The State of Data Clean Rooms

Written By Jonathan D'Souza-Rauto

Data Clean Rooms

Data privacy regulations and signal degradation have fundamentally changed how brands measure and optimise media campaigns. Data Clean Rooms (DCRs) have emerged as the benchmark infrastructure for privacy-safe data collaboration across a variety of vendors, both giants of data infrastructure / cloud computing to more nible providers. This post explores what good looks like for organisations leveraging DCRs to achieve different use cases and considers whether DCRs actually have been successful since the existence of it as a martech buzzword.

The Basics - Data Clean Room Architecture

At its core, a Data Clean Room allows two or more parties to join first-party datasets in a secure environment where raw customer data is never exposed. Modern data clean rooms rely on differential privacy, cryptographic hashing (e.g. SHA-256) and query-level aggregation controls to output actionable insights while preventing data extraction. But increasingly the requirement to just depend on a deterministic signal is becoming less relevant, although any DCR output will always be stronger on an authenticated, aggregated input.

Query execution inside a clean room often relies on privacy-restricted SQL that enforces strict aggregation thresholds (e.g. minimum group sizes of 50 or 100 users) to prevent re-identification or individual tracking.

-- Example: Aggregated Overlap Query in Snowflake / AWS Clean Room
SELECT 
  a.segment_name,
  COUNT(DISTINCT a.hashed_email) AS intersected_users,
  SUM(b.purchase_amount) AS total_revenue
FROM brand_first_party_data.customers a
INNER JOIN publisher_data.impressions b
  ON a.hashed_email = b.hashed_email
GROUP BY 1
HAVING COUNT(DISTINCT a.hashed_email) >= 100;

Data clean room environments broadly fall into 4 primary categories:

DCR Category Description & Key Platforms
Walled Garden DCRs Single-platform ecosystems (Google ADH, Amazon AMC, Meta AA) built to analyse platform ad spend against conversions.
Neutral Cloud Infrastructure Cloud-native environments (Snowflake, Databricks, GCP BigQuery Clean Rooms, AWS Clean Rooms) allowing custom multi-party joins.
Agency & Independent Solutions Turnkey applications (InfoSum (part of WPP), LiveRamp/Habu (part of Publicis), Decentriq) offering UI query builders, identity matching or confidential computing.
MMP Clean Rooms Mobile Measurement-focused clean rooms (Appsflyer) tailored for in-app attribution and privacy-safe app cross-promotion.

Measurement & Activation Use Cases

A Data Clean Room is not always straight forward to get up & running, therefore having set use cases is fundamental to understanding the value it could give.

1. Advanced Cross-Channel Measurement:

  • Incrementality & Lift Modelling: Run matched-market or test-vs-control query frameworks directly against exposure logs to quantify true incremental sales.
  • Multi-Touch Attribution (MTA) & Path-to-Purchase: Query impression, click and conversion timestamps across channels to build custom attribution models outside ad platform UI defaults.
  • MMM Calibration: Feed granular, privacy-cleansed aggregate reach/frequency and conversion baselines straight into Marketing Mix Models.
  • 1P Audiences: Combine your digital and offline footprint to dig into deeper insights around ad exposure, customer journeys & overlap.

2. Privacy-Safe Activation & Targeting:

  • Lookalike & Propensity Modelling: Build cohort models inside the DCR by intersecting high-LTV CRM segments with publisher data, pushing target flags back to DSPs via identifiers or Deal ID activation or even something more agentic.
  • Suppression & Frequency Capping: Exclude existing customers or over-exposed audiences across multiple platforms where supported.
  • Commerce Collaboration: CPG brands join first-party data directly with retailer POS datasets to attribute offline in-store sales to digital campaigns.
-- Example: Generating an Activation Cohort via Privacy-Safe Match Keys
SELECT 
  user_match_id,
  CASE WHEN ltv_score > 80 THEN 'High_Value_Lookalike' ELSE 'Standard' END AS activation_segment
FROM clean_room_collaboration.audience_matches
WHERE opt_in_flag = TRUE
  AND last_conversion_days <= 90;

Deep-Dive: Amazon Marketing Cloud (AMC)

Among all walled garden clean rooms, Amazon Marketing Cloud (AMC) has emerged as arguably the most mature and commercially actionable. Built on AWS cloud infrastructure, AMC provides pseudonymous, event-level impression, click and conversion logs across the entire Amazon Ads ecosystem—including Sponsored Products, Sponsored Brands, Sponsored Display, Amazon DSP and Prime Video. This is especially powerful for Endemic brands who sell on Amazon.

