The Product Feed Fields AI Uses to Decide What to Recommend (And Which Ones You're Missing)

By Steve Merrill, Founder of WRKNG Digital — June 21, 2026

What fields does AI actually use to evaluate a product?

AI shopping assistants don't browse your store. They work from structured data — and if your feed doesn't have what they need, your product doesn't exist to them.

That's not a metaphor. When ChatGPT recommends a product, it's pulling from indexed product data that's been validated, structured, and cross-referenced. The stores that show up aren't just "on the internet." They have the right fields populated, in the right format, at the right level of detail.

I've audited 40+ Shopify stores in the last six months. The pattern is almost identical every time: feeds exist, but they're thin. Title, price, one blurry description, maybe a color variant. That's it.

Here's what AI actually needs — and what it's almost certainly not getting from your feed right now.

The fields every AI shopping system looks for first

The baseline fields are ones you probably think you have covered. You don't — not really.

Product title. Most Shopify titles are written for humans skimming a category page. "Blue Merino Sweater." That's not enough. AI systems need brand + product name + key attributes in the title itself. Google's Merchant Center guidelines recommend this structure: Brand + Product Name + Key Differentiating Attributes. Most stores skip the brand and the attributes. Both matter.

GTIN (Global Trade Item Number). This one surprises people. If your product has a barcode — and most manufactured products do — AI systems use GTINs to match your product against the same product listed elsewhere. Without it, you're an unknown quantity. According to Google's product identifier policies, products with valid GTINs receive better placement in Shopping results. The same logic applies downstream to AI-powered recommendations.

Brand. A required field that's often left blank or filled with the store name instead of the actual manufacturer brand. Not the same thing. AI systems use brand as a trust signal and a disambiguation tool.

Product category (Google product taxonomy). This is where a lot of feeds break down. The taxonomy has over 5,400 categories. Most stores pick something vague near the top — "Clothing & Accessories" — and call it done. That's like telling a librarian your book is "fiction." Technically true. Useless for placement.

Which product feed fields are almost always missing?

The required fields above are just the floor. AI recommendation systems — especially conversational ones like ChatGPT and Perplexity — need more to make confident recommendations.

Semantic descriptions that answer buyer questions

Most product descriptions are feature lists. "Soft merino wool. Machine washable. Available in 6 colors." Fine for a product page. Not useful for an AI trying to answer "what's the best sweater for someone who runs warm."

AI systems are trying to match your product to a conversational query. They need descriptions that read like answers. Who is this for? What problem does it solve? How does it compare to alternatives? Under what conditions does it perform best?

We ran a test last month where we rewrote 15 product descriptions from feature-list format to answer-style format — same products, same stores. The AI citation rate for those products in Perplexity Shopping responses went from 0 to 4 out of 15 within three weeks. That's not a controlled study. But it's not nothing, either.

Structured product attributes

Attributes like material, pattern, age_group, gender, size_type, and fit are optional fields in most feed specs. They don't feel urgent. Nobody's yelling at you to fill them in.

But these are exactly what AI systems use to filter products when responding to specific queries. "Gifts for a 10-year-old boy." "Activewear for women with a relaxed fit." If those attributes aren't in your feed, the AI can't make the match — even if your product is perfect for the use case.

Missing. Almost universally.

Product ratings and review data

AI recommendation systems factor in social proof. Not the reviews on your site — the structured review data in your feed, submitted through a product ratings program like Google's or surfaced via third-party integrations with platforms like Yotpo, Stamped, or Judge.me.

A 4.8-star product with 200 reviews doesn't automatically get that signal into your feed. You have to connect it. Most stores haven't.

Custom labels and use-case tagging

Custom labels (custom_label_0 through custom_label_4 in Google's spec) are free-form fields you can use however you want. Smart operators tag products by use case, occasion, customer type, or season. "gift_under_50." "summer_outdoor." "beginner_friendly."

These don't directly feed into AI recommendation engines the way attributes do — but they're part of a broader structured data picture that makes your catalog more machine-readable overall. And that matters as AI agents start doing more autonomous product discovery.

Why does your product title matter more than you think?

The title is the first field every AI system processes. It's also the field most stores treat like an afterthought.

Think about what happens when someone asks ChatGPT: "What's a good lightweight running shoe for someone with wide feet?" ChatGPT isn't searching your store. It's working from indexed product data that includes titles, attributes, and descriptions. If your product title is "Men's Running Shoe — Style 42B," you're invisible to that query. If it's "Brooks Ghost 15 Men's Wide Running Shoe — Lightweight, Neutral Cushion," you're at least in the conversation.

That's not an exaggeration. That's how the matching works.

How do you find out what your feed is actually missing?

Start with Google Merchant Center. If you have a Shopping feed connected, the diagnostics tab will show you missing or flagged fields for every product. It's free. It's brutal. And it's the clearest picture most store owners have ever had of their product data quality.

Export the feed itself — most Shopify apps like Simprosys or DataFeedWatch give you a CSV — and just look at the columns. Count how many are blank. For most stores, more than half of the "recommended" attribute columns are empty across their entire catalog.

The fix isn't fast. But it's not mysterious either. You're filling in structured data that already exists somewhere in your business — you just haven't put it in the feed yet.

Where to start if your catalog is large

Don't try to fix everything at once. Prioritize your top 20% of products by revenue. Get GTINs added. Rewrite descriptions for your ten best sellers. Fill in the structured attributes for your top category.

That's enough to see a measurable difference in feed quality scores — and in how AI systems engage with your products when they're evaluating recommendations.

The stores winning in AI-driven discovery aren't necessarily the biggest or the most established. They're the ones whose data is cleanest.