AEO for ecommerce: how to optimise product pages for AI shopping

AI shopping agents do not read your product page the way a shopper does. They parse the markup, weigh the evidence and hand back a recommendation - or they skip you. This is a practical guide to product schema, reviews and Q&A, and the support pages that quietly win a large share of citations, so assistants extract, trust and recommend your products.

AI has become a checkout, not just a research tool

For most of ecommerce history, the job of a product page was to persuade a human. Someone landed on it, scanned the images, read a few reviews and clicked buy. The page was the destination. That assumption is quietly breaking. A growing share of buyers now start with an assistant - "what is the best cordless vacuum under 200" - and the model does the scanning for them, then hands back a shortlist with a recommendation attached.

The scale of the shift is no longer speculative. Adobe Analytics found that traffic to US retail sites from generative AI tools grew 693% year on year during the November to December 2025 holiday season, and that AI-referred shoppers converted 31% more often than visitors from other sources. Salesforce, tracking Cyber Week 2025, reported that AI agents influenced roughly 67 billion dollars in global sales, touching about one in five orders worldwide.

When an assistant both recommends the product and, increasingly, routes or completes the purchase, being named in the answer moves very close to being bought. The uncomfortable corollary: if the model cannot read, trust or extract your product data, you are invisible at the exact moment the decision is made. Optimising the product page for a machine reader is now a commercial job, not a technical nicety.

Product schema is the machine-readable layer AI reads first

An AI shopping agent does not experience your product page the way a shopper does. It parses the underlying markup, and structured data is the cleanest signal it has. Product JSON-LD (schema.org/Product) is how you hand the model the facts in a form it does not have to guess at: name, brand, images, price, currency, availability and reviews, spelled out unambiguously.

Google's own merchant listing requirements are a useful floor here, because the same structured fields that qualify you for Shopping and rich results are the ones assistants lean on. The core required fields are the product name, an image, and an offer containing price, priceCurrency and availability. Beyond that floor, three fields do disproportionate work for machine trust and disambiguation:

• Brand. A clear brand value ties the product to a known entity, which strongly influences how it is ranked and recommended rather than treated as an anonymous listing.
• GTIN or MPN. A global identifier lets the model reconcile your page with the same product elsewhere - marketplaces, review sites, price comparison - so it can corroborate rather than treat your claims in isolation.
• AggregateRating and Review. Structured ratings give the model a number it can cite directly, instead of hoping it infers sentiment from prose it may not parse.

Two rules matter more than completeness. First, the structured data must match the visible page exactly; a price in JSON-LD that differs from the price on screen is a trust failure that can get the whole page discounted. Second, keep availability and price genuinely current. Google's AI shopping surfaces weight real-time accuracy heavily, and a model that catches you serving stale "in stock" data has a reason to prefer a competitor it can trust.

Curious how AI engines describe your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

The catch: much of your catalogue is not machine-readable

Here is the gap most teams underestimate. In mid-2026 Adobe scored retail product pages at only 66% on machine readability using its content visibility checker, meaning roughly a third of product content is effectively invisible to large language models. The worst-performing categories - apparel, sporting goods, grocery, furniture - sat closer to half.

The usual culprits are familiar to anyone who has audited a modern storefront. Specifications buried in JavaScript tabs that only render on click. Key details baked into product images as text the model cannot read. Sizing, materials and compatibility notes living in an accordion the crawler never expands. Reviews loaded from a third-party widget in an iframe the agent treats as a separate, uncited document.

"If a third of the page is invisible to the model, you are being judged on the two-thirds it happened to read."

The fix is not glamorous but it is decisive: get the load-bearing facts into server-rendered HTML and mirror them in structured data. If a spec determines whether the product is recommended - waterproof rating, cable length, allergen, fit - it should exist as plain text and as a schema field, not only as a pixel in a lifestyle photo or a value hidden behind a tab interaction.

