PARAKEET was a proposal from the Microsoft Edge team for serving targeted advertising without third-party cookies. The acronym stands for Private and Anonymized Requests for Ads that Keep Efficacy and Enhance Transparency, and the mechanism at its center was a cloud service, trusted by the browser, that sat between the user's device and the ad network. Every ad request would be routed through that service, which stripped or coarsened the signals that could identify an individual, added the user's anonymized interest data, and only then forwarded the request onward. It existed because Microsoft disagreed with the approach Google had taken. Google's competing design moved the ad auction into the browser itself. Microsoft argued that doing so would force the ad tech industry to rebuild its infrastructure from the ground up, and proposed keeping the auction where it already ran, on ad network servers, while changing what those servers were allowed to see.

Kelda Anderson, a program manager on Edge product development, published the draft to the W3C's public web advertising mailing list on February 19, 2021, with a presentation to the Web Advertising Business Group scheduled for March 2. Wendy Seltzer, then at W3C, replied the same day that the group would look forward to discussing "this new bird". The avian naming was already a convention: Google's TURTLEDOVE, Criteo's SPARROW, Google's Dovekey. PARAKEET joined a crowded aviary of competing designs, and never left it.

How the anonymizing service worked

The design had three moving parts: interest collection, request anonymization, and rendering.

Advertisers added users to interest groups through a JavaScript call, navigator.joinAdInterestGroup(), deliberately modeled on the equivalent function in TURTLEDOVE. The object passed in carried an origin identifying the advertiser domain, a business name for user-facing transparency, a list of interest labels, a list of ad networks permitted to read those interests, and optionally a set of vector representations. Those representations were the more ambitious element: natural language models such as DSSM or Doc2Vec encoding page content into a fixed-length vector, typically 128 dimensions. Membership carried a hard expiry of 30 days.

The browser stored those interests locally, then worked with the PARAKEET service to convert them into a differentially private form. The explainer described clustering all users' interest vectors into K clusters, then sharing the nearest cluster centroid with probability p and a random centroid otherwise, yielding a formal privacy budget expressed as epsilon.

At ad serving time, the publisher's page called navigator.createAdRequest() with a configuration object. That object named the ad network origin, a reporting worklet URL, a publisher code, an ad unit label, contextual targeting terms, and an array of ad slot properties covering orientation, size, language, creative type, and a per-slot bid floor. A separate field, anonymizedProxiedSignals, listed which signals the service was authorized to add: coarse geolocation, coarse user agent, contextual targeting, and cross-site ad interests.

The browser then sent the request to the PARAKEET service rather than to the ad network directly. The service removed the page URL and page title, truncated the user agent string, and replaced the IP address. Location handling was quantified: the service would track the user population at country, state, county, city, postcode and coordinate granularity over the preceding minute, and reduce precision in the outgoing request until at least 10,000 users shared the same value. If the resulting bundle of signals was still too rare, the service dropped the interest data entirely and passed through only what the page itself could have supplied.

Bid floors were subject to the same logic. A floor of 3.12 that no other seller had ever sent was itself an identifier, so the service could round it down to a commonly seen value. If the winning bid cleared the rounded floor but not the real one, PARAKEET returned nothing and the seller fell back to its own auction.

The winning creative arrived as an opaque URL, unreadable by the publisher page, rendered inside a fenced frame.

From a proxy to a ranking engine

Momentum was rapid in 2021. Microsoft opened biweekly public calls in March. In April it published MaCAW, short for Multi-party Computation of Ads on the Web, a supplementary design meant to reduce how much the trusted service needed to know. Masked LARK, a proposal for aggregate conversion measurement, followed in the summer. Brandon Maslen filed an Intent to Prototype with the Chromium project on August 23, 2021, listing a Chrome flag named Parakeet, no specification, no TAG review requested, and no signal from either Mozilla's Gecko or Apple's WebKit.

Then the design changed. In April 2022 Microsoft published PARAKEET with Noisy Ranking, which the document states supersedes the original. The trusted service stopped being a proxy and became a ranking engine. Instead of forwarding anonymized signals for the DSP to act on, it mapped user data into a low-dimensional embedding using a model the ad tech supplied, added calibrated Gaussian noise, dispatched the vector inside a randomly ordered batch, then ran the buyer's ranking logic itself. The paper proposed concrete settings: a privacy parameter delta of one in a thousand, epsilon of 10, a batch size of 1,000, producing noise with a standard deviation near 0.1 and a bounded ranking error of the same magnitude.

That shift moved PARAKEET closer to SPARROW, the Criteo proposal it had originally distinguished itself from, and further from the argument that had made it attractive to ad tech in the first place.

Public meetings continued into 2022, with the last archived minutes dated August 24 of that year. On October 13, 2023, Microsoft moved the PARAKEET, Noisy Ranking, MaCAW and Masked LARK documents into an archive folder in the W3C community group repository. On March 5, 2024, the same repository was repurposed to describe the Ad Selection API, which PPC Land covered as a new privacy-centric API for Edge. The successor abandoned the browser-vendor-operated service in favor of trusted execution environments run by ad tech companies themselves, and aligned its syntax with Google's design rather than opposing it.

Where it broke down

The explainer was unusually candid about its own weaknesses, listing four attacks in a security section. Repeated requests could leak interest data if contextual signals were unique enough. Timing correlation could let an adversary match an outgoing anonymized request to the incoming one it came from. Contextual signal stuffing could game the uniqueness threshold by flooding many clients with identical signals, then querying one client of interest. And the service itself had to be trusted absolutely, a point the document conceded without resolving.

