Taste Graph is Pinterest's internal map of what its users are interested in, built by attaching interest labels to images saved on the platform and then attaching those labels, weighted by how recently and how often a person engaged, to the people who saved them. It exists because Pinterest holds an unusual record: not a list of friendships, and not a browsing trail gathered on other companies' websites, but a corpus of images people deliberately filed under headings of their own invention. Multiplied across hundreds of billions of saves, those small statements of preference become the asset the company sells to advertisers.

The phrase predates Pinterest. In recommendation engineering it describes any structure mapping people to items by predicted affinity. In advertising it now refers almost always to the Pinterest product: the interest vocabulary, the user profiles built on it, and the automated campaign machinery reading both.

How interests get attached to content

Pinterest Engineering set out the architecture in a post published on 21 September 2017 by Brian Johnson, then head of knowledge engineering. Three components make up the graph. Pins carry scored metadata signals including user activity, interests, category relationships, demographic affinities, graph walks and vector embeddings. Users carry aggregated signals boosted by recency and frequency of engagement. Interests sit in a multi-level hierarchy of thousands of frequently encountered terms, organised by similarity and user activity.

The content pipeline runs in three stages. Text is extracted for every pin inside a PinJoin, the collection of pins sharing one image, drawing on title, description, link text, board name, alt text, page title, meta description, meta keywords and structured Schema.org data. That text is lemmatised and matched against a dictionary, so that car, cars and their possessive variants collapse to one entry. Pinterest's English dictionary held roughly 3 million interests at the time, about half in production use. Interest text is then discarded, processing continuing on numeric identifiers alone for speed and storage.

Scoring selects the best labels. A supervised model picks the top 25 interests per language for each PinJoin, chosen empirically because fewer than 10 produced a sparse signal and more than 25 cost too much to serve. Features include word embeddings, frequency counts across pin, board and link text, normalised TF-IDF scores, category affinities, position within the text, whitelists drawn from Wikipedia titles and vertical glossaries, graph queries and spam blacklists. Pinterest tested support vector machines, decision trees, random forests and feed-forward networks before settling on logistic regression, training it on crowdsourced human judgements of whether a keyword was relevant to a given image.

The published results were specific. Precision improved from 71 percent to 79 percent, coverage averaged around eight interests per pin per language, and dictionaries existed for 32 languages with 23 active in production.

On the user side, the distinguishing mechanic is decay. John Milinovich, the product manager who joined through Pinterest's 2016 acquisition of the mobile ad tech firm URX, described the advance over earlier systems as understanding recency and frequency well enough to drop stale affinities: someone who stopped saving wedding content would stop seeing it.

What an advertiser can actually buy

Interest targeting reaches people based on their interests, tastes and life events, according to Pinterest Business help documentation, which recommends testing broad terms such as living rooms alongside precise ones such as mid-century modern. In the bulk editor the same selection is written as JSON in column AP, pairing a numeric identifier such as 934876475639 with a label such as diy and crafts.

Interests sit alongside keywords, audiences, demographics, placement and Performance+ as one of six targeting controls, and Pinterest's guidance is to combine the first two rather than run either alone.

The automated path removes the choice. Performance+, launched globally on 1 October 2024, cut required campaign inputs by half and handed targeting to the model. Promote a Pin went further in March 2026, using the graph to decide which users are likely to convert with no manual targeting at all: the advertiser supplies an image and a budget.

Origin and evolution

The term entered commercial use through Hunch, founded in 2007 by Chris Dixon, Caterina Fake and Tom Pinckney and launched publicly in June 2009 as a question-and-answer decision engine before pivoting to affinity prediction. eBay acquired the company on 21 November 2011 at a price reported then as around $80 million. Dixon later described the scale in an eBay interview: predicted information on roughly 500 million people, 200 million items and 30 billion connections between them, compressed to 50 numbers per person, which Hunch called a Taste Fingerprint.

Pinterest built its own version on different raw material, publishing interest assignment work from 2014. The advertising launch came on 21 September 2017, when targetable interests rose from roughly 400 across 29 top-level categories to more than 5,000. Pinterest reported early tests showing click-through rate gains above 50 percent and cost-per-click reductions above 20 percent, figures it published itself and which were not independently verified.

The machinery underneath then changed shape. PinSage, published in 2018, applied a graph convolutional network to 3 billion nodes and 18 billion edges, trained on 7.5 billion examples. PinnerSage followed in 2020, representing each user through multiple clustered embeddings rather than a single vector across a base of more than 400 million monthly users. SearchSage added query representations in 2021. The symbolic interest vocabulary survived as the advertiser-facing layer while retrieval beneath it moved to learned vectors.

PinRec, a single generative retrieval model serving search, home feed and related pins, was extended site-wide in the first quarter of 2026, improving search fulfilment by roughly 180 basis points and cutting CPA and CPC by a similar margin. The search context window was stretched to 18 months, moving ranking further from query intent and closer to accumulated taste.

Why it matters for the marketing community

Distribution explains the weight. More than 72 percent of Pinterest impressions occur on search surfaces, against 80 billion-plus monthly searches, so the same profile that orders organic results now orders paid ones. Performance+ accounted for 30 percent of lower-funnel revenue by the first quarter of 2026, on revenue of $1.008 billion and 631 million monthly users.

Interest targeting also needs neither a conversion tag nor customer data. Everything else on the platform assumes infrastructure: catalogues for Top of Search and commerce formats, hashed customer files for third-party audience matching, conversion signals for automated bidding. The graph works from platform behaviour alone.

