Anatoliy Dankov
CEO
A merchandiser checks how the catalog performs inside ChatGPT and finds a gap that shows up in no analytics report. The product is in stock, the price sits below two competitors who do appear, and the product page converts perfectly well when a human lands on it. The assistant still builds its shortlist from three other retailers. Nothing broke, no ranking dropped, and no error surfaced anywhere in the stack.
Products get skipped for one of two reasons. Either the record failed validation and never entered the index at all, or it entered with too few attributes to answer the question a shopper asked. Both failures happen inside the product feed, which is a separate system from the storefront a search engine crawls.
The channel is small but no longer marginal. Salsify's 2026 Consumer Research report, based on a survey of roughly 3,000 shoppers across the US, UK and Canada, found that 22% now use AI search tools for product research, putting them ahead of review sites at 19% and forums at 14%.
The same report puts a second number next to it: only 14% of shoppers trust an AI recommendation enough to buy on it alone, and 31% say detailed descriptions and specifications are what convince them. Visibility inside the assistant and completeness of the underlying data turn out to be the same problem.
The selection happens before a shopper ever reaches a product page, and it runs on a different input than search ranking does. An assistant building a shortlist works from a structured catalog file the merchant supplies, weighs the attributes in it against the question asked, and returns a handful of products with prices and availability attached.
The outcome turns on whether the record entered the index at all, how completely it describes the product, and how much supporting evidence sits alongside it.
Product Feed vs Product Page: Why Crawling Your Storefront Is No Longer EnoughSearch optimisation was built around a page a crawler visits. Titles, copy, markup, and links all shape what an engine understands about a product.
None of that governs what an assistant sees. Product data in this channel is not crawled. The merchant pushes a structured file to a secure endpoint, and that file is what the assistant reads for pricing, availability, and every other value it reports.
Two consequences follow. A product page that ranks well can still be invisible inside an assistant, because the two systems read different inputs. And ownership moves: marketing owns the page, but whoever owns catalog management now owns the feed, which in most retailers means the content or merchandising team. [лінк 3 → catalog management solutions hub]
Results in this channel are not ads and are not influenced by paid placement. They rest on feed quality, product relevance, and user context.
The OpenAI Product Feed and What ChangedThe change is a reversal of direction. A search engine came to the storefront and took what it could parse. An assistant waits for the merchant to deliver a structured file, and a product absent from that file does not exist inside the assistant regardless of how well the storefront is built.
Merchants share that data with OpenAI through a secure, regularly refreshed feed in CSV or JSON, covering identifiers, descriptions, pricing, inventory, media and fulfilment options. OpenAI ingests the file, validates the records, and indexes the metadata for retrieval and ranking inside ChatGPT.
Merchants already running a Google-compatible product data format get a head start, because OpenAI will use that formatting. The mapping is familiar, and that familiarity is exactly what causes teams to assume the work is done.
One scope note worth keeping in view: selected US merchants can submit product data today, with shopping and instant checkout rolling out to more merchants and to countries outside the United States through 2026.
How to Submit Your Products to ChatGPTThe onboarding path is short and mostly mechanical.
The mechanical part ends there. Everything that decides whether a product gets recommended happens in the catalog behind the file.
Required fields exist to make price and availability render correctly. Recommended attributes such as media, reviews and performance signals improve ranking, relevance and user trust.
The distinction produces two different failures. Miss a required field and the record never enters the index. Supply every required field and skip the recommended ones, and the product is indexed but rarely selected.
Required, and they decide whether the product exists:
Recommended, and they decide whether the product is chosen:
Completeness and quality get measured with the same dashboard and behave nothing alike. A thin record limits how many questions a product can answer, and the damage is bounded by the questions it misses. A wrong record answers the question confidently and incorrectly, and the damage arrives later, in a different department's numbers.
Most catalog audits chase the first problem because it is visible. Percentage of filled attributes is easy to count and easy to improve. Product data quality has no equivalent metric, which is why a catalog can reach ninety percent completeness and still lose money on the ten percent of values nobody verified.
The Cost Lands After the RecommendationBeing surfaced by an assistant is not the end of the journey. Salsify's research puts the share of shoppers who verify an AI answer elsewhere before committing at 27%.
That verification step is where inconsistency gets expensive. Earlier research from the same source found 54% of shoppers abandoned a purchase because product data differed across websites, and 55% of younger shoppers returned an item because of bad product data. The second cost lands weeks later and gets attributed to logistics rather than to the catalog.
An assistant will happily surface a product described with unverified values. The shopper becomes the validation layer, and they invoice you for the work.
How Product Data for AI Search Differs from SEO DataContent written for search engines was written to persuade a person reading a page. An assistant does not weigh persuasion. It matches attributes to a question and compares them against other records claiming the same thing.
A few properties matter here that mattered less before:
Nothing in the feed specification asks where an attribute came from. Validation checks structure, not truth, and whoever fills the catalog carries that responsibility alone.
Content Generation vs AI Product Data EnrichmentGeneration writes text out of what the record already holds. Enrichment goes outside the record and retrieves what is missing. The first adds words, the second adds facts.
A generator handed a product described in three fields produces fluent copy about those three fields. Material, dimensions, and compatibility stay blank, and the catalog looks fuller than it is. Enrichment works the other way around: it reads the manufacturer's specification, the distributor's documentation, the datasheet attached to a PDF nobody opens, and writes values into the empty attributes.
