Retailers lose AI visibility at the product data layer, and then spend the budget on content. An assistant answering “waterproof walking boots, wide fitting, under £150” is matching four attributes. If those attributes exist only as prose in a description, your product cannot be matched, however good the buying guide next to it.
This page is retailer-level. If you sell a small range direct to consumer, the product-level version is how to get products recommended by ChatGPT, and the service page is GEO for e-commerce.
What a retail answer is actually made of
When somebody asks an assistant for a product recommendation, four things happen in sequence, and retailers typically only invest in the last one.
Qualifier extraction. The question is decomposed into attributes. Category, price ceiling, use case, constraint, brand preference. Each one becomes something to match against.
Attribute matching. Products are matched on those attributes as data. Not as adjectives in a description. As fields.
Corroboration. Independent sources are checked. Reviews, comparison articles, marketplace listings, retailer aggregators. This is where a model decides whether to trust the match.
Answer assembly. Two or three products named, usually with a reason attached to each.
Almost every retail visibility failure happens at stage two, and it is invisible from the inside, because the information genuinely is on the page. It just is not in a form that can be matched.
The attribute gap
Here is the test. Take your ten best-selling products. Write down every attribute a customer might specify when asking for something like them. Then check which of those exist as structured data rather than as sentences.
For footwear that might be: size range, width fitting, material, waterproof rating, sole type, weight, intended terrain, break-in period, care requirements. Most retailers have three of those in structured form and the rest in a paragraph, or missing entirely.
The paragraph is not useless. It is simply much weaker than a field, because a field is unambiguous and a sentence needs interpretation. Given two candidate products and a shopper who specified a wide fitting, a system will favour the one where “wide fitting” is a value rather than an adjective in the middle of a description.
The work is unglamorous: complete the feed, one category at a time, starting with the categories that carry the most revenue. It is data entry. It is also the highest-return AI visibility work a retailer can do, and it is why this page opens with it rather than with content.
Schema, and what most retailers are missing
Product schema is widely implemented and rarely complete. The common version carries a name, an
image and a price. The useful version carries:
name,description,imagebrand, andmanufacturerwhere differentgtin,mpnorsku, which is how a model reconciles your listing with the same product elsewhereofferswithprice,priceCurrency,availabilityandpriceValidUntilaggregateRatingandreview, but only where those are genuinely published and visibleadditionalPropertyfor the attributes that have no dedicated field, which is where most of the gap above gets closedisVariantOfandhasVariantfor size and colour ranges
Two cautions worth stating.
Stock and price accuracy matter more here than in traditional search. A model that recommends an out-of-stock product produces a bad outcome the platform gets blamed for, so stale availability data is a reason to be deprioritised. If your feed updates nightly and your stock moves hourly, that is a visibility problem as well as an operational one.
Never mark up ratings the page does not display. It is detectable, and it is a manual action as well as a credibility problem. The general standard is in the schema markup guide.
The marketplace tension
This one is uncomfortable and it does not fully resolve.
If you sell on Amazon, eBay or a sector marketplace, those listings are strong, well-crawled sources for the same products. An assistant may recommend the product and point at the marketplace rather than at you. You are competing with your own distribution.
There is no way to make a model prefer your site by asking. What you can do is be the better source on everything a marketplace listing cannot carry:
- Fit, sizing and compatibility advice, which marketplaces handle badly and shoppers ask about constantly
- Honest comparisons between products you sell, including who each one is wrong for
- Aftercare, servicing, spares and warranty specifics
- Stock reality, including lead times on made-to-order items
- The expertise layer: why this product for this use, written by somebody named
That is a content strategy, and it is the point at which content becomes the right investment rather than the premature one. It works because it answers qualified questions that a listing cannot, and qualified questions are where the valuable retail traffic sits.
Buying guides that actually get cited
Most retailer buying guides are unusable to a model because they refuse to choose. “Consider your budget and requirements” resolves nothing.
A citable guide does four things:
- Names specific products, with prices and the date the prices were checked
- States trade-offs explicitly. This one is lighter and less durable. That one is heavier and will outlast it
- Says who each product is wrong for. The single most reliably cited element, because almost nobody publishes it and a model answering “is X right for me” needs both directions
- Carries a visible updated date, because a price-bearing page with no date reads as stale
The general principles are in how to write content AI models actually cite.
Reviews, in retail specifically
Retail reviews do a job here that no other content can, because they are the only source describing the product in a customer’s own words on a surface you do not control.
What matters is the text. “Great boots” answers nothing. “Wore these on the Pennine Way in persistent rain, feet stayed dry, wide fitting was genuinely wide, needed about two weeks to break in” answers waterproofing, fit, terrain and break-in period, and it does so independently.
