Cohort-based targeting is the practice of choosing an advertisement for a group of people who share an inferred trait, rather than for an individual recognised by a persistent identifier. The buyer bids on "browsers whose recent history resembles cluster 4,812" or "readers a publisher classifies as car intenders", not on a cookie ID attached to a stored profile. The approach exists because the identifiers that made individual profiling cheap - third-party cookies and mobile advertising IDs - came under pressure from browsers, operating systems and regulators, and the industry wanted interest-based advertising to survive their loss. A cohort is the compromise: enough signal to price an impression, and supposedly too little to single out a person.

How a cohort is built and used

Every implementation follows four steps. An observer first collects behaviour: pages visited, content consumed, purchases, survey answers. An algorithm or a rulebook then sorts that behaviour into a group. A privacy control is applied next, usually a minimum membership size, the principle known as k-anonymity, sometimes with added noise. Finally, only the group label travels to the buyer, who decides whether to bid.

Designs differ on who does the grouping and where. Browser-side cohorts are computed on the device from local history. Seller-side cohorts are computed by a publisher or its data partner from first-party interactions and passed in the bid request. Platform-side cohorts live inside walled gardens and clean rooms, where a buyer receives a segment name and aggregate reporting but never the member list.

In OpenRTB, the real-time bidding protocol most exchanges use, seller-side cohorts ride in existing objects. Guidance published by the IAB Tech Lab in February 2022 places audience signals in user.data with three required inputs: name, the domain of the entity making the determination; ext.segtax, an integer identifying the taxonomy; and segment.id, the node values. A segtax of 4 denotes Audience Taxonomy 1.1, 7 denotes Content Taxonomy 3.0, and values from 500 upwards are reserved for vendor taxonomies. A bid request can therefore state that mydmp.com places a browser in two Audience Taxonomy nodes and say nothing else about the person. Audience Taxonomy 1.1 holds more than 1,600 attribute nodes, according to the document, running six tiers deep as PPC Land's taxonomy explainer sets out.

The same guidance expects such requests to be devoid of user IDs, cookies, IP addresses and deal IDs. That separation is the entire privacy claim. When a cohort label and an identifier travel together, the label merely enriches a profile.

FLoC: the browser as profiler

The first large-scale attempt put the grouping inside Chrome. Federated Learning of Cohorts (FLoC), incubated in the W3C's Web Incubator Community Group, defined a cohort as a short name shared by thousands of people and derived from browsing history. Websites read it through a JavaScript call, document.interestCohort().

In the trial, a Chrome-operated server pipeline clustered SimHash values from synced browsing data using an algorithm called PrefixLSH, according to analysis by MIT researchers Alex Berke and Dan Calacci. Each browser then hashed the registrable domains it had visited over the previous seven days and looked up its cohort ID in a file shipped with Chrome. Browsers with fewer than seven qualifying domains received no cohort. Despite the name, no federated learning was involved.

That origin trial ran across Chrome 89 to 91 in 2021, with a minimum cohort size of 2,000, and produced 33,872 cohorts. Cohorts that correlated too closely with sensitive categories were suppressed under a t-closeness threshold of 0.1; about 2.3% were blocked. Site owners could exclude their pages by sending the header Permissions-Policy: interest-cohort=(), which Drupal began sending by default.

Topics and seller-defined audiences

Google replaced FLoC with the Topics API on January 25, 2022. Instead of an opaque cluster number, the browser maps hostnames to a public, human-curated taxonomy. At the end of each one-week epoch it computes the user's top five topics. A caller using document.browsingTopics() receives up to three, one per recent epoch, and only topics that caller had itself observed. Each epoch carries a 5% chance that the topic returned is random. A revision enlarged the taxonomy from 349 topics to 469, and Chrome sorted its 22 root topics into two utility buckets using industry input. PPC Land reported Canary testing of Topics alongside FLEDGE and Attribution Reporting in March 2022.

