How structured retailer data supports discovery in AI search

Set aside the branding for a moment. Whatever an AI search feature is doing at the end, the input is largely familiar: it reads public web pages that a crawler was allowed to fetch, and it works with what those pages made explicit.

That means the practical inputs are the ones you already influence:

• Visible page text — the product name, the description, the specification table, the price, the availability statement.
• Structured markup — schema.org Product data, where it matches the visible content.
• Page structure — the title element, headings, tables and lists, which signal what is a fact and what is decoration.
• Link context — how you link to the page, and with what anchor text.
• Identifiers — GTIN and MPN, which allow your listing to be recognised as the same product a shopper saw elsewhere.

The important consequence: anything only implied is likely to be lost. A capacity that exists only inside a variant dropdown, a noise level that appears only in a product photograph, a stock status conveyed only by a greyed-out button — these are legible to a human looking at the page and absent from the extracted record.

Crawlability and index eligibility

This is the unglamorous foundation, and it is where more retail catalogues fail than anyone expects — usually as a regression rather than a decision.

• Product pages return 200 and are not blockedCheck robots.txt and meta robots tags on the live product template, not on a page you happened to have open.
• Key content does not require interactionSpecifications behind a tab that only loads on click may not be seen at all.
• Products are reachable through linksNot only through internal search or a filter interface. If the only route to a product is a search box, assume it is undiscoverable.
• Faceted URLs do not overwhelm the crawlCanonicalise or exclude filter and sort permutations so crawl budget reaches actual products.
• Sitemaps list canonical product URLsCurrent, complete, and reflecting the canonical form.
• The crawl posture is reviewed after every releaseAn accidental disallow or a template-level noindex can remove an entire category from every downstream system in one deployment.

Product structured data that matches visible content

Schema.org Product markup remains worth implementing, with one rule above all others: the markup must describe what the page actually shows. Google's guidance is explicit that structured data should reflect visible content, and mismatches risk losing rich-result eligibility entirely — a worse outcome than not marking up at all.

Two habits prevent most of these: validate markup after every release, because template changes break it routinely; and treat markup as a projection of your product data rather than as a separate artefact maintained by hand. Hand-maintained markup drifts from the page within a quarter.

Internal links and descriptive page titles

Internal linking does two jobs at once: it makes products reachable, and it tells a reader — human or machine — what the destination is about.

• Descriptive anchor text. “Acme SlimWash 45 slimline dishwasher” carries meaning; “view product” carries none. Repeated across a category, the difference is substantial.
• Title elements that name the product specifically. Brand, product name and the distinguishing variant attribute. Title templates that produce “Dishwashers | Buy Online | Retailer” on every page waste the single most-read line on your site.
• Category and buying-guide pages that link into products. These give a crawler a route and give a reader context about how products relate to each other.
• Related-product links. Comparable models, accessories and successor products establish relationships that would otherwise have to be inferred.
• Headings that mark real structure. A specification section under a heading called Specifications is easier to extract than the same table under a styled div.

Stable product and variant identity

Identity is what allows your product to be recognised as the same product a shopper is looking at elsewhere. Without it, every other improvement is attached to something that cannot be matched.

• Publish GTINs wherever the manufacturer assigns one. It is the only genuinely cross-retailer key.
• Publish MPN or model number for categories where GTIN assignment is inconsistent.
• Give each purchasable variant its own canonical URL where price or identifier differs.
• Keep URLs and identifiers stable across site releases. Where a URL must change, use a permanent single-hop redirect to the equivalent product, not to the category page.
• Never reuse a retired identifier. It corrupts every downstream record built from it.

The cost of getting this wrong is quiet. A retailer with strong specifications and no GTINs simply does not appear in comparisons that other retailers appear in, and there is no error message anywhere to tell them why.

Freshness, availability and price context

Downstream systems treat retailer prices as observations with a timestamp — simlir returns a price snapshot with the retailer it was seen at and an as_of date, for exactly this reason. That convention works in your favour, but only if your page is accurate when it is read.

• Push price changes and stock-outs to the live page quickly. The gap between your system and your page becomes a wrong number in someone else's answer.
• State availability explicitly rather than implying it through a button state.
• Distinguish discontinued from temporarily unavailable. They deserve different treatment everywhere.
• Keep discontinued product pages resolving sensibly — with a clear status, or redirected to a successor. Silent 404s remove accumulated context.
• Show unit pricing where the category calls for it. It supports comparison across pack sizes and is a legal requirement in many UK categories.

Measuring referral and search visibility

Measurement is harder here than in classic search, but it is not the black box it is often described as. Two concrete instruments exist.

The most useful metric is the least glamorous one: the percentage of products in a category carrying a GTIN, a durable hero image and the deciding specifications. It is entirely within your control, it moves when you do the work, and it is the input every downstream system depends on. Track it alongside the visibility signals rather than instead of them.

What cannot be guaranteed

Anyone promising you placement in AI search results is selling something they cannot deliver. Here is the honest boundary:

• Inclusion is not guaranteed. No amount of structured data entitles a page to appear in any search feature or assistant answer.
• Ranking is not purchasable or configurable. There is no submission process or setting that determines position.
• Rich results are eligibility, not entitlement. Google states this directly about structured data.
• Behaviour changes. Surfaces, formats and citation conventions are moving quickly, and a tactic tuned to today's behaviour may not survive the year.
• Attribution will stay imperfect. Plan for directional measurement rather than precise channel accounting.

Things you can safely ignore

Google has been unusually direct about tactics that do not help on its surfaces, which is useful because several of them are actively sold as services:

• llms.txt and similar files. Google Search does not use them. Maintaining one for other systems is fine and will neither help nor harm your visibility in Google Search.
• “Chunking” content into small pieces. Not required. There is no ideal page length.
• Rewriting pages specifically for AI systems. They understand synonyms and intent, so you do not need to capture every phrasing a shopper might use.
• Overfocusing on structured data. It is not required for generative AI features. Keep using it because it supports rich-result eligibility in ordinary Search — not because it is an AI tactic.
• Chasing inauthentic mentions. Spam systems and quality ranking both apply to generative features.

Set against that list, the recommendations in this article are deliberately boring: crawlable pages, explicit facts, stable identity, accurate markup, prompt freshness. That is the work.

What you can control is whether your products are readable, matchable and accurately described when they are read. That is a smaller claim than most of what is written about this subject, and it is the part that actually compounds — because the same work improves your own site search, your feeds, your comparison partners and your conventional SEO at the same time.

Frequently asked questions