Ask for specifics at the point of use rather than the point of delivery. A review request three weeks after a walking boot arrives produces better content than one sent the day it ships, and better content is the entire point. The mechanics are in review platforms and trust signals.
Physical retail and the local layer
If you have shops, a large share of your AI visibility is local rather than catalogue.
“Where can I buy X near me” is answered by combining place data with availability. That means the Google Business Profile for each branch has to be complete and accurate, the store locator has to be crawlable rather than a JavaScript widget, and stock or range information at branch level is a genuine advantage where you can expose it.
Full mechanics in AI search for local business.
Agentic commerce, without the hype
Agents that browse, compare and complete part of a purchase are real, growing and not yet a large share of transactions. The honest position is that nobody knows the timeline.
What makes the question easy is that preparing for it requires nothing speculative. An agent needs complete structured attributes, accurate stock and price data, clean schema, a checkout that works without a human reading a modal, and reliable delivery information. Every one of those improves ordinary AI visibility, conversion and operations today.
So the recommendation is not to build for agents. It is to do the product data work you already needed to do, and note that it happens to be the same work. If the agent shift accelerates you are ready, and if it does not you have still fixed your feed.
The order for a retailer
| Priority | Work | Timescale |
|---|---|---|
| 1 | Attribute audit on top revenue categories. Fill the gaps in the feed | 4 to 8 weeks |
| 2 | Complete Product schema including identifiers and additionalProperty |
2 to 4 weeks |
| 3 | Stock and price feed frequency raised to match reality | Engineering |
| 4 | Review requests rewritten to ask for use-case specifics | Ongoing |
| 5 | Category and branch pages made crawlable without JavaScript | 2 weeks |
| 6 | Buying guides that name products and state trade-offs | Ongoing |
| 7 | Entity and off-site work: registers, directories, consistency | Ongoing |
Content is sixth. That ordering is the argument of this page.
MarGen’s pricing is published, the method is published, and the results are in the case studies.
AI Search for Retail: Common Questions
How do AI assistants recommend retail products?
They match structured product attributes against the qualifiers in a question, then check whether independent sources agree. A query for a waterproof walking boot in a wide fitting under £150 needs four attributes to be present as data. Retailers whose product pages carry that detail in prose but not in structured fields lose to competitors whose data is machine-readable.
What is the biggest AI search mistake retailers make?
Treating it as a content problem. Retailers commission buying guides while their product feed is missing materials, dimensions, compatibility and fit data. The guides get crawled and the products still cannot be matched to a qualified question, because the attribute a shopper asked about does not exist anywhere in structured form.
Does a retailer need product schema for AI visibility?
Yes, and more completely than most implement. Product schema with price, availability, currency, GTIN or MPN, brand and review data is the machine-readable version of your catalogue. Partial schema with only a name and a price is common and it means your products cannot be matched on the attributes shoppers actually specify.
Do marketplaces compete with a retailer’s own AI visibility?
They do, and it is a real tension rather than a solvable problem. Your listings on Amazon or eBay are strong, well-crawled sources for the product, and they can be recommended instead of you. The counter is to be the better source on things a marketplace listing cannot carry: fit advice, compatibility, comparisons, aftercare and honest limitations.
What is agentic commerce and should retailers prepare for it?
Agentic commerce is an AI agent completing part or all of a purchase on a shopper’s behalf, which shifts the buying decision to whichever product data is clearest to a machine. Preparing for it is not speculative work, because everything it requires, meaning complete attributes, accurate stock and price data and clean schema, already improves ordinary AI visibility.
How important are reviews for retail AI visibility?
Very, and the text matters more than the score, because assistants read reviews to answer qualified questions. Reviews that mention fit, durability, sizing and use case give a model evidence it can match to a query. Aggregate ratings alone answer nothing specific, which is why detailed reviews outperform higher averages.
Should retailers write buying guides for AI search?
Yes, but second. A genuinely useful buying guide that names specific products, states trade-offs and says who a product is not for is highly citable. It just will not compensate for missing product data, so the order is fix the feed, then fix the schema, then write the guides.
Does AI search matter for physical retail?
Yes, and differently. Assistants answering where can I buy X near me combine local place data with product availability, so a retailer with accurate location data and stock information is answerable and one without is not. For physical estates the local data layer usually matters more than the catalogue does.
Where to Go Next
Product-level and service pages: GEO for e-commerce · how to get products recommended by ChatGPT · do AI SEO services work for e-commerce
The layers this depends on: schema markup guide · review platforms and trust signals · AI search for local business
Plan it: AI visibility strategy · what happens to AI visibility when I delete product pages · GEO site migration
Have it delivered: MarGen is a UK GEO agency with published pricing, a published method and documented results.