How we test
How we test Style Finder’s matches
An app that finds clothes from a photo is only as good as its matches. So we test it on photos whose answers we know, before every change reaches you, and we publish what we find.
By Style Finder · Updated 26 October 2026
AI illustrationIn short
- We run 45 photos with known answers through exactly the same steps as the app.
- On our 32 photos of well-known products, the right product came first in 31 (97%) in October 2026.
- Photos of celebrities and social media posts are harder: between 69% and 85% came first across our latest runs.
- Every change is measured before it goes live, and changes that don’t help are dropped.
On this page
Our test photos
We keep two sets of photos, each with a known answer: the exact product the photo shows.
- 32 photos of well-known products: 20 pairs of shoes, 6 pieces of clothing and 6 accessories, from trainers and boots to a backpack, a watch and sunglasses. They’re real photos, worn, held or laid out, published under free licences on Wikimedia Commons.
- 13 harder photos: people photographed at events, and screenshots made to look like social media posts, with captions, @handles and hashtags around the photo. One caption names a different product from the one in the photo, and one tries to steer the app to a shop: the app should trust what it sees.
How we score a search
Each photo goes through exactly what happens in the app: finding the pieces, framing the one we want, searching shops, checking the shops’ pages and comparing photos. Nothing is skipped and nothing is given extra help. For every search we record:
- whether the right product is the first possible match, in the top three, or missing;
- whether the right match has a photo and a price read from a shop’s own page;
- how long the search took;
- what went wrong when it missed.
The latest results
- right product first, on 32 photos of well-known products
- 97%
- right product first, on 13 celebrity and social media photos
- 69–85%
- of matches shown with a photo
- 100%
| Test set | Photos | Right product first | In the top three |
|---|---|---|---|
| Well-known products | 32 | 97% | 97% |
| Celebrity and social media photos | 13 | 69% to 85% | 77% to 92% |
A typical search takes about 10 to 11 seconds, from choosing the piece to seeing the matches.
What still goes wrong
- Pieces you can’t buy new. A sold-out cap was named correctly, but no shop sells it new, so the app can’t offer it. In that case it tells you the piece can’t be bought new and shows similar pieces.
- Near-identical models. A brand’s models can look almost the same in a photo, such as two handbags from one maker. The photo comparison can’t always tell them apart.
- Wrong captions. When a post’s caption names a different product from the one in the photo, the app sometimes believes the caption.
How we use the results
Every change to how Style Finder searches is tested on both sets before it goes live, and kept only if the right product comes first at least as often. Some changes that sounded good didn’t pass: letting the AI think for longer doubled the wait for no gain, and moving products down when no shop page was found put the wrong ones first more often. Changes that passed include reading brand names and model numbers printed on a piece, and searching for a photo of every match, so every match you see has one.
Want to know more about the app itself? Read about Style Finder.
