6 min read
In brief
Pinterest image search used AI to guess the searches behind each photo, and visits rose 18 percent.
A photo may show a green dress, but the person searching could want an office outfit, a particular style or something to wear to an event. Search engines can connect the picture to those needs only when enough language and structure exist around it. That was the gap Pinterest image search needed to solve.
Pinterest is built around visual discovery, and many of its images are useful for the ideas, occasions and projects they suggest as well as the objects they show. A literal label can identify the dress while missing every reason someone might want it.
Pinterest researchers asked whether AI could supply that missing context at scale. In a February 2026 paper, they described a system that generated likely searches for each image, used those searches to create collection pages and connected the pages through internal links. The reported 18 percent lift came from a four-week test of this complete setup against Pinterest’s previous system.
How Pinterest image search learned intent #
The previous system matched images to searches through visual similarity. Its replacement generated five to seven likely searches for every image. Pinterest trained the model on query-image pairs with recorded impressions, clicks or top-ten rankings, then added synthetic examples for intentions underrepresented in the historical data.
The generated searches fell into three groups. Thirty percent described the subject, such as “soft green knit dress”. Another 30 percent added style or detail, such as “sage green monochrome look”. The remaining 40 percent described a use case, such as “modern office outfits for women”. Pinterest gave this final group the largest share because its researchers found that use-case searches produced disproportionate incremental traffic.
The Pinterest image search system gathered relevant images onto collection pages, then linked each page to its images and to related collections. This structure showed search engines how one picture fitted into a broader topic and gave them a route through the library.
Putting it to the test #
Pinterest tested the complete package, running three versions side by side for four weeks. The first had no annotation system or internal links. The control used queries and metadata-based links from the previous system. The treatment used the model-generated queries and their corresponding links.
The treatment scored 1.18, compared with 1.00 for the previous system, an 18 percent gain. The version without annotations or links scored 0.82. These are index scores, not raw visit counts.
The difference was much larger within traffic Pinterest classified as generative search, meaning visits from AI-powered search. The three groups scored 0.04, 1.0 and 9.2. That covers one traffic category, so total visits did not increase ninefold. The distinction matters because AI systems can mention, cite and recommend brands in different ways, even when they draw on the same underlying pages.
The collection pages also needed the right images. Pinterest compared two models for selecting them. PinCLIP performed better in an offline assessment of query fit. SearchSAGE, which was trained with engagement and graph signals, produced slightly higher sign-up and click-through rates in a separate month-long online test. The paper links the difference to SearchSAGE’s training, although the experiment did not isolate the cause. In this comparison, the stronger offline score and the stronger live engagement measures pointed to different models.
Beyond these tests, Pinterest reports that the system powers tens of millions of pages and runs at roughly 1/94 of the cost of commercial vision-language model APIs. It also reports a 20 percent increase in organic search traffic. A later table shows total sessions rising by 1.20 percent, and the paper does not reconcile the two figures. They may cover different deployments, periods or denominators, so treat them as separate measures.
The gap it filled #
For Pinterest image search, the generated queries supplied the missing language. The collection pages gave each query a destination, and the internal links tied individual images to broader topics. Use-case searches carried the largest share of the mix. They describe the decision behind a search, where a literal label describes only what appears in the picture. The experiment tested these elements as a package. It shows that the new combination outperformed Pinterest’s previous system, but it does not reveal how much of the gain came from the generated queries, the collection pages or the links.
Try it yourself, but smaller #
A smaller company can test the same Pinterest image search mechanism within one product category. The question to answer is whether pages organised around customer intent attract more qualified search traffic than the existing catalogue structure.
The primary measure should be additional qualified organic sessions or contribution margin from search visitors. Referrals from AI search systems, indexation and collection click-through can help explain the result, but they should not replace the business outcome.
The third group separates the effect of creating the pages from the added effect of linking them. Compare it with the unchanged group to estimate what the pages contributed, then with the fully linked group to see whether the links added more.
Source note. Results are reported by Pinterest in a company-authored paper posted to arXiv on 3 February 2026 and checked on 20 September 2026. The production tests have not been independently replicated. The paper does not disclose traffic allocation, sample counts or confidence intervals for the four-week test.