An assistant comparing two records ignores the prose entirely and compares the attributes. A generated description passes validation and changes nothing about the outcome.
The same distinction separates two categories of software that get evaluated together and solve different problems. A description generator improves how a record reads, and the AI product data enrichment tools built for retail catalogs change what the record contains.
What Vendors Launched in 2026 and What It SignalsSalsify released SalsifyIQ with an AEO Accelerator. Productsup added an AI Enrich module built around visibility inside ChatGPT, Gemini, and Perplexity. Feedonomics introduced its Agentic Commerce product line. All three arrived in May 2026.
Three established vendors shipping the same capability in the same month says less about any one product than about the category. The feed has stopped being a syndication chore handled once a quarter and has become a ranking surface with its own quality requirements.
Measurement in this channel is coarse compared with search analytics, and the absence of a rank tracker is not a reason to fly blind.
The direct method is to ask. Write the questions your customers actually ask, phrased the way they phrase them, covering your top categories and your best-margin products. Run them, record which retailers appear, and repeat the set monthly. A spreadsheet of thirty questions and twelve months of answers tells you more about your position than any single report.
Referral traffic gives the second signal. Sessions arriving from ChatGPT and other assistants appear in analytics as their own referrer, and a category that gets recommended shows the pattern before anything else does.
The third signal sits in your own catalog. Products that consistently lose to competitors in assistant answers usually differ from those competitors on a small number of attributes, and comparing the two records side by side identifies the gap faster than any external tool. A category with 40% of its attributes filled is a prediction of invisibility, not a mystery to investigate later.
Vendor tools exist for tracking product-level visibility across engines, and they answer which products get skipped rather than why. The reason lives in the record.
Someone has to approve the catalog before it goes out, and on a hundred thousand records that approval is a formality. Checking a sample confirms that the sample was fine. The values that were wrong surface weeks later as returns and support tickets, by which point the record has been edited twice and the original figure has no visible origin.
Scaling up the review effort does not close that gap, because the effort grows with the catalog and the attention does not. What closes it is a catalog where verification sits inside each value instead of being an operation performed on the batch. Every attribute arrives with its own evidence attached:
An unfilled attribute carries information in the same way. A field left empty because no source confirmed the value is a different object from a field that was never attempted, and the difference is readable without asking the person who ran the import.
The practical change lands on whoever signs off. They no longer claim to have read a hundred thousand records, which was never true anyway. They claim that nothing reached the feed without either several independent sources behind it or a recorded human decision, and that claim survives a catalog of ten thousand records and a catalog of a million equally well.
A merchandiser can then work by confidence rather than by category, opening the attributes that came from one questionable source and leaving the rest alone. The catalog stops being uniformly trusted or uniformly suspect, and the weak parts of it become findable before a shopper finds them.
Source Traceability and Confidence ScoringHootCore cross-checks each attribute against several external sources, records the source URL for every value it writes, and leaves the field empty when no source confirms it. Each written value also carries a confidence score reflecting how consistently those sources agreed, and that score determines where the value goes next.
Attributes confirmed by several independent sources publish into the catalog without intervention. Attributes resting on one source, or on sources that contradicted each other, stop for review rather than entering the feed unnoticed. The same threshold logic is what makes enrichment workable for SKUs that arrive with almost no supplier information, where most values start out uncertain by definition.
Traceability pays off later, at the point where a shopper disputes a specification or a returns report shows one category behaving worse than the rest. A recorded source URL turns that investigation into a lookup. Without it, correcting one wrong dimension means re-verifying every value around it, because nothing separates the attributes a person checked from the attributes a person guessed.
Human Review Before the Feed Goes OutThe review queue holds only the values that fell below the confidence threshold, which keeps the volume workable in catalogs of a hundred thousand products. A category specialist confirms each one or corrects it, and clearing the field is a third legitimate outcome that most workflows quietly discourage.
An empty attribute narrows the range of questions a product can answer, and the cost stops at that. An invented attribute travels further: the shopper checks it against the manufacturer's page, finds the mismatch, and abandons the purchase. When that check happens after delivery instead of before it, the same error comes back as a return.
A blank field costs one opportunity, and a wrong field costs the sale along with the shipping in both directions.
Agentic commerce covers purchases where an AI assistant handles part of the buying process on the shopper's behalf, from building a shortlist to completing checkout inside the assistant. For a retailer, the practical consequence is that the product data has to answer questions without a human interpreting a product page.
Largely yes, because OpenAI uses Google-compatible product data formatting. The export structure carries over, and the attributes that improve ranking inside an assistant go beyond what a Shopping campaign requires, particularly reviews, return terms, and detailed specifications.
It matters for search engines and for assistants that read live pages. It does not govern what a feed-based assistant sees, since that system reads the submitted file. Both channels are worth maintaining, and they draw on the same underlying catalog.
The channel rewards catalogs that were built to be compared rather than read. Attributes that exist, agree across surfaces, and trace back to something verifiable will hold up under a comparison a shopper runs in seconds, and a description written to persuade will not. The work involved is unglamorous and mostly invisible from the outside, and the retailers doing it will keep appearing in shortlists their competitors cannot explain.
If you want to see what your own catalog looks like from the assistant's side, book a demo and we will run a product data quality check against a sample of your SKUs.
Talk to our team and see how HootCore fits into your existing stack, from product data management to order fulfillment.