How AI Describes Ecommerce Products—and Gets Them Wrong
How AI Describes Ecommerce Products—and Gets Them Wrong
See why AI can describe ecommerce products incorrectly when PDPs, feeds, retailer listings, reviews and variants disagree—and how to diagnose the source.
Haritha Kadapa
How Does AI Describe Your Products—and When Is It Wrong?
AI does not necessarily describe a product the way the brand describes it.
A shopping assistant may combine structured merchant data, product pages, retailer listings, public reviews and other sources before turning that information into a shorter answer for the shopper. OpenAI, for example, says ChatGPT can generate simplified product titles and descriptions from third-party product information, and its review summaries can draw from reviews on public websites. Google says its Shopping Graph receives product information from brands, retailers and other content providers and can use that information in generative-AI shopping experiences.
That creates an ecommerce problem that is easy to miss:
The product your company sells and the product an AI system describes are not always the same information object.
Before asking how to improve AI visibility, commerce teams need to know whether the product being represented is accurate in the first place.
One SKU can have several versions of the truth
Imagine a footwear brand selling a trail-running shoe.
Its current PDP says the shoe has a waterproof membrane, comes in regular and wide fits, and weighs 280 grams.
The merchant feed contains the same product, but the description has not been updated since the previous season. It calls the upper “water resistant” and does not include the wide-fit option.
A large retailer carries the product too. Its listing has the correct price and images, but copied an older specification showing a 300-gram weight.
Customer reviews repeatedly say the shoe “runs narrow,” even though a wide variant exists.
Now a shopper asks:
“Is the Trail X2 a good waterproof running shoe for wide feet?”
An AI assistant has several pieces of evidence available, but they do not say exactly the same thing.
Which one becomes the answer?
There is no universal rule.
OpenAI explicitly says its shopping descriptions may be generated from third-party information rather than simply reproducing the merchant's own wording. Google similarly says it uses information from multiple product-data sources and may not use the data supplied by the merchant for every product listing.
So the ecommerce team should not assume that correcting the PDP automatically corrects the product everywhere AI may encounter it.
Product information can drift before the AI generates anything
Some apparent “AI errors” begin much earlier.
Google Merchant Center documentation is unusually useful here because it shows how ordinary commerce infrastructure already encounters conflicting versions of product truth. Google checks landing-page information against product feeds and structured data, and it explicitly documents price and availability mismatches between those surfaces. Its automatic-update systems exist partly to correct cases where submitted product data has fallen out of sync with the landing page.
Availability can drift too. Google tells merchants that product availability needs to remain consistent across the landing page, checkout, structured data and product data source.
Those examples are not evidence about how every AI assistant ranks products. They show something more basic: a single ecommerce company can already expose contradictory machine-readable facts about the same SKU.
Once generative systems begin retrieving, comparing or summarizing those facts, the inconsistency becomes visible to the shopper in a new way.
Price is easy to notice.
Other attributes are harder.
A product may be described as leather in one listing and synthetic leather in another. An older retailer page may contain last year's battery capacity. One color variant may inherit the dimensions of another. A marketplace description may omit a compatibility limitation that the brand added months ago.
Those differences rarely appear on the executive dashboard.
They can materially change whether the product appears suitable for a shopper's question.
Sometimes nothing is stale—the AI interpretation is still wrong
Source consistency does not remove every error.
Generative systems can summarize, infer and compress information rather than simply copy database fields.
Microsoft warns users directly that Copilot Shopping can occasionally make mistakes and advises them to verify important product details, pricing and purchase terms with the retailer.
OpenAI makes a similar distinction with generated shopping labels. A label such as “Budget-friendly” is generated from information available to the model and is not a verified guarantee that the product has the lowest price.
This distinction matters.
Consider:
“This shoe has a wide-fit option.”
That is a product fact.
“This shoe is good for wide feet.”
That is an interpretation.
The first can often be checked against a catalog, PDP or manufacturer specification.
The second may depend on sizing information, reviews, comparison evidence and the shopper's circumstances.
A brand can therefore be represented inaccurately even when every individual source contains a piece of defensible information. The error can emerge from how those pieces are combined.