Centralization was the structural objection. Because the service held more sensitive data than any other party, Microsoft wrote that it expected the service to be operated by the browser vendor rather than federated across certified providers. Critics saw that as replacing many trackers with one, run by a company that also sells advertising. Microsoft also said publishers and advertisers would fund the service through fees comparable to existing exchange and platform charges, meaning the ecosystem would pay a browser maker for permission to transact.

Mozilla, writing in October 2021, named PARAKEET alongside Google's FLoC as proposals it would judge against a single standard of whether privacy improved for everyone, and cautioned that such infrastructure raised centralization risks.

PARAKEET's own sharpest criticism ran the other way. Google's design separated contextual and interest-based requests into two uncorrelated calls. The PARAKEET explainer called that "a weak form of privacy protection, if any", arguing that untreated interest signals could still be used to exchange identity between parties whether they arrived together or apart.

Nothing shipped. There was no Chrome implementation, no Firefox or Safari signal, no W3C specification, and no origin trial in Edge under the PARAKEET name.

Disambiguation

PARAKEET and TURTLEDOVE. TURTLEDOVE was Google's originating proposal for on-device interest-group advertising, later prototyped as FLEDGE and renamed Protected Audience in August 2023. The distinction is auction location: TURTLEDOVE and its descendants ran the auction in the browser, PARAKEET ran it on ad network servers behind an anonymizing proxy.

PARAKEET and SPARROW. Criteo's SPARROW also used a trusted intermediary, but that gatekeeper hosted the bidding models and ran the final auction. PARAKEET originally left bidding and auction logic with the ad network. The 2022 Noisy Ranking revision narrowed that gap considerably.

PARAKEET and MaCAW. MaCAW was not an alternative to PARAKEET but an extension of it, applying multi-party computation so that no single service held the complete picture.

PARAKEET and the Ad Selection API. The Ad Selection API is the successor, not a rebrand. It uses trusted execution environments operated by ad tech firms, adopts k-anonymity and differential privacy, and is deliberately code-compatible with Protected Audience.

Outside advertising, Parakeet also names an Australian Army battlefield communications system introduced in the mid-1990s and, more recently, a family of open speech recognition models. Neither relates to the ad tech proposal.

Recent developments

The lineage that began with PARAKEET has quietly wound down. Google retired most Privacy Sandbox technologies on October 17, 2025, including Protected Audience and Topics, citing low adoption. Microsoft's Ad Selection API origin trial in Edge carried an expiry of November 30, 2025 in the browser's release notes, and the trial does not appear in the origin trial list published with Edge 147 in April 2026. Microsoft archived its PARAKEET repository on GitHub on June 15, 2026, making it read-only.

The market context also changed. Microsoft shut down the Xandr demand-side platform on February 28, 2026 and handed its programmatic buyers to Amazon, leaving Microsoft Monetize as its remaining programmatic asset. The company that spent 2021 arguing the industry should not have to rebuild its ad stack now operates a substantially smaller one.

Timeline

  • January 23, 2021: The WICG privacy-preserving-ads repository is created
  • February 19, 2021: Microsoft publishes the PARAKEET draft and announces it to the W3C web advertising mailing list
  • March 2, 2021: PARAKEET is presented to the W3C Web Advertising Business Group
  • March 15, 2021: First archived biweekly PARAKEET open discussion
  • April 8, 2021: Microsoft publishes MaCAW as a supplementary proposal
  • August 3, 2021: Masked LARK is presented to the Web Advertising Business Group
  • August 23, 2021: Brandon Maslen files an Intent to Prototype for PARAKEET with the Chromium project
  • April 2022: PARAKEET with Noisy Ranking is published, superseding the original proposal
  • August 24, 2022: Final archived PARAKEET open discussion minutes
  • October 13, 2023: PARAKEET, Noisy Ranking, MaCAW and Masked LARK are moved to an archive folder
  • March 5, 2024: Microsoft announces the Ad Selection API and repurposes the repository
  • October 8, 2024: Ad Selection API enters limited developer preview in Edge 130, excluding the EEA and the UK
  • October 17, 2025: Google retires Protected Audience, Topics and most other Privacy Sandbox APIs
  • November 30, 2025: Listed expiry date for the Ad Selection API origin trial in Edge
  • June 15, 2026: Microsoft archives the PARAKEET GitHub repository

Summary

Who: The Microsoft Edge team authored PARAKEET, with Kelda Anderson running the public process and Brandon Maslen filing the Chromium prototype intent. It was debated in the W3C Web Advertising Business Group and the Web Platform Incubator Community Group, against competing proposals from Google and Criteo.

What: A set of browser APIs plus a cloud service, trusted by the browser, that anonymized ad requests. The service stripped identifying context, applied differential privacy to cross-site interest data, enforced population thresholds on location and bid floor values, and forwarded a single combined request to the ad network, which kept its auction on its own servers.

When: Published February 19, 2021. Revised as PARAKEET with Noisy Ranking in April 2022. Archived October 13, 2023. Superseded by the Ad Selection API on March 5, 2024. The GitHub repository was made read-only on June 15, 2026.

Where: In Microsoft Edge, in principle. In practice the proposal lived in GitHub repositories, W3C mailing lists and biweekly community calls. No browser shipped it.

Why: Because Google's plan to move ad auctions into the browser would have required the ad tech industry to rebuild its bidding, pricing, brand safety and measurement systems. PARAKEET was the argument that privacy could be enforced at the network boundary instead, letting existing infrastructure survive. That argument lost on adoption rather than on merit, and the mechanism it opposed has since been retired too.