Your customers are asking assistants what to buy: here is where those answers come from, and what decides whose product gets named.
How many shoppers are actually using AI to buy?
Enough that it has stopped being a question about early adopters. NIQ’s research found that 42% of consumers have used at least one AI tool to shop within the past month, 17% have used AI for product recommendations, and 5% have used fully autonomous AI agents to place orders on their behalf.
Read those numbers with their scope attached. NIQ’s findings come from “a monthly sample of approximately 500 consumers across the United States in early 2026”, a small sample in one market, so treat it as direction rather than precision. The direction is not ambiguous.
Most merchants can explain in detail how they rank on Google. Very few can explain how one of their products ends up inside an AI answer, or why a competitor’s did instead.
Where do AI assistants get product data?
Three ways, and a merchant controls each to varying degrees. That matters, because the effort you can usefully spend is not spread evenly across them.
Product feeds you submit
This is the route most like the one merchants already know. Google’s Shopping Graph “houses more than 35 billion product listings (and counting)”, and those listings are “updated constantly - either through information retailers share directly via Google Merchant Center or from what retailers and brands post across the web”.
OpenAI runs its own product feed under the Agentic Commerce Protocol. Merchants push full catalogue snapshots by SFTP, with a recommended cadence of at least daily. ChatGPT Ads has a product feed of its own. Perplexity has its own route: “we’re launching the Perplexity Merchant program to make it easy for large retailers to share their product specs with us”.
Three platforms, three feeds, one catalogue sitting behind all of them.
Product page crawlers read
Nothing needs to be submitted for this route to work, which is why most merchants underestimate it. OpenAI describes its shopping research as organic: “Results are organic and based on publicly available retail sites-reading product pages directly, citing sources, and avoiding low-quality or spammy sites.” Google draws on the same well, using “what retailers and brands post across the web”.
Your product pages are being read and judged today, whether or not anyone at your company has decided to act on AI. A thin product description is not neutral here. It is evidence.
Reviews, video and third-party pages
The Shopping Graph also pulls “relevant data from the web - like images, descriptions, reviews and YouTube videos”.
This is the route with the least direct control and often the loudest voice, and it matters most where your own data is weakest. When a description omits the material, fit or compatibility, a reviewer’s blog post or a video transcript fills the gap, and that source wasn't written to sell your product.
What does a product feed actually require?
Less than almost anyone expects, and that gap is where the trouble starts.
OpenAI’s product feed specification requires four fields in the feed header, an ID and a list of variants at product level, and, at variant level, an ID and a title. That is the whole requirement. Description, barcodes, price, availability, condition, categories, images and seller are all optional. A feed can pass validation completely and still tell an assistant almost nothing about what it is selling.
Yet the platforms are explicit that fuller data performs better. OpenAI’s own documentation states: “Recommended attributes-like rich media, reviews, and performance signals-improve ranking, relevance, and user trust.” Perplexity makes a near-identical argument for joining its merchant programme, promising “increased chances of being a ‘recommended product’ because the products will be in our index, and when we have more robust details, we can better determine if a product is high quality and relevant to a user’s query.”
Two platforms, independently, telling merchants the same thing: the bar to be listed is not the bar to be chosen.
Why does a thin catalogue lose the comparison?
Because an assistant answering a buying question is doing the job a category merchandiser used to do, and it can only weigh what it can read.
Ask an assistant for a waterproof jacket under $200 that packs into its own pocket, and it must resolve three constraints against a set of candidates. A listing carrying packed dimensions, weight, fabric, seam construction and a waterproof rating can be matched against all three. A listing carrying a name, a price and a paragraph of brand copy can be matched against one, and only loosely.
Neither product was rejected on quality. One was simply unusable as evidence.
So the question stops being “are we integrated with the right platform” and becomes “does our catalogue describe our products well enough to win a comparison we will never see”. That is merchandising work. It is measured in fields, not in connectors.
What does product data enrichment involve?
Enrichment is the unglamorous work of making a catalogue legible to a machine. In practice, it has five parts.
Attributes an assistant can compare
Dimensions, weight, material, compatibility, certifications, care instructions, what is in the box. Values that line up against a competing product. Brand copy reads well to a human and compares to nothing.
Identifiers that hold a product together
GTINs and MPNs (the barcode and manufacturer part numbers that identify a product across systems) plus a consistent variant structure are how an assistant recognises your product as the same one it saw on a marketplace and in a review, and how it merges ratings, pricing and availability into a single judgement. Optional in the specification; decisive in practice.
SKU-level coverage, not just hero products
Brand-level presence does not predict product-level selection. An assistant can speak warmly about a brand and then name three competitor products in the same answer. Coverage has to reach the long tail, where the specific questions land and where attributes are usually thinnest.
A refresh cadence that matches the feed
Feeds run on daily snapshots, and OpenAI retains a processed record for up to 14 days after it drops out of one. Enrichment that ran once is a snapshot that starts decaying the day it ships.
One source of truth across channels
The same product has to hold together across a shopping feed, an ads feed, a marketplace listing and your own product page. Where those diverge, one product starts looking like several, and none of them looks authoritative.
Taken together, that is a recurring data operation with an owner, not a tidy-up squeezed into the end of a quarter.
How do you keep AI-written product data accurate?
Most teams will use a model to write these attributes, because the alternative is writing thousands by hand. That is reasonable, and it carries an obvious failure mode: a model will confidently invent a fabric composition or a compatibility claim. Push that into a feed, and you have not improved your data; you have industrialised an error and shipped it to every surface the feed reaches.
Four controls keep it in check.
Confidence scores on generated values
Every generated value carries a score, so output you can trust is separable from output that needs a person to look at it.
A review threshold you choose deliberately
Auto-accept above a chosen bar, review below it. Where that bar sits is a business decision, not a default someone else sets for you.
Provenance for every value
Record what each value was derived from. You can then trace and correct a wrong value, rather than guessing a second time.
Extra scrutiny on regulated attributes
Materials, safety, compatibility and anything carrying legal weight get a stricter threshold than colour or style.
How do you know whether any of it worked?
Mostly, you don't, and this is the part merchants find hardest to accept.
Every route described above is one-way. A feed is submitted, and nothing comes back. A product page is crawled, and nobody is notified. A review is quoted, and no report is issued. You can do all of this properly and still not know which of your products entered an answer, on which assistant, or which competitor appeared instead. Assistants also disagree with one another, so a result on one says little about the rest.
That makes measurement a separate discipline from the work itself, and unlike the feed specifications, nobody has published instructions for it.
The platforms have told us, in some detail, what they read. None of them tell you what they picked. Until that changes, the only reliable thing a merchant controls is a catalogue good enough to win the comparison on its own.