Limitations and disputes

Accuracy is the first constraint, and Pinterest's own disclosure sets the ceiling. Precision of 79 percent at the content layer means roughly one label in five was wrong when the system launched for advertisers, and no comparable figure has been published since.

Inference is the second. A board is an aspiration as often as a purchase plan, and the graph cannot separate the two. The decay logic acknowledges the problem rather than solving it.

Verification is the third, since performance claims attached to the graph come from the seller. Applying it on top of tvScientific's connected television algorithms produced a stated 27 percent increase in outcomes and 65 percent increase in purchases, with one home furnishings partner reporting 190 percent higher incremental reach and 159 percent higher incremental sales in early testing.

Control is the fourth. As automation absorbs targeting, the taxonomy becomes less auditable: buyers running broad or interest-based campaigns were advised to watch performance through the PinRec transition because the signal underneath had changed.

Then there is who sits in the graph. Pinterest's advertising guidelines prohibit targeting on gender identity, race, ethnicity or religious belief; profiling rules elsewhere bite harder. Under the United States children's privacy regime, persistent identifiers on child-directed inventory cannot be used to build a profile. Pinterest's own teen wellbeing research published in September 2026 documented distinct behaviour among younger users without addressing whether those signals reach the same advertising infrastructure.

Cost closes the list. Retraining the graph and running inference on every session sits behind a planned $4 billion commitment to Amazon Web Services through 2031, agreed while the company was cutting roles and defending a securities class action over advertising disclosures.

Not the same as

An interest graph is the generic category. Taste Graph is one proprietary implementation, owned by one company and not portable.

A social graph maps people to other people and infers interests from those connections second-hand. Pinterest skips the intermediate step, having little social connection data to work with.

An identity or device graph resolves identifiers to a person or household across companies. Taste Graph resolves nothing; it scores affinity for accounts Pinterest already holds.

Keyword targeting is a different signal again: a query states intent at a moment, an interest states affinity over time, and contextual targeting reads the content environment rather than any user profile.

Recent developments

Extension beyond Pinterest properties is the live theme. tvScientific was acquired for $465.1 million, closing on 17 February 2026, putting graph-derived audiences into connected television inventory Pinterest does not own. Four AI products announced on 17 June 2026 included a Model Context Protocol server exposing campaign and keyword data to external agents, with six agency and ad tech firms named as alpha partners.

Commercial pressure arrived alongside it. European advertising growth fell to 12 percent in the second quarter of 2026, 7 percent in constant currency, on group revenue of $1,179.7 million and 640 million monthly users. Audience research continues in parallel, with Pinterest reporting in September 2026 that Gen Z men account for more than 62 percent of male monthly searches, alongside annual consumer trend forecasts and seasonal planning guidance drawn from the same search corpus.

Timeline

  • 2007: Hunch founded by Chris Dixon, Caterina Fake and Tom Pinckney
  • June 2009: Hunch launches publicly as a decision engine, later pivoting to affinity prediction
  • 21 November 2011: eBay acquires Hunch, price reported at around $80 million
  • 2014: Pinterest publishes early work on its interests platform
  • 2016: Pinterest acquires URX, bringing John Milinovich to the interest targeting effort
  • 21 September 2017: Pinterest opens more than 5,000 interests for advertiser targeting, up from roughly 400
  • 2018: PinSage published, applying graph convolutional networks to 3 billion nodes and 18 billion edges
  • 2020: PinnerSage introduces multi-embedding user representations
  • 2021: SearchSage adds query representations
  • 1 October 2024: Performance+ launches globally
  • 11 December 2025: Pinterest agrees to acquire tvScientific
  • 17 February 2026: tvScientific acquisition completes for $465.1 million
  • 24 March 2026: Promote a Pin launches, targeting handled entirely by the graph
  • 5 May 2026: Q1 2026 results report PinRec extended site-wide
  • 4 June 2026: Pinterest announces a planned $4 billion AWS commitment through 2031
  • 17 June 2026: Pinterest announces its Model Context Protocol server and three further AI products
  • 5 August 2026: Q2 2026 results show European advertising growth down to 12 percent

Summary

Who: Pinterest, Inc. owns and operates the Taste Graph. Advertisers and agencies consume it through Ads Manager interest targeting, Performance+, Promote a Pin and, since 2026, tvScientific connected television buys. Brian Johnson led the knowledge engineering work published in 2017; John Milinovich ran the product effort that turned it into a targeting surface. The generic term originated with Hunch, founded by Chris Dixon, Caterina Fake and Tom Pinckney.

What: A three-part structure linking pins, users and a hierarchy of interests, populated by extracting and scoring text attached to saved images, then scoring user affinity for those interests with weighting for recency and frequency.

When: The term dates to Hunch in the late 2000s and eBay's 2011 acquisition. Pinterest opened the graph to advertisers on 21 September 2017, expanding targetable interests from roughly 400 to more than 5,000, and has layered learned embeddings over it since 2018.

Where: Across Pinterest search, home feed and related pins, in Ads Manager and the bulk editor, and, following the February 2026 tvScientific acquisition, in connected television inventory outside Pinterest properties.

Why: Pinterest holds behavioural data that other platforms do not, because users file images by intent rather than react to a feed. The company argues that signal is more commercially valuable than passive engagement. The counterweights are inference error the platform has disclosed at roughly one label in five, performance figures supplied by the seller, and an automation layer that increasingly places the taxonomy beyond advertiser inspection.