@lumine and @clickfromai have the schema and feed side covered, so here are the two things I’d check first — both checkable on your own store this afternoon.
1. Is the answer text actually in the server-rendered HTML?
GPTBot, PerplexityBot and Google-Extended fetch your page and parse what comes back. They don’t wait around while a tab or accordion loads its contents after paint. On a lot of Shopify themes the specs, the shipping details and the FAQ block are exactly the parts that arrive late or sit behind a click — so what a shopper sees and what a crawler gets are two different pages.
Check it the boring way: view-source on a product page (or curl it) and search for a sentence you know is in your FAQ or spec tab. If it isn’t in the raw HTML, no assistant has ever read it. That one check tends to surface more than any schema tweak.
2. Whose words are on the page — yours, or the shopper’s?
People ask assistants things like “will this fit a 15 inch laptop” or “is this ok for sensitive skin”. Merchant-written copy almost never uses that phrasing, it uses product-marketing phrasing. Pages carrying real customer questions in the customer’s own words match those prompts better, for the dull reason that the wording is the same.
That’s the honest case for a shopper-submitted Q&A block on product pages rather than only a hand-written FAQ: you stop guessing which questions to answer, and the phrasing comes free. We build one of those (AC – Questions & Answers), so weigh that accordingly — the principle holds whatever you use, including a plain metafield you fill in by hand.
On FAQ schema specifically: add it, but know what it buys you now. Google restricted FAQ rich results to well-known government and health sites in August 2023, so it won’t earn you a SERP accordion anymore. It’s still worth having here for a different reason — it hands a crawler an explicitly labelled question-and-answer pair instead of a wall of divs.
On tracking: referrer filtering for chatgpt.com and perplexity.ai works, but treat the number as a floor rather than a total. A good share of assistant traffic arrives with no referrer at all, and if you judge the whole effort by that figure you’ll undercount it and kill something that was working.
