The 11 Product Page Fields Language Models Actually Read (And Which Ones They Ignore)

The 11 Product Page Fields Language Models Actually Read (And Which Ones They Ignore)

I’ve now pulled apart the product data of several hundred Shopify stores, and I’ve reached a conclusion that annoys me slightly: most product page optimisation advice is written for a version of the internet that stopped existing.

Keyword density in the description. Character counts for the title tag. Making sure the H1 contains the target term. None of it is wrong, exactly. It’s just aimed at a system that reads a page as a document, when the systems that increasingly decide whether you get recommended read a page as a record.

So here’s the field-by-field version. Eleven fields, in the order I’d fix them, plus the ones I’ve stopped bothering with.

Why the PDP is now the landing page

One number reframes this entire discipline. In Shopify’s Q1 2026 data, **55% of AI-referred sessions start on a product detail page**, versus **20% for organic search**.

Organic traffic has always arrived through the front door — homepage, collection page, blog post — and been merchandised toward a product. AI traffic arrives at the product. There’s no journey to design, no upsell path to guide, no category page to persuade with. The PDP is the entire experience, and increasingly it’s the entire experience for a shopper who has already been told by a machine that this is the right product.

Which means the product page’s job has changed from *persuading a browser* to *being legible enough to have been chosen in the first place*.

Fields 1–4: identity

**1. Product title.** Structure it as attributes, not marketing. `Brand + Product + Key Differentiator + Variant` beats a clever name every time. “Northwind Merino Base Layer — Midweight 260gsm — Men’s Navy” tells a model four things. “The Explorer” tells it nothing, and no amount of description recovers what the title threw away.

**2. GTIN / MPN / barcode.** The single most under-filled high-value field in ecommerce. A GTIN lets a system match your product to the same product elsewhere — reviews, price comparisons, spec sheets, retailer listings. Without it your product is an orphan record that has to be understood entirely from your own copy. With it, it inherits a web of corroboration. If you resell branded goods and haven’t populated barcodes, stop reading and go do that.

**3. Brand / vendor.** Populated properly and consistently. “Northwind”, “Northwind Outdoors” and “northwind” across your catalogue are three brands as far as a matcher is concerned.

**4. Variant naming.** Variants should describe the axis they vary on. “Blue / Large” is good. “Option 1 / Option 2” is a data loss event. Colour names in particular: “Slate” is a marketing name, “Slate (grey)” is a marketing name a machine can act on when someone asks for a grey jacket.

Fields 5–8: substance

**5. Description structure.** Shopify’s own semantic search draws directly on product descriptions — and here’s the part most merchants don’t know, on **image data too, including text within images and colours**. If Shopify’s own index is reading those, assume everything downstream is.

Write descriptions as short declarative sentences with explicit subjects. “This jacket is waterproof to 10,000mm and seam-sealed throughout” extracts cleanly. “Whatever the sky throws at you, you’ll stay dry” does not. Keep a paragraph of brand voice at the top if you must. Put the facts underneath in plain sentences.

**6. Structured attributes / metafields.** The single biggest lever. Every question a knowledgeable customer would ask before buying should exist as a discrete field: material, weight, dimensions, capacity, compatibility, power, care, certification, origin, what’s in the box.

The critical thing to understand is that a blank attribute isn’t a weak signal. When an agent filters for a specification, a product with no value for that field is excluded, not demoted. There’s no partial credit. I’ve seen 4,000-SKU catalogues averaging six populated attributes where the category norm is twenty. That store is syndicated to every AI channel and invisible on every specific query.

**7. Images and alt text.** Given that Shopify explicitly reads text within images and colour data, alt text is not an accessibility box-tick — it’s a description field you’re leaving empty. Describe what’s shown, including colour and context. Also: if a critical spec exists only inside a graphic on your image carousel, it’s semi-legible at best. Put it in a field.

**8. Availability and price accuracy.** Shopify Catalog verifies pricing and inventory in real time before syndicating. A store whose schema claims in-stock while the storefront says sold out is teaching every crawler that your data can’t be trusted, and trust is not a per-product property.

Fields 9–11: trust

**9. Product and Offer schema.** JSON-LD, complete, matching the visible page. Name, image, description, brand, sku, gtin, offers with price, priceCurrency, availability and priceValidUntil. Most themes emit some of this. Very few emit all of it correctly, and almost none keep it in sync after an app is installed. Verify in Google’s Rich Results Test rather than assuming, and re-verify after every theme change.

**10. AggregateRating and Review markup.** Only if the reviews are real and on the page — faking this is both a policy violation and easily detected. Where it’s genuine, it’s one of the strongest trust signals you can emit, and Shopify names real-time trust signals and engagement metrics explicitly among the components of GEO.

**11. Shipping, returns and delivery data.** Increasingly asked about directly. “Free returns within 30 days” as structured data rather than a link to a policy page in the footer is the difference between an assistant being able to answer the objection and having to say it doesn’t know.

The three fields I’ve stopped worrying about

**Meta description as a ranking play.** Still worth writing — it’s your snippet copy and it does get read — but the hours some teams spend agonising over it would be better spent populating three metafields.

**Keyword density anywhere.** I have never once seen a product page fail to be cited because a term appeared four times instead of six. I see products fail constantly because the fabric weight isn’t recorded.

**Long-form PDP content for its own sake.** A 900-word product description full of lifestyle prose adds noise. The same 900 words split into a 100-word intro plus fifteen populated attributes is a materially better record.

A copy-paste audit checklist

Run this on your top 20 products by revenue:

– [ ] Title contains brand, product type and key differentiator
– [ ] GTIN or MPN populated
– [ ] Vendor consistent across catalogue
– [ ] Variants named by axis, colours include a plain-English colour word
– [ ] Description contains factual declarative sentences, not only brand voice
– [ ] Every attribute a customer would ask about exists as a metafield
– [ ] All images have descriptive alt text including colour
– [ ] No critical spec exists only inside an image
– [ ] Product + Offer schema validates and matches the live page
– [ ] Price and availability in schema match the storefront right now
– [ ] Review markup present and genuine
– [ ] Shipping and returns available as data, not just a policy link

If you get through that on twenty products and find fewer than fifteen gaps, you’re in better shape than the vast majority of stores I’ve looked at.

This is exactly the work I automated into GEOptimisation, because after doing it manually enough times it became obvious it’s mechanical rather than strategic. But the checklist works fine on its own, and doing it by hand once for your best sellers will teach you more about your own catalogue than any tool’s dashboard will.

*Sources: Shopify Q1 2026 commerce data; Shopify Help Centre, Search & Discovery documentation; Shopify, “Agentic Commerce on Shopify” (2026). Verify current structured data requirements at developers.google.com before implementing.*