplier, which is the language your pages need to echo.
Match four things to the query, in this order
Once you've got the query ChatGPT ran, line up four things on your page to sit as close to it as you can: the URL slug, the page title, the H1, and the first sentence of your copy. Those are the four spots a machine weighs most when it decides what a page is about.
If ChatGPT's query is '18mm marine plywood 2440x1220 uk', then a page at /products/18mm-marine-plywood-2440x1220, titled '18mm Marine Plywood 2440 x 1220mm', with a matching H1 and a first line that reads '18mm marine-grade plywood sheets, 2440 x 1220mm, in stock for next-day delivery', is a clean, liftable answer. The same product sat at /ProductDetails/?Code=PLY18MRN under a title like 'PLY MRN 18 2440X1220' is not, no matter how good your price is.
You're not stuffing keywords. You're removing the gap between what the AI searched for and what your page plainly says it is.
Why merchant ERP pages fail this test
This is exactly where ERP-driven catalogues fall down, and it's fixable. Two problems show up again and again. First, the page title is really a stock description: 'T&E 2.5MM 6242Y 100M GREY' means something to your warehouse and nothing to a machine trying to match 'where to buy 2.5mm twin and earth cable 100m'. Rewrite the visible title to the words the buyer searches, keep the code in the body for your team.
Second, the URL is a code. Something like /ProductDetails/?Code=6242Y25 tells ChatGPT nothing about what the page is. Where your platform lets you, use a descriptive slug like /cable/2.5mm-twin-and-earth-100m. And keep one intent per page: a page trying to be the plywood category and the plywood buying guide at once is a weak match for both. One page, one job.
What to do this week
Pick your three best-selling lines. For each, ask ChatGPT the question a buyer or a builder would ask to end up buying it, then pull the real queries using the steps above. Write them down next to the product.
Then check the four spots, slug, title, H1, first sentence, against that query and close the gap. You don't need to touch the ERP's guts, most systems let you override the visible title and lead copy. Just make the page an unmistakable answer to the search ChatGPT is actually running, and re-test in a few weeks to see if you've entered the answer.