Five questions an agent checks before it'll recommend a product.
An AI agent shopping on someone's behalf doesn't browse a product page the way a person does. Before it will recommend anything, it needs five separate questions answered: does the exact variant exist, is it currently in stock, will it work for this specific case, what happens if it's wrong, and will it arrive in time. A page built to persuade a browsing human answers few of those directly.
8 min read
A product page is built to persuade a browsing human
Someone asks an AI agent to find a waterproof jacket, women's size 14, that'll keep her dry on a hike in the Brecon Beacons this weekend, and can arrive by Friday. The jacket's product page might have excellent photography, a founder's story about the fabric, a run of five-star reviews. None of that answers the question the agent is holding: does a size 14 exist in this colour, is it in stock right now rather than usually in stock, and will it arrive by Friday to this postcode.
Ecommerce design has spent two decades optimising one thing: the journey of a person already looking at a product, deciding whether to want it. Photography, brand story, social proof, all aimed at persuasion. None of it persuades an agent, because persuasion was never the job it's doing. It's checking a small number of facts directly, and if even one of them isn't answered without inference, it moves to the next retailer rather than guess on the buyer's behalf.
The five questions an agent resolves before it recommends anything
The variant has to exist, not generically as 'the jacket' but as this exact size and colour, confirmed specifically. Stock is a different question again: a true, current count for that variant, not a general in-stock badge left over from a busier day. Whether it'll actually work for this case depends on fit, use and compatibility, whatever the specific ask turns on, and what happens if it doesn't is answered by returns terms stated plainly enough to spell out a real risk. And will it arrive in time? That's a delivery estimate tied to the buyer's actual location, not the blanket '2-3 days' printed regardless of where it's actually going.
Miss any one and the agent either asks a follow-up question the buyer didn't want to answer, or drops the product from consideration rather than risk a bad recommendation.
The five hold regardless of category. A parent buying a cot mattress needs the exact size, 120 by 60 centimetres rather than the nearly identical 140 by 70, confirmed in stock, confirmed safe for the cot they actually own, with a return route if the fit's wrong, arriving before the baby does. Swap jacket for mattress and the five questions an agent resolves don't change. Only the specific facts filling them in do.
Where this differs from "write better copy for AI"
The generic advice going round is to write plainer product copy so a language model can parse it. That's the shallow half of the job. The deeper problem isn't how the copy reads, it's whether the product data underneath holds the answer to a real decision, structured so an agent can check it rather than infer it from marketing prose. A jacket buyer wants to stay dry this weekend, not a page optimised for a crawler. Evidence a fact is true beats a well-written claim that it might be, every time an agent has to decide whether to stake a recommendation on it.
Where this shows up for a business
An agent can check the specific answer to each of the five questions directly, rather than infer one from a category page or a generic in-stock badge. Selfe connects the live variant, stock and delivery data straight from a retailer's own systems, and surfaces the fit and returns facts, the sizing notes, compatibility details and actual return terms an agent needs to weigh, rather than the vague reassurance most catalogues default to. The rest of this hub goes through each of the five in more detail, one spoke at a time.
Does this mean we need to rewrite our product pages?
Not necessarily. The gap is usually in the underlying data, not the prose, whether the five questions have a checkable answer behind the page.
Which of the five matters most?
It depends on the category. A made-to-order item lives or dies on delivery time; a sized item lives or dies on variant and fit. This hub's other spokes go through each in turn.
What makes one listing easier for an agent to sell than another.
Given two nearly identical products, an AI agent recommends the one with fewer unresolved questions, not the one with better marketing copy. The gap between them is almost always one missing, narrow fact, not a missing feature.
Returns and sizing: the two questions agents ask that catalogues rarely answer.
Sizing and returns are the two questions that decide whether an AI agent will risk a recommendation at all, and most catalogues answer neither with any real specificity. "True to size" and "easy returns" aren't answers, they're the absence of one.
Variants, stock and delivery: what an agent checks before it recommends.
Three facts gate whether an AI agent will recommend a product at all, regardless of anything else about it: does the specific variant exist, is it there right now, and will it arrive in time. Get any one wrong and the rest of the listing doesn't matter.
What a product feed needs that a product page doesn't.
A product page is written for a person looking at one item. A product feed is what an AI agent reads to compare many at once, and it needs a different kind of precision: structured, decision-relevant fields, not the same copy reflowed into a spreadsheet.
What "buyable" needs beyond stock.
An AI agent has already checked that a product is in stock. Buyable asks more of it than that: is the price the one that'll actually be charged, and can this specific item reach this specific buyer at all. Miss either and "in stock" stops meaning anything useful.
Multi-brand retailers: keeping every label distinct when an agent is comparing.
A retailer selling forty brands under one site template often ends up saying the same thing about all of them. An AI agent comparing two products from two different brands needs the fact that's true of each one, not a house style that's quietly smoothed the real differences away.