FMCG · Retail media and clean rooms: sharing data without giving it away
A snack manufacturer wants to know if its in-store display and app ads actually sold more chips last quarter. The retailer holds the answer inside its loyalty card and point-of-sale (POS) data. Neither side wants to hand over raw files. The brand doesn't want to expose its media spend and audience logic, and the retailer won't release shopper-level purchase records. The deal happens anyway, inside a "clean room." Nobody sees the other's raw data. Both walk away with a lift number.
This setup is now the default structure for retail media deals, so it's worth knowing how it works and where it can still fail.
What a clean room actually is
A data clean room is a controlled environment where two or more parties bring their data, run agreed-upon queries against it, and only see the aggregated output. Neither party ever sees the other's underlying records.
Technically, this relies on a few tools. Hashed matching turns identifiers such as email or loyalty ID into irreversible codes, so both sides can match records without revealing them. Differential privacy adds statistical noise so no individual can be re-identified from results. KKThe average number of new users each existing user generates through referrals. Above 1.0, growth compounds on itself and becomes exponential.View full definition →-anonymity thresholds are a rule that no output can be shown unless it represents at least *k* people, commonly 25 to 100 depending on the retailer.
Major implementations: Amazon Marketing Cloud, Walmart Connect's clean room, Kroger Precision Marketing (via 84.51°), and Google's Ads Data Hub. Snowflake and Databricks also offer neutral clean room infrastructure that brands and retailers can jointly license instead of trusting either party's platform.
Why this exists: the regulatory squeeze
Several forces pushed the industry here at the same time.
Signal loss, even without the cookie cutoff. For years the industry planned around Google removing third-party cookies from Chrome. That never happened. In July 2024 Google dropped its deprecation plan, in April 2025 it said it would not add a separate cookie-choice prompt either, and in October 2025 it retired most of its Privacy Sandbox replacement technologies. Cookies still work in Chrome. Cross-site tracking is still weaker than it used to be, though: Safari and Firefox block third-party cookies by default, Apple's App Tracking Transparency limits mobile IDs, and many shoppers decline consent banners. Brands still lose sight of shoppers across sites and apps, and first-party datafirst-party dataData collected directly from your own customers and prospects through your own channels: your most reliable and privacy-compliant source.View full definition → matters more as a result.
Privacy law. In the EU, the GDPRGDPREU regulation governing how organizations collect, store and use personal data, with fines tied to global revenue for breaches.View full definition → (General Data Protection Regulation, enforced by national Data Protection Authorities) requires a lawful basis to process personal data and gives shoppers rights to access, correct, or delete it. The US has no single federal privacy law, but state laws like the CCPA/CPRA (California Consumer Privacy Act, updated by the California Privacy Rights Act, enforced by the California Privacy Protection Agency) give similar rights and impose obligations when data is "sold" or "shared." That term is broad enough to cover many ad-targeting arrangements. Close to 20 other states now have comparable laws.
Retailer leverage. Retailers like Kroger, Walmart, Tesco (Clubcard), and Carrefour realized their first-party loyalty data is more valuable and more durable than any cookie. Retail media is one of the fastest-growing ad segmentssegmentsDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.View full definition → globally. eMarketer puts US retail media spend at roughly $60 billion a year as of 2025, with Amazon holding the largest single share.
Clean rooms let retailers monetize that data without a "sale" in the legal sense, since raw personal data never leaves their environment.
How a lift measurementlift measurementThe share of results (sales, conversions, revenue) that only happened because of a marketing action, not what would have occurred anyway.View full definition → actually runs
Say the snack brand ran a four-week retail media campaign. The workflow:
- Brand uploads its campaign exposure log (which loyalty IDs, hashed, saw the ad) into the clean room.
- Retailer's POS data (same hashed loyalty IDs, purchase history) sits in the same environment.
- A pre-agreed SQLSQLSales Qualified Lead: a prospect the sales team has validated as ready for direct outreach and a proposal, having passed clear qualification criteria.View full definition →-style query joins the two on the hashed ID and buckets shoppers into "exposed" vs "control."
- Only the aggregate result returns: incremental sales lift, spend efficiency, category halo effect.
A simplified version of the query logic:
SELECT
campaign_group,
COUNT(DISTINCT hashed_loyalty_id) AS shoppers,
SUM(basket_value) AS total_sales
FROM matched_exposure_and_pos
WHERE k_anon_check >= 25
GROUP BY campaign_group;That k_anon_check >= 25 line is the governance control. If a segment (say, a niche flavor SKU in a small store cluster) has fewer than 25 matched shoppers, the query returns nothing rather than a number that could be reverse-engineered to identify individuals.
