Why does an AI shopping agent return wrong products for detailed queries?

An AI shopping agent can return wrong products when the catalog lacks structured evidence for constraints such as care instructions, age suitability, materials, or safety. Without that evidence, the agent must either return nothing or guess from product-description text, and text guesses can produce unsafe or incorrect recommendations.

How should I test a conversational commerce vendor on my product catalog?

Test the conversational commerce vendor on a real catalog export that includes long-tail products, not a cleaned sample. Use twenty unannounced queries with three or four stacked constraints, score recall separately from correctness, and require the vendor to identify which catalog attribute resolved each constraint.

What product data does an AI shopping agent need?

An AI shopping agent needs structured and normalized attributes that provide evidence for shopper constraints. Common commerce fields such as title, price, GTIN, brand, image, and category are not enough when queries depend on material, care instructions, age range, weight limits, non-toxic finishes, or detachable-part details.

Why should recall and correctness be scored separately for AI product recommendations?

Recall measures whether the system returns products, while correctness measures whether the returned products actually satisfy the query. A confident but incorrect recommendation is worse than no recommendation, especially when the query includes safety constraints such as toddler suitability.

How does product feed enrichment improve conversational commerce?

Product feed enrichment fills missing attributes, normalizes inconsistent attribute names and values, and derives fields that suppliers or commerce channels did not populate. The resulting catalog can support queries that previously lacked enough structured evidence, while the enriched data can also support advertising and marketplace feeds.

Your AI Shopping Agent Can Only Answer What Your Catalog Already Knows

Every conversational commerce demo follows the same script. A shopper types something a search box could never handle. The agent thinks for a beat, asks one clarifying question, returns three products that fit perfectly. The room nods.

Then you go home, ask the same question of your own catalog, and nothing comes back. Or worse, something comes back and it is wrong in a way nobody in the room would have caught.

This wasn’t a vaporware demo. It was running on a catalog that had been prepared for it.

Take one query apart

"Machine-washable, toddler-safe, under $80."

Only one of those constraints is likely to exist in your feed as a structured field. Price is always there, because price is what commerce systems were built around. Machine-washable lives in the care instructions, which means it lives in a supplier PDF, or in the third paragraph of a description, or nowhere. Toddler-safe is not an attribute at all. It is an inference sitting on top of several attributes you probably also lack: no detachable parts under a certain size, non-toxic finish, weight limit, an age range someone actually populated instead of leaving as "all ages."

An agent can’t reason its way to an answer it has no evidence for. When the evidence is missing it can decline, which shoppers read as a broken product. Or it can guess from description text, which shoppers read as a working product right up until a parent buys a choking hazard because the model saw the word "soft."

The second failure is the expensive one, and it is invisible in a demo.

Nor is this a problem confined to thin catalogs. Adobe scored retail product pages at 66% machine-readable on average, against 75% for homepages. A third of the content on the page where the purchase decision happens is invisible to the model reading it.

What the feeds actually look like

I spend my working days on AI data enrichment for advertising and marketplace product feeds at Feedonomics, which means I see catalogs in the state they arrive in rather than the state they demo in.

The pattern holds across verticals. A handful of fields are populated well because some channel required them: title, price, GTIN, brand, primary image, one category path. Everything past that thins out fast. Material appears on half the SKUs and is spelled six ways across suppliers. Care instructions are prose when they exist. Age and safety attributes are sparsest of all, because until recently nothing downstream punished you for leaving them blank.

This is not just my sample. Baymard's benchmark found that 38% of ecommerce sites cannot filter on attributes they already display in their own product listings, and their researchers trace it to product data that was never normalized into common attribute names. The filters were missing because the data underneath them was.

Then enrichment runs, fields get derived and normalized, and the catalog starts answering questions it could not answer the week before.

Change what the demo has to prove

Run it on your catalog. A real export, long tail included, not a cleaned sample.

Bring twenty queries you did not share in advance. Three or four stacked constraints, at least one a soft inference like "good for a small apartment."

Score recall and correctness separately. A system that returns confident wrong answers scores worse than one that returns nothing, even though it demos better.

Ask which attribute resolved each constraint. If the vendor cannot say, the answer came from a text guess. Fine for "cozy sweater," unacceptable for "toddler-safe." Then ask what share of your catalog would hit that path today.

The thing under the chat window

Conversational commerce is not a fake category. The agents shipping right now are real engineering, and the good ones move real numbers.

But the interface is the last mile, and the last mile is not where the risk lives. Two retailers can license the identical agent from the identical vendor and get completely different results, because one has a catalog that can answer questions and the other has a catalog that can only list prices.

The demo worth trusting shows you attribute coverage first and the conversation second. If a vendor will not show you the layer underneath, it is because that layer is yours to build, and they would rather you find out after signing.

That layer is buildable, and it is a smaller project than most teams assume. Feedonomics Data Enrichment fills missing attributes, normalizes the ones that arrived spelled six different ways, and derives the fields nobody populated because no channel used to ask for them. The catalog gets deeper, and it pays off in ads and marketplaces even if you never implement agentic chat.

Disclosure: I work on that product, so weigh this accordingly.

David works on AI data enrichment for advertising and marketplace product feeds at Feedonomics.

Sources

AI traffic grows but retail sites lag in AI search visibility, Adobe, 2026
Have Filters for All Displayed List Item Info, Baymard Institute

Further reading (from my employer)

Agentic catalog exports, Feedonomics, June 2026, on the attribute, taxonomy, and structured content gaps that determine whether an agent surfaces a product at all
Data enrichment for scalable, high-quality product content, Feedonomics, January 2026