Often the problem is not the product. It is what an AI assistant can read about it: a thin description, no answers to the questions buyers ask, incomplete structured data, or crawlers that are not allowed in.
I built an AI agent, product_geo_agent, that checks these things on a Shopify product and scores it from 0 to 100. Here is what it found on three products, and how you can check your own store.
What an AI assistant needs from a product page
Before anything else, the page has to be reachable. If a store’s robots.txt blocks OAI-SearchBot, the crawler OpenAI uses for ChatGPT search, that crawler will not read those pages. My post on GPTBot and robots.txt covers which bots do what.
After that, my agent looks at seven things. The weights are my own choice, not a published standard:
| What it checks | Points |
|---|---|
| Does the description answer common buyer questions? | 20 |
| Description quality and length | 15 |
| FAQ content (a store FAQ page, or a product FAQ) | 15 |
Complete Product structured data |
15 |
| Is the product named when an AI model is asked for a recommendation? | 15 |
| Clear specs and variants | 10 |
| Image alt text | 10 |
The AI model judges the writing. Plain Ruby code does the math, so the score is the same sum every time for the same judgments.
Three real audits: 70, 33 and 10 out of 100
These are single runs on a Shopify development store with Shopify’s sample snowboard products, recorded on 2 October 2026.
| Product | Score | Where it lost points |
|---|---|---|
| The Complete Snowboard | 70 / 100 | Complete Product schema. No FAQ (15) and not recommended by the AI model (15). |
| The Videographer Snowboard | 33 / 100 | Product schema present, but the text left out buyer details like sizing and binding compatibility (20). No FAQ (15), not recommended (15). |
| The Out of Stock Snowboard | 10 / 100 | Empty description, so it lost points for that and for unanswered buyer questions. Product schema present but incomplete. No FAQ. |
Two honest notes. First, single runs move a little: across five evaluation runs, the complete snowboard scored between 65 and 70, because the model rated its specs “fair” three times and “good” twice. Second, the recommendation check is simple. The agent picks a buyer-style category, asks a Gemini model one general question, and checks whether the product name appears in the answer. In a run on 29 September 2026 the question was “What’s a good all mountain snowboard you’d recommend?”, and the answer named well-known brands instead. Sample products on a test store were never going to be named, so all three lost those 15 points.
The fixes that matter most
Going by where these three products lost points:
- Put the buyer’s questions into the description. Sizing, materials, compatibility, what is included. This was the biggest single loss for the middle product.
- Add FAQ content. A store FAQ page, or FAQs on the product itself.
- Make the structured data complete. The
Productblock needs a real description and offer details, not just a name. - Write alt text for product images.
- Check robots.txt so the search crawlers you want can reach your pages.
What nobody can promise
Fixing these does not guarantee that ChatGPT, Perplexity or Google AI will recommend a product. Each assistant decides on its own. The fixes remove what stops them from reading and understanding the product.
How to check your own store
- Open
yourstore.com/robots.txtand look for AI bot names such as OAI-SearchBot, GPTBot or PerplexityBot, and whether they are disallowed. - Read one product page as a buyer. Write down five questions a buyer would ask, and check whether the page answers each one in text.
- Run the product URL through Google’s Rich Results Test and check that the
Productdata has a description and an offer. - Ask ChatGPT or Perplexity the question your buyer would ask, and note whether your store appears and how it is described.
If you want this done for you, see AI discoverability and a sample report.