Replacing Cookies Moved the Trust Boundary, Not the Tracking

The third-party cookie was never the tracking; it was the cheapest storage slot for the tracking. Replacing it with browser-mediated advertising APIs does not delete the capability — it moves the profile from a vendor's server into the browser, and hands the browser vendor the job of deciding which questions advertisers are allowed to ask. That is a real change in who you have to trust. It is not the same thing as a change in how much you are tracked.

Chrome spent years announcing the death of the third-party cookie, then in 2024 said it would give users a choice instead, which is the clearest signal available that the replacement was harder to make economically acceptable than to specify. The technical reason is not mysterious: measurement and frequency capping are load-bearing parts of the ad market, and any API that supports them necessarily leaks something about cross-site behaviour. The design question was never "does information flow?" It was "who holds it, and under what budget?"

The Cookie Was a Scoping Bug, Not a Feature

A cookie is set by an origin and returned to that origin, which sounds properly scoped until you notice that a script embedded across thousands of publishers can set and read the same cookie on all of them. The identifier is first-party to none of those sites and third-party to all of them, and the profile built from it spans everything the tag was present on. The privacy problem was always the embedding pattern, not the file format.

That is why the industry's two answers diverge. Partitioning — keying storage by the top-level site, so the embedded script gets a different bucket on every publisher — attacks the linkage directly and needs no API at all. Safari and Firefox went that way. Mediation, which is the Privacy Sandbox model, keeps the linkage possible in aggregate and instead constrains how it can be used: on-device computation, noisy aggregates, and a fixed menu of outputs. Partitioning removes the capability. Mediation rations it.

What the APIs Expose by Design

The API surface is worth reading as a list of things the browser decided to allow. Interest-based advertising is expressed as topics derived on-device from recent browsing history, with the caller receiving a small subset rather than the history itself. Remarketing is expressed as interest groups the browser stores and auctions locally, with the bidding logic running in an isolated worklet that the advertiser cannot read the profile from. Conversion measurement is expressed as attribution reports, split into event-level reports that are delayed and noised and aggregate reports that are computed by a separate aggregation service.

In every case the raw signal stays inside the browser or inside a service the browser cooperates with, and the advertiser gets a bounded output. That is a genuine architectural improvement over handing a tag your whole browsing trail. It is also, in every case, the browser vendor deciding the bounds — and a browser vendor that also sells advertising is deciding the shape of a market it competes in. The UK competition authority extracted commitments over exactly this concern, which is a governance answer to a governance problem.

"Aggregated" Is Doing Heavy Lifting

Aggregation sounds like anonymisation and is not. If a report says three people in a segment converted and you control two of them, you have learned about the third — the classic differencing attack. Real systems defend against it with three mechanisms acting together: contribution bounding, so no single user contributes more than a capped amount; noise calibrated to the sensitivity of the query; and a privacy budget that depletes as queries are answered, so the same slice cannot be interrogated forever.

Each of those is a policy choice with a business consequence. Cap contributions low and the heavy users disappear from your measurement. Calibrate noise to a strict budget and small advertisers get numbers indistinguishable from zero, because noise that is negligible against a million conversions is fatal against forty. Set the budget generously and the aggregate reports become a slow-motion reconstruction of individual behaviour, one query at a time.

So when a vendor says the reporting is "aggregated," the only follow-up worth asking is: with what contribution bound, what noise, and what budget across what time window? Without those three numbers, the word describes the transport format and nothing about the privacy property.

What Actually Changes, and for Whom

For a user, the practical change is that the shared identifier is gone or partitioned while browser-mediated inference takes its place. That is progress: the same profile can no longer be read by any script that shows up on enough pages. What it does not do is reduce the total surface. Fingerprinting signals — canvas rendering, font enumeration, screen geometry, timezone, IP — are untouched, and their relative value rises as cookies stop being useful. First-party identity, meaning a login, becomes more valuable because it is the one linkage no browser policy can partition away.

For an ad tech vendor, the change is a loss of raw access and a gain in permitted queries. The firms that win are the ones large enough for noisy aggregates to be statistically useful and integrated enough to work inside the browser's boundaries. The firms that lose are the long-tail data brokers whose entire product was a join key. That is a market restructuring presented as a privacy feature, and both descriptions are accurate.

For a publisher, nothing gets simpler. Consent banners still gate the legal basis, the APIs still need integration work, and the measurement numbers still do not match the old cookie-based ones — which is a fight about attribution methodology being conducted as if it were a fight about privacy.

What I Think

I think this is a net improvement and a smaller one than it is advertised as. Moving the profile into the browser genuinely removes a class of abuse — the tag that sees everything and reports to whoever pays — and the browser is at least a party whose interests are partly aligned with the user's, because it has to be installed by choice. That is better than trusting a data broker whose product is the profile.

But I do not accept the framing that tracking has been solved, and I distrust any roadmap where one company writes the API, operates the aggregation service, and sells the ads the API measures. The honest summary is that the trust boundary moved from a vendor you have never heard of to a vendor whose browser you are running, and the useful question for the next five years is not whether cookies are gone. It is what the privacy budget is, who audits it, and whether anyone outside the vendor can check the noise was actually added.