A separate lineage began within weeks. The IAB Tech Lab published seller-defined audiences (SDA) as version 1.0 in February 2022, drafted by its Rearc Addressability Working Group of 295 individuals from 146 organisations. Benjamin Dick, senior director of product at the Tech Lab at launch, said "the basic concept is straightforward": publishers classify their own users, map them to the Audience Taxonomy, document each segment under the Data Transparency Standard and relay only taxonomy IDs. The specification names integration with Prebid.js, the open-source header bidding wrapper, as a prerequisite. In November 2024, Tech Lab chief executive Anthony Katsur proposed renaming SDA as Curated Audiences.

Origin and evolution

Grouping audiences is as old as media planning. What changed was the constraint. Google introduced the Privacy Sandbox in 2019 and in January 2020 set a two-year horizon for making third-party cookies obsolete in Chrome. A 2021 Google Canada strategy document, later disclosed in antitrust litigation, framed FLoC as an evolution of interest-based targeting, with TURTLEDOVE playing the same role for remarketing.

On January 25, 2021, Chetna Bindra, a Google group product manager, wrote that simulations of FLoC for in-market and affinity audiences showed advertisers could expect "at least 95% of the conversions per dollar spent" compared with cookies, a claim PPC Land covered at the time. Chrome 89, released on March 2, 2021, carried the trial.

Microsoft proposed PARAKEET, in which a browser-trusted service anonymised ad requests before ad networks saw them. The IAB Tech Lab opened its ID-Less Solutions Guidance for comment on November 21, 2024 and finalised it on July 17, 2025; the 57-page document lists Cross-Context Deterministic Cohorts and Probabilistic Cohorts among its methods.

Why the approach matters to marketers

For buyers, cohorts decide what stays addressable where identifiers are absent: Safari and Firefox traffic, people who refuse consent, and much connected TV and app inventory. For publishers, SDA promised a way to sell first-party knowledge without leaking it. The Tech Lab warned that if device data accompanied a cohort, an audience identified on a premium site could be re-found on a cheaper one.

The browser version was expected to protect revenue. Regulator-supervised tests pointed the other way. The UK Competition and Markets Authority (CMA), overseeing tests of Protected Audience, Attribution Reporting and Topics between November 2023 and June 2024, found publisher revenue per impression around 30% lower without third-party cookies even with the APIs available. Criteo estimated declines of up to 78%, while Index Exchange recorded a 33% fall in CPMs and questioned the value of Topics signals. The 95% figure came from Google's own simulations, not from independent testing.

Limitations and disputes

The core dispute is whether a cohort is private at all. Bennett Cyphers of the Electronic Frontier Foundation argued in March 2021 that a cohort ID gives fingerprinters a head start, since a label shared by a few thousand browsers narrows the search, and urged the industry to stop "re-inventing the tracking wheel". Berke and Calacci later showed that each weekly FLoC cohort was k-anonymous while the sequence was not: more than 50% of devices in their sample carried a unique cohort sequence after three weeks, and more than 95% after four.

Sensitive inference is the second problem. Browsing correlates with health, income and ethnicity, and clusters formed without labels can reproduce those traits. Topics responded with a curated taxonomy, whereas FLoC relied on suppression after the fact.

Granularity cuts both ways. Ari Paparo, then of FreeWheel, said FLoC's granularity meant "it was only going to be useful to ML models". Topics' few hundred categories are legible but blunt.

Seller-side cohorts face a verification gap. The SDA document itself warns that self-attested labels invite a publisher to present a games enthusiasts cohort as high net worth individuals. It also declines to fix a minimum cohort size, citing the 50-user query floor in Google's Ads Data Hub only as a directional reference.

The legal status is unsettled too: draft Australian legislation would treat targeting a class of people as direct marketing.

Not the same as

Interest groups. Protected Audience, originally FLEDGE, stored advertiser-defined audiences on the device for remarketing and ran the auction locally. Membership was never broadcast as a label.

Contextual targeting. Selection is based on the page, not the reader's history. SDA's context signals are contextual; its audience signals are cohorts.

Lookalike audiences. A model finds individuals resembling a seed list, and the output usually resolves to identifiers.

