Data Clean Rooms

The Problem They Solve

A brand and a publisher both hold customer data. Combining it would reveal which ad exposures produced which purchases. Neither can legally or commercially hand the other their raw customer records.

A clean room lets both upload data into an environment where only aggregated, privacy-protected outputs can be extracted. Matching happens on hashed identifiers; neither party can query individual records.

What They Are Used For

Measurement. Attributing conversions to media exposure where the exposure and conversion data sit with different parties.

Audience overlap analysis. Understanding how much of a publisher's audience you already reach.

Activation. Building audience segments from combined data for targeting, without transferring the underlying records.

The Main Types

Walled garden clean rooms — Google Ads Data Hub, Amazon Marketing Cloud, Meta Advanced Analytics. Access to the platform's own data under its rules. Powerful within that platform and unable to compare across platforms, which is the point.

Neutral or independent clean rooms — InfoSum, Habu, LiveRamp and similar. Operate between parties without owning the media, permitting cross-platform analysis.

Cloud-native — Snowflake, BigQuery and Databricks offer clean room functionality within data platforms organisations already use.

The Limitations

Aggregation thresholds mean small segments return no results, which frustrates granular analysis.

Query restrictions prevent many analyses you might want to run, and the restrictions differ by provider.

Match rates are rarely complete — matching depends on shared identifiers, and a meaningful proportion of records fail to match.

Walled garden clean rooms answer questions about that garden only. They cannot tell you whether Google or Meta deserved the credit.

Who Should Consider Them

Clean rooms suit large advertisers with substantial first-party data and media budgets big enough to justify the operational cost. They require data engineering capability and are not a self-service marketing tool.

For most organisations, improving first-party data collection and running incrementality tests delivers more measurement value per unit of effort than a clean room implementation.

What Makes One Work in Practice

Clean rooms fail on the unglamorous parts, not on the privacy technology.

Match rate decides everything. If only a minority of your customers can be matched to the partner's data, the analysis describes that minority — which is systematically the more digitally engaged, more frequently logged-in part of your base. A clean room with a low match rate does not produce a weaker answer; it produces a confidently wrong one about a skewed group. Ask for the match rate before the contract, not after the first analysis.

Your own data has to be ready. Identifiers normalised, hashed in the agreed way, deduplicated, and reflecting a consistent definition of a customer. Most of the effort in a clean-room project is this, before anything joins.

Somebody has to be able to write the query. The interfaces are SQL-shaped and constrained by aggregation thresholds, so a team with no analyst will buy access and not use it.

The Limits That Decide Whether It Is Worth It

The constraints that make a clean room private also make it awkward, and they are what determine whether one earns its cost.

Aggregation thresholds mean small segments return nothing. Queries below a minimum audience size are suppressed, so the narrow, high-value analyses people most want are often exactly the ones that cannot be run.

You cannot export the joined data. Outputs are aggregates. Anything you want to use downstream has to be expressible as an aggregate.

Each partner is a separate room. Cross-platform measurement — the reason most people want one — usually means several rooms with incompatible outputs, which is the problem you started with in a more expensive form.

And the costs are ongoing: the licence, the engineering to keep data flowing, and the analyst time to use it.

The honest test: name the specific decision the clean room will settle, and what you will do differently depending on the answer. If that is vague, the project will produce interesting charts and no decisions — the same test that governs a segmentation or a dashboard.