The broader factuality problem is well established in language-model research. Ecommerce-specific research has also studied hallucinations in LLM-enriched product listings because generated product attributes can become unfaithful to the original source information. That research does not directly measure ChatGPT or Gemini shopping results, but it reinforces why product claims should be checked at the attribute level rather than assumed correct because the output sounds plausible.
The most useful audit starts with claims, not mentions
A brand-level visibility report might tell you that the Trail X2 appeared in 37% of tracked answers.
That still does not tell you whether those answers described the shoe correctly.
For product intelligence, the more useful unit is the claim.
Take the product statements appearing in AI answers and compare them with the best available first-party product truth.
For the trail shoe, the audit might look like this:
AI-stated claim | Authoritative product truth | Result |
Waterproof membrane | Confirmed on current PDP/spec | Supported |
Available in wide fit | Confirmed for selected variants | Supported |
300 g weight | Current specification says 280 g | Outdated / conflicting |
Good for wide feet | Wide variant exists, but reviews are mixed | Interpretation |
$129 current price | Brand site currently shows $139 | Stale or merchant-specific |
The objective is not to label every interpretation a hallucination.
It is to separate four very different situations:
Supported: the AI statement matches authoritative product evidence.
Conflicting: credible sources disagree.
Stale: the statement once may have been correct but no longer is.
Unsupported: the statement cannot be traced to adequate evidence.
That classification immediately makes the problem more actionable.
If the AI is reading an outdated retailer listing, the fix is different from a missing feed attribute.
If the product feed and PDP disagree, the commerce-data team has work to do.
If the answer is an interpretation synthesized from reviews, changing schema may accomplish nothing.
And if all available product information is correct but the assistant still invents a specification, the team needs monitoring and re-testing rather than another content rewrite.
Variants deserve their own truth check
Variant-heavy catalogs make this harder.
OpenAI's current commerce product specification treats variants as distinct objects with stable identifiers and fields for options such as color or size, along with variant-level price and availability.
Google likewise recommends explicit variant identification and distinguishes parent product information from variant-specific titles and attributes.
This matters because the shopper rarely asks about an abstract parent SKU.
They ask for:
“the black one in size 10”
or:
“the 256 GB version”
or:
“the fragrance-free formula.”
An answer can be technically correct about the parent product and still be wrong about the purchasable configuration the shopper actually means.
For large ecommerce catalogs, product-description accuracy eventually has to become SKU- and variant-level, not merely brand-level.
What should an ecommerce team measure?
I would resist turning this immediately into one synthetic “AI product accuracy score.”
Start with observable diagnostics.
For a defined set of commercially important SKUs and shopping questions, track the proportion of factual product claims that can be verified, the attributes where different sources disagree, how often stale values appear, and whether the same SKU is described differently across AI engines or markets.
The denominator matters.
If you test only five hero products, do not report the result as catalog accuracy.
If you test price, availability and size but not performance claims, say so.
And if the answer involves interpretation—“comfortable,” “premium,” “best for sensitive skin”—keep that separate from directly verifiable facts.
This is where the difference between product truth and product perception becomes useful.
Product truth asks:
What is this product actually specified to be?
Product perception asks:
What does the evidence ecosystem lead AI to say about it?
Both matter, but they require different fixes.
The practical goal is not to control every AI answer
Brands cannot guarantee that every generative system will describe every SKU exactly as their internal catalog does.
They can make the evidence less contradictory.
That means keeping PDPs, feeds and structured product information aligned; maintaining clear variant identities; correcting outdated retailer content where possible; monitoring how important product claims appear in AI answers; and tracing incorrect claims back to the source or inference that seems to be driving them.
Google's own commerce infrastructure already treats product-data consistency as an operational requirement. OpenAI and Microsoft both tell shoppers that generated shopping information should still be verified.
For commerce teams, the logical next step is to treat AI product representation the same way they would treat any other customer-facing product surface: measure what it says, compare it with product truth, and fix the part of the evidence chain that is actually wrong.
Gravton's current public platform positioning follows that SKU-level logic: its Shopping Intelligence focuses on which SKU wins and why, while its catalog workflow connects shopper questions, product attributes, feeds, structured data and channel consistency.
Before optimizing how often AI recommends a product, make sure the product AI is describing is the one you actually sell.
Win your AI search demand universe
Understand how to turn AI narrative insights into prioritized content that improves how AI platforms describe your brand.
VISIBILITY & CONTENT STRATEGY