Demographic and generational cohorts. Planning groups such as Gen Z are bought through platform inference rather than a privacy mechanism. Cohort analysis in product analytics, which groups users by sign-up date to track retention, is a separate discipline.

Recent developments

Google confirmed on April 22, 2025 that Chrome would keep third-party cookies. On October 17, 2025 it announced the retirement of Topics, Protected Audience and most other Privacy Sandbox APIs, citing low adoption.

In November 2025 the IAB Tech Lab published analysis promoting on-device cohort modelling for publishers. Ogury took its persona cohorts to connected TV on February 24, 2026, combining survey answers with contextual and Mastercard payment signals. Cognitiv pitched synthetic consumer journey profiles against static segments a month later. Separately, the Australian exposure draft, opened for consultation on August 31, 2026, takes submissions until September 18, 2026.

Timeline

  • 2019: Google introduces the Privacy Sandbox, including the FLoC proposal
  • January 2020: Google sets a two-year plan to make third-party cookies obsolete in Chrome
  • January 25, 2021: Google reports FLoC simulations reaching at least 95% of cookie-based conversions per dollar
  • March 2, 2021: Chrome 89 is released, starting the FLoC origin trial that ran through Chrome 91
  • March 3, 2021: The Electronic Frontier Foundation publishes its critique of FLoC
  • January 25, 2022: Google announces the Topics API as the replacement for FLoC
  • February 2022: The IAB Tech Lab publishes seller-defined audiences version 1.0 with the segtax extension
  • March 2022: Chrome Canary begins testing Topics, FLEDGE and Attribution Reporting
  • 2022: Alex Berke and Dan Calacci publish their post-mortem analysis of FLoC privacy
  • August 2023: FLEDGE is renamed Protected Audience
  • November 2023 to June 2024: CMA-supervised testing of Topics, Protected Audience and Attribution Reporting
  • November 2024: Anthony Katsur proposes renaming seller-defined audiences as Curated Audiences
  • November 21, 2024: The IAB Tech Lab opens its ID-Less Solutions Guidance for public comment
  • April 22, 2025: Google confirms Chrome will keep third-party cookies
  • June 13, 2025: The CMA publishes test results showing roughly 30% lower publisher revenue per impression
  • July 17, 2025: The IAB Tech Lab releases the final ID-Less Solutions Guidance
  • October 17, 2025: Google announces the retirement of Topics and most Privacy Sandbox APIs
  • November 11, 2025: The IAB Tech Lab publishes analysis on on-device cohort modelling
  • February 24, 2026: Ogury extends persona-based cohort targeting to connected TV
  • August 31, 2026: Australia opens consultation on a bill defining direct marketing to include group-level targeting
  • September 18, 2026: Australian consultation closes

Summary

Who. Browser makers, led by Google with FLoC and Topics and Microsoft with PARAKEET, designed device-side cohorts. The IAB Tech Lab standardised seller-side cohorts through seller-defined audiences. Publishers, data providers, supply-side platforms and demand-side platforms operate them, and privacy advocates, researchers and regulators including the CMA have tested and challenged them.

What. A way of selecting advertising for a group sharing an inferred trait instead of for an identified individual, relying on minimum group sizes, curated taxonomies or noise to limit re-identification, and transmitted as a label such as a FLoC cohort ID, a Topics value or a segtax-coded taxonomy node in OpenRTB.

When. Group buying predates digital media, but the privacy-driven form emerged with the Privacy Sandbox in 2019, was trialled in Chrome in 2021, split into Topics and seller-defined audiences in early 2022, and lost its browser implementation when Google retired Topics on October 17, 2025.

Where. In the browser for FLoC and Topics, in publisher systems and Prebid.js for seller-defined audiences, inside walled gardens and clean rooms for platform segments, and increasingly on devices and in connected TV.

Why. Third-party cookies and mobile identifiers faced restrictions, and the industry sought interest-based advertising that did not track individuals. Measured revenue shortfalls, privacy attacks on cohort sequences and legal definitions that treat group targeting as personal data processing have kept the approach contested.