Catch product data errors before returns rise: check 30 items with AI
Start small: pick the attribute behind most returns, have AI compare 30 products, and let a person make the fixes.
Key points
- Define the attribute behind most returns
- Have AI list mismatches for 30 products
- Confirm materials with suppliers
- Log returns and queries before and after
Contents
Define the attribute behind most returns, such as width
A&A perspective
Start by picking one attribute that drives returns and writing down what it means in a single line. An attribute is an item listed on a product page, such as dimensions, color or material.
A&A perspective
Line up a few months of return and inquiry reasons, such as "wrong size" or "wrong color," and start with the most common one. If you pick width, settle a rule like "for extendable dining tables, list the width before extension."
From the sources
Wayfair, a US online retailer of furniture and home goods, illustrates the cost of wrong dimensions with a hypothetical: a sofa listed at six feet arrives measuring closer to seven, and the customer debates whether to return it and start the search over.
From the sources
The company writes that passing supplier dimensions on to customers gets complicated in practice. A single product can carry many measurements in different units, provided by more than one supplier.
From the sources
Wayfair may want a consistent reading of a sofa's height, width and depth, but an individual supplier may not follow it. A value in a listing can also change over time without being checked against the rest of the listing.
From the sources
Wayfair writes that it is a mistake to think a product's width is always self-apparent. For a dining table with an extension leaf or a sectional sofa, even people can struggle to agree on the primary dimensions.
From the sources
So Wayfair put dimension standards defined by its merchandising team, along with case-specific instructions, directly into the AI's prompt.
From the sources
OpenAI, the company behind ChatGPT, has also published a Wayfair customer story. It says Wayfair's system first uses internal definitions and web information to build a guide explaining what each label means. The models then use that guide to review product information.
A&A perspective
Both systems settle the meaning before the AI looks at anything. Without that, a person cannot tell whether a mismatch the AI flags is an error or just a different way of measuring.
Have AI list mismatches for 30 products
A&A perspective
Next, take the 30 products with the most returns for your chosen attribute and have AI list mismatches between each product page and your product list. By product list we mean the sheet where your company records dimensions, materials and other specs.
A&A perspective
Do not let the AI fix anything; ask it only for each mismatch and the evidence behind it. Give it text such as the product name and description, plus any images that show the dimensions.
From the sources
Wayfair's system gives the AI two kinds of input: text such as product names, descriptions and feature bullets, and images that show the product's measurements. The model first extracts dimensions from the text and the images separately.
From the sources
It provides a value only when the two sources do not conflict. That value is compared with the supplier's dimension, and cases where the model's confidence is low or the difference is small are dropped.
From the sources
Wayfair writes that small differences are dropped to avoid exhausting suppliers with requests for tiny changes.
From the sources
It also writes that flooding suppliers with false positives risks losing their trust and engagement. For that reason it prioritizes precision over recall.
A&A perspective
When you test 30 products, look at whether each flagged mismatch is right rather than how many there are. Have the AI attach its evidence to every mismatch: which sentence, or which image.
A&A perspective
With the evidence attached, the person checking only needs to look at that spot. If most flags turn out wrong, revisit the instructions to the AI or the one-line definition you wrote first.
A&A perspective
Your product list can be the outdated side, so finding a mismatch does not yet tell you which side is right. That is decided in the next step, after a person checks.
| Kind of mismatch | How Wayfair handles it | In a 30-product test |
|---|---|---|
| Small dimension difference | Dropped, to avoid exhausting suppliers with tiny requests | A person looks and decides whether to fix it |
| Low-confidence or high-risk change | Not applied until the supplier confirms | Confirm with the supplier or the item, then fix |
| High-confidence change | Updated directly, and the supplier is notified | A person checks until you can tell confidence apart |
Confirm materials with suppliers when photos can't tell
A&A perspective
A person checks every mismatch the AI lists before anything is fixed. For things photos cannot settle, such as material, ask the supplier or look at the actual item yourself.
A&A perspective
A photo may not show whether a board is solid wood or printed wood grain, yet AI can still state a material from an image. Automate & Augment (A&A), which writes this article, thinks a page should not be rewritten on that basis alone.
From the sources
Wayfair writes that when its process flags a problem, it asks the supplier to review the issue and change the product data. It would like its suggested corrections to become reliable enough to fix values without the supplier's validation.
A&A perspective
Read the other way round, even the dimension system still routes fixes through the supplier today.
From the sources
OpenAI's customer story says Wayfair associates inspect product samples to check the output. Suppliers also help validate changes.
From the sources
When confidence is high, the system updates product content directly and notifies the supplier. Low-confidence or high-risk changes require supplier confirmation before taking effect.
A&A perspective
At the 30-product stage, you have no yardstick yet for telling high confidence from low. So a person checks everything, and materials are confirmed with the supplier or the actual item before any fix.
A&A perspective
Until the supplier answers, leave the page as it is. Keep a one-line note per product of what you asked and what came back, so the next mismatch on that product shows what was already checked.
Log returns and queries before and after fixes
A&A perspective
After fixing, count returns and inquiries about that attribute over a period as long as the one before. Take the before figures from the same length of time up to the day of the fix.
A&A perspective
Count orders, returns, inquiries and time spent. First, orders for the 30 products, the base you compare returns against. Next, returns that cite the attribute, such as "wrong size."
A&A perspective
Then inquiries about the attribute, such as "how wide is it?" Last, the time it took to check and fix, a guide to the effort of keeping this going.
From the sources
According to the customer story, before working with OpenAI, Wayfair's tagging relied on suppliers and customers telling it that something looked wrong. Manual effort could not keep up with the volume.
From the sources
Wayfair's associate director of catalog merchandising says better data quality builds more trust with customers. It helps shoppers make the right buying decisions, the director says, directly reducing costly issues such as returns from misrepresented products.
A&A perspective
Counting returns and inquiries moves your own company a step past waiting to be told. But 30 products over a few weeks produce few returns, and a before-and-after gap can appear by chance.
A&A perspective
So before asking whether returns fell, check whether mismatches were found and fixed. If they were, expand to the next 30 products or the next attribute.
A&A perspective
The time spent tells you how to expand. Knowing how many minutes one check takes lets you size the next batch to what your people can handle.
Summary
A&A perspective
Checking product data does not have to be a big job of handing every product to AI. Pick the attribute behind most returns and define it, have AI list mismatches for 30 products, and confirm materials that photos cannot settle with suppliers before fixing. Counting returns and inquiries before and after leaves your company the evidence to decide whether to continue.
Product data errors can be tested on one attribute and 30 products, with AI listing mismatches and a person fixing them. The next decision is which attribute to start with.
FAQ
Can I apply the mismatches AI finds as they are?
At first, a person checks before fixing. In the customer story, Wayfair updates content directly and notifies the supplier when confidence is high, and waits for supplier confirmation when it is low. Until your company can tell high confidence from low, having a person review everything is safer.
Why 30 products?
The number 30 has no source. It is a working figure this article sets as an amount a person can check one by one against the evidence. If you carry only a few dozen products, reviewing all of them by hand without AI may be faster.
What if we make our own products and have no supplier?
Instead of asking a supplier, measure the actual item or ask the person who made it. Not deciding material from photos, and counting before and after the fix, stay the same.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- How We Validate Product Dimensions at Scale With Multimodal GenAI
Wayfair · 2026-01-15
Accessed 2026-10-28 - Wayfair boosts catalog accuracy and support speed with OpenAI
OpenAI · 2026-09-25
Accessed 2026-10-28