Search that sees you: Pinterest's inclusive search

#sizeandshape; #ethnicity

Type "wedding dress" into most search engines and you'll get an avalanche of results — but how many of them actually look like you? For years, that mismatch between "search results" and "myself" was just background noise online. Pinterest decided to do something about it, and the result is a quietly significant case study in what inclusive product design looks like when it's built into the core of a search engine rather than bolted on as an afterthought.

There are now three separate filters threaded through Pinterest's search experience: skin tone ranges, hair pattern search, and body type ranges.

The three tools

Skin tone ranges is the oldest of the three, dating back to 2018. Search a beauty term — eyeshadow, red lipstick, eyeliner — and you can narrow results to one of four skin tone palettes. It's the feature with the longest track record: Pinterest has rebuilt the underlying model at least once, moving from raw pixel analysis to an embedding-based approach that, by the company's own account, cut the time needed to classify skin tone across billions of images from nearly a week down to under an hour.

Hair pattern search, launched in 2021, lets you refine a broad search like "summer hairstyles" or "glam hair" into one of six categories: protective, coily, curly, wavy, straight, and shaved/bald. It was built using computer vision-powered object detection and, notably, developed with BIPOC creators and stylists rather than designed in isolation and shipped outward.

Body type ranges is the newest addition, tested in late 2023 and rolled out to the US in 2024 before reaching the UK, Ireland, Australia, New Zealand, Germany and France in 2025. It currently only applies to women's fashion and wedding searches, where you can filter into one of four body type ranges. Pinterest has said it plans to extend this to men's fashion and further markets, though that rollout has been gradual rather than sweeping.

The mechanics are deliberately consistent across all three. You search a relevant term, a filter bar appears above the results, you pick a range, and your selection can optionally be remembered on that device for future searches. None of it is automatic or assumed — you have to actively search a relevant term and then opt in to the filter.

How it actually works

Underneath all three sits the same basic idea: machine learning models trained to classify visual attributes — skin tone, hair texture, body shape — across Pinterest's enormous image library (the body type model alone has been applied to over 3.5 billion images). Pinterest calls the team behind this its Inclusive Product team, a deliberately cross-functional group sitting between engineering and the company's inclusion and diversity function. That structural detail matters more than it might seem — it's an admission that "make search more representative" isn't purely a modelling problem or purely a policy problem, but both.

The skin tone work in particular shows the unglamorous reality of getting this right: dealing with shadows and lighting variance, running qualitative reviews with diverse panels of evaluators, building dashboards specifically to monitor content diversity rather than just accuracy. It's the kind of investment that's easy to skip and hard to notice when it's done well.

The case for it

Pinterest's own numbers, while self-reported, are hard to wave away. Internally, nearly 60% of top beauty search terms once contained a skin tone qualifier — "dark skin," "olive skin," "pale skin" — meaning people were already doing this work manually by stuffing it into the query. The feature simply turned an existing workaround into a first-class part of the product. Engagement metrics tell a similar story: Pinterest reported a 70% year-on-year increase in users saving Pins across skin tone ranges after international rollout, and a 66% higher engagement rate per session among people using body type ranges.

There's a neat design principle buried in this: the filter doesn't try to know who you are. It asks. You select the range that's relevant to you, search by search, rather than the platform inferring and locking in an identity on your behalf. That's a meaningfully different (and more respectful) posture than most "personalisation" features, which tend to quietly profile you rather than ask.

Where it gets harder

None of this is finished, and Pinterest is upfront that these are "a work in progress." A few things stand out from a product-design lens:

• Coverage gaps persist. Body type ranges still only works for two verticals — women's fashion and wedding — years after launch. Hair pattern and skin tone are wider but still region-gated, and availability genuinely varies by country, which the help pages flag bluntly rather than glossing over.

• The selection doesn't travel with you. Your chosen range is saved on-device, not to your account. Switch from your laptop to your phone and you're starting over. Convenient for privacy, mildly annoying for consistency — a trade-off Pinterest seems to have made deliberately rather than by accident.

• Someone has to be classified before you can find them. All three filters depend on Pinterest's models tagging the content, not just the searcher. That means real images of real people get sorted into a skin tone, hair pattern, or body type bucket by an algorithm, whether or not the person pictured asked for that. Pinterest's answer is a route for people to object to or request removal of how their image has been categorised — a necessary safety valve, though one that puts the burden of correction on the person rather than the system.

• Self-identification and computer vision sit in tension. A user choosing "coily" as a search filter is an act of self-description. A model labelling someone else's photo as "coily" is a judgement call made about a third party. Running both through the same six-category system is pragmatic, but it's worth being honest that they're not quite the same kind of inclusion.

If you're working on anything that touches representation in search or recommendations, that's the template worth borrowing: ask, don't assume; start narrow and earn the right to expand; and always leave a door open for someone to say "that's not me" — because eventually, for any model classifying people at scale, someone will need to.