What happens inside a clean room

A brand knows who bought. A publisher knows who saw the ad. Neither should hand the other its customer list. A clean room lets both put their data in one place, join it on a hashed ID, and run only the queries they agreed on in advance. What comes out is aggregates: counts, rates and averages over groups big enough that nobody can be singled out. Every query is logged.

That's the whole trick, and it's a genuinely useful one for campaign measurement and audience overlap. What it isn't: a data warehouse, a real-time system, or a way to make a weak measurement design strong. The lift number above is a correlation — people who saw the ad were not chosen at random. And if your question involves one company's data, you don't need a clean room at all.

Real products use far bigger thresholds (tens or hundreds of people, not 5) and more safeguards against clever query combinations. A simplified illustration — not how any specific product works.

Read the longer take: Most clean rooms pitches are lying to you →

• Run the spend-by-category analysis, then slide k up and watch small groups vanish.
• Push k above 10 on the exposure analysis: no lift can be published at all.
• Ask for the matched customers. The room refuses, and the refusal is logged.
• Turn on noise and re-run a few times. The answer wobbles, but the story holds.
• Notice the faded IDs: customers outside the match never enter the result.