Reviews and Q&A are the evidence models trust most

Assistants are cautious about a brand's own marketing copy, because every brand claims to be the best. What moves a recommendation is corroboration - independent voices saying the same thing. That is why review content and customer Q&A punch far above their weight in AI shopping answers.

Make reviews extractable, not just visible. Render review text and ratings in HTML rather than injecting them client-side, and expose the aggregate as structured data so the model can quote a specific score. A body of substantive reviews that mention concrete attributes - "runs half a size small", "battery lasted the whole weekend" - gives the assistant the exact phrasing it needs to answer a buyer's real question, and those attribute-level mentions are frequently what gets your page pulled into the response.

Customer Q&A does something reviews cannot: it captures the objection in the buyer's own words. "Does this fit a standard UK plug?" "Is it dishwasher safe?" These are the long-tail buying questions people now type into ChatGPT verbatim. A product page that already contains the question and a clear answer is a near-perfect match for the query, and it saves the model from having to infer or hedge. Seed genuine questions from your support inbox rather than inventing them, and answer them plainly.

Support, sizing, shipping and returns pages do more than you think

This is the least intuitive finding, and the most useful. When you look at which ecommerce pages AI engines actually cite, a striking share are not product or category pages at all. Aleyda Solis's analysis across five ecommerce subverticals found that assistants cite a broad "evidence layer" - size guides, support and repair articles, policies, returns and shipping pages, buying guides - because those are the pages that resolve the buyer's uncertainty before, around and after the purchase. Shopping was the vertical with the highest citation density of any she studied, roughly twice that of real estate.

The priorities differ by category in ways that should shape where you invest. Fashion and apparel sees size guides and returns policies cited heavily, because fit and the cost of getting it wrong are the dominant anxieties. Consumer electronics leans on support and repair documentation. General marketplaces surface policy and store or location pages. In each case the assistant is answering the question sitting behind the purchase, and it cites whoever answers it best.

The practical implication is that a returns page written as legal boilerplate is a wasted asset. A returns page that states the window, who pays for return shipping, how refunds are issued and how long they take - in plain, extractable language - is exactly the kind of page an assistant will cite when a buyer asks "how easy is it to return something to this shop". The same goes for shipping timelines, warranty terms and sizing. Treat these as answer pages, not compliance pages.

How to structure a PDP so an agent extracts and trusts it

Pulling this together, an AI-ready product page has a recognisable shape. The facts a buyer needs to make a decision are present as text and as structured data, they agree with each other, and they are corroborated by evidence the model can reach without leaving your domain.

A workable sequence for a single PDP looks like this. Lead with a concise, factual description that states what the product is, who it is for and its two or three defining attributes, in the first paragraph, because models weight the top of the page. Follow with a complete specification block in plain HTML. Expose reviews and their aggregate rating server-side. Answer the common buying questions inline. Then link clearly to the sizing, shipping and returns pages so the assistant can follow the thread of uncertainty to a page you control, rather than to a marketplace or a Reddit thread that mentions your product in passing.

Underneath all of it, the Product JSON-LD carries the same facts in machine form, with brand, a GTIN, current price and availability, and the aggregate rating. Consistency across the visible page, the structured data and your marketplace listings is what earns trust: when the model sees the same identifiers, price and rating in three places, it treats your product as a real, verifiable thing and is far more willing to recommend it.

You cannot fix what you cannot see

The final piece is measurement, because none of this is a one-off. AI answers are volatile: the model that recommends you this month may favour a competitor next month after a crawl refresh or a new batch of reviews. The only way to know whether your PDP work is landing is to track whether the engines actually name and cite your products for real buying questions - and where a rival wins the recommendation instead.

That is a different measurement problem from classic rank tracking. You are watching for whether ChatGPT, Gemini, Perplexity and Google's AI surfaces mention you, which of your pages they cite, and how that shifts as you improve schema, reviews and support content. Get that feedback loop running and the PDP work stops being guesswork: you can see which changes move you from unmentioned to recommended, and defend the position once you have it.