Wayfair
Wayfair uses Gemini to catch wrong product dimensions across its catalog
Wayfair uses a multimodal AI model to compare supplier-provided product dimensions with the product's text and images and flag conflicts for suppliers to fix. Wayfair reports over 85% precision and more than 70% recall, far above its earlier machine learning methods.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
over 85%
Precision of conflict detection on product dimensions1
more than 70%
Recall of conflict detection on product dimensions1
less than 50% precision and 7% recall
Precision and recall of earlier machine-learning methods1
over 100 thousand
Products in the pilot where suppliers reviewed the flagged cases1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Wrong product dimensions frustrate customers and can lead to returns. Wayfair has a catalog of over 40 million products from over 20 thousand global suppliers, with thousands of new items added daily. Suppliers give measurements in different units and conventions, and listings change over time, so text, images and measurements can contradict each other.
What they built
Wayfair built a validation model on Google's Gemini 2.5 Flash. It reads a product's text (names, descriptions, feature bullets) and its dimension images, then checks whether they support or conflict with the dimensions the supplier provided. Conflicts go to the supplier to review and fix.
How they did it
The model extracts dimensions from text and images separately, with a confidence score for each. It then reconciles them into one suggested value and compares it with the supplier's value. Low-confidence cases and tiny differences are dropped. Prompts use step-by-step instructions, a few examples and Wayfair's merchandising definitions. It runs through the Gemini Batch API as the best balance of reasoning strength, cost and turnaround.
Results
Wayfair says the model consistently achieved over 85% precision and more than 70% recall in detecting conflicts. Its earlier machine-learning methods reached less than 50% precision and 7% recall. After a pilot on over 100 thousand products, where suppliers reviewed flagged cases, Wayfair judged the model ready for full production release.
What to copy
Decide in advance whether false alarms or missed errors cost more. Involve the business owner in the definitions. Test on a pilot where the affected parties confirm the flags.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Precision of conflict detection on product dimensions | over 85%1 | |
| Recall of conflict detection on product dimensions | more than 70%1 | |
| Precision and recall of earlier machine-learning methods | less than 50% precision and 7% recall1 | |
| Products in the pilot where suppliers reviewed the flagged cases | over 100 thousand1 | |
| Human-labeled records in the first evaluation set | approximately 5,0001 |
Lessons for your project
Favour precision over recall when humans must act on each flag. Wayfair asks suppliers to fix flagged data, and false alarms would cost their trust.
Add tolerance rules for small differences so suppliers are only asked to fix meaningful conflicts.
Write clear definitions into the prompt with the business team. Wayfair's merchandising team defines what counts as height, width or depth for odd shapes.
Validate in phases: a labelled test set, internal review of flagged items, then a pilot where the people affected check results.
Caveats
- All figures are Wayfair's own, from one company tech blog post. No outside party confirms them.
- Precision and recall are reported as thresholds the model consistently achieved, not as a single measured value, and the post does not state which phase produced each.
- The post gives no go-live date. It says the results made Wayfair confident the model was ready for full production release and that new products now enter with accurate dimensions. The deployed date shown is the post date.
- Suppliers, not Wayfair, make the corrections. The post does not give how many errors were fixed, returns avoided or money saved.
- The cost claim (cheaper than hiring people to review the data) is stated without a figure. Wayfair also lists cost per use as a continuing concern.
Sources
- How We Validate Product Dimensions at Scale with Multimodal GenAI
Wayfair Tech Blog · 2026-01-15 · Company blog · aboutwayfair.com
Summarized in our own words from the sources above. Spotted an error? See corrections.