Worked example: the exposed group of 50,000 shoppers spends an average of $4.20 on the category post-campaign. A matched control group of 50,000 spends $3.80. Lift = ($4.20 - $3.80) / $3.80 = 10.5%. That's the only number the brand's media team sees. They never see which 50,000 people, what else was in their baskets, or their names.
Where governance still has to do real work
Clean rooms reduce risk. They don't remove the need for oversight. Three checks matter most.
Query auditing. Someone should log every query run against the clean room and review it for "creep": a brand asking overly narrow questions designed to isolate small, identifiable segments. This is a classic re-identification attack.
Consent lineage. The loyalty data feeding the clean room has to trace back to a valid consent or legitimate-interest basis under GDPR, or a proper CCPA opt-out mechanism. If the underlying consent was defective, aggregating it in a clean room doesn't fix that. Governance teams should be able to answer "what did the shopper agree to" for any data reaching the room. The UK ICO's guidance on anonymisation and pseudonymisation is a solid free reference on where "anonymised enough" legally starts.
Vendor and cross-border checks. Many clean room providers are US hyperscalers (Amazon, Google, Microsoft, Snowflake, Databricks). A European retailer running EU shopper data through a US-headquartered clean room needs a valid transfer mechanism. The 2020 *Schrems II* ruling struck down the old EU-US Privacy Shield. Its replacement, the 2023 EU-US Data Privacy Framework, survived a first challenge at the EU General Court in September 2025, but that case is under appeal to the Court of Justice. Many companies keep Standard Contractual Clauses (SCCs) in place as a fallback, and that is sensible while the question stays open.
Knowledge check
1. What is the core structural feature that makes a data clean room acceptable to both a brand and a retailer who don't trust each other with raw data?
2. Why would a retailer use a k-anonymity threshold (e.g., no output shown unless it represents at least 25-100 people) in a clean room query?
3. A brand wants to know exactly which individual loyalty card holders saw its app ad and later bought chips, so it can build a personalized retargeting list. Why would a clean room NOT be the right tool for this goal?
4. Select ALL correct answers about why retail media clean room deals have become the industry default.
Select all the correct answers.
5. Select ALL correct answers about the technical mechanisms that make clean rooms privacy-safe.
Select all the correct answers.
The limits of the model
Clean rooms solve the "don't hand over raw data" problem. They don't solve everything:
- Output leakage over time. Running many slightly different queries against the same population can let a sophisticated party statistically reconstruct individual-level insight. This is a known weakness called a "differencing attack." Retailers cap query frequency and log query history to catch it.
- Asymmetric power. Amazon and Walmart set the rules of their own clean rooms. A mid-sized snack brand has little say over which queries are even allowed.
- No single global standard. IAB Tech Lab and others have published frameworks, but implementations differ by retailer. A brand working with five retailers manages five clean room setups, five data dictionaries and five governance postures.
🎬 [VIDEO: "What Are Data Clean Rooms? (And Why Retailers Love Them)" - youtube.com - search for recent explainers from IAB Tech Lab or major ad-tech conference talks (LiveRamp, Snowflake, InfoSum) for a visual walkthrough of the matching and query flow described above]
Key Takeaways
- A data clean room lets a brand and a retailer join and query data without either seeing the other's raw, identifiable records. Outputs are aggregated and threshold-gated (commonly k ≥ 25 to 100).
- The model grew because of privacy law (GDPR, CCPA/CPRA and other US state laws), lost tracking signal from browsers, apps and consent refusals, and retailers treating loyalty data as a durable asset. Google kept third-party cookies in Chrome, but first-party data still matters more each year.
- Real infrastructure includes Amazon Marketing Cloud, Kroger Precision Marketing/84.51°, Walmart Connect, Google Ads Data Hub, and neutral providers like Snowflake and Databricks.
- Governance work continues inside the clean room: query auditing, consent lineage back to a GDPR or CCPA basis, and cross-border transfer checks (the Data Privacy Framework is still being litigated, so many keep SCCs as backup).
- Re-identification risk drops but doesn't disappear. Repeated narrow queries can still leak individual-level insight, so query logging and frequency limits matter as much as the initial architecture.