Whilst Google were first to market with its Ads Data Hub (ADH) solution, it has had its fair share of struggles and has seen Amazon overtake it, coinciding with Amazon's adtech ambitions growing to take on Google as the top dog across Search & Programmatic.

Rather than relying on pre-aggregated dashboard reporting, AMC gives brands direct query access to analyze raw consumer paths to purchase and join Amazon's ad signals with their own off-Amazon 1st-party CRM data via AWS S3 uploads. But of course there are privacy limits in place in order to not pull individual user data out.

The Role of AMC in the Modern Tech Stack

Event-Level Analysis
  • Cross-Media Overlap Analysis: Measure the exact sales lift achieved when shoppers are exposed to both upper-funnel Streaming TV / DSP ads and lower-funnel Sponsored Search ads versus single-channel exposure.
  • Path-to-Conversion Mapping: Uncover the multi-touch sequencing of ad formats that yields the highest New-to-Brand (NTB) purchase rates.
  • Closed-Loop Activation: Query custom audience cohorts (e.g. high-value cart abandoners or non-converting high-frequency exposed users) and push them directly into Amazon DSP as custom targeting / suppression rules / bid modifiers.

Sample AMC SQL Queries

AMC uses a specialised dialect of SQL executed against pseudo-anonymous event tables like sponsored_ads_traffic, dsp_impressions and amazon_attributed_events_by_conversion_time. All output queries must pass strict privacy checks (>100 distinct users per aggregate group). Nowadays, Amazon have made it easier for non-technical users to use it, by providing templates or using their Gen AI builder to write queries from a prompt. Here are a few examples of queries:

1. Ad Format Overlap & Conversion Rate Lift

This query evaluates how conversion rates and total purchase value differ between users exposed strictly to Sponsored Ads versus users exposed to both Sponsored Ads AND Amazon DSP.

-- Measure conversion overlap between Sponsored Ads and Amazon DSP
WITH user_exposure AS (
  SELECT 
    user_id,
    MAX(CASE WHEN dataset = 'sponsored_ads' THEN 1 ELSE 0 END) AS exposed_sponsored_ads,
    MAX(CASE WHEN dataset = 'dsp' THEN 1 ELSE 0 END) AS exposed_dsp
  FROM (
    SELECT user_id, 'sponsored_ads' AS dataset FROM sponsored_ads_traffic
    UNION ALL
    SELECT user_id, 'dsp' AS dataset FROM dsp_impressions
  )
  GROUP BY user_id
),
conversions AS (
  SELECT 
    user_id,
    COUNT(DISTINCT conversion_event_id) AS total_orders,
    SUM(product_sales) AS total_revenue
  FROM amazon_attributed_events_by_conversion_time
  WHERE tracked_asin IN ('B00EXAMPLE1', 'B00EXAMPLE2')
  GROUP BY user_id
)
SELECT 
  CASE 
    WHEN e.exposed_sponsored_ads = 1 AND e.exposed_dsp = 1 THEN 'Both (Sponsored Ads + DSP)'
    WHEN e.exposed_sponsored_ads = 1 THEN 'Sponsored Ads Only'
    WHEN e.exposed_dsp = 1 THEN 'DSP Only'
  END AS exposure_segment,
  COUNT(DISTINCT e.user_id) AS total_exposed_users,
  COUNT(DISTINCT c.user_id) AS converting_users,
  ROUND(COUNT(DISTINCT c.user_id) * 100.0 / COUNT(DISTINCT e.user_id), 2) AS conversion_rate_pct,
  SUM(c.total_revenue) AS aggregate_revenue
FROM user_exposure e
LEFT JOIN conversions c ON e.user_id = c.user_id
GROUP BY 1
HAVING COUNT(DISTINCT e.user_id) >= 100;

2. Generating an Audience Segment for DSP Retargeting

Unlike standard analytics tools, AMC allows you to output custom user cohorts (using AMC Audience features) to instantly create high-intent audiences. Below is a query identifying users who added items to their cart in the last 14 days but have not yet purchased.

-- Generate an AMC Custom Audience for DSP Abandoned Cart Activation
SELECT DISTINCT 
  user_id
FROM sponsored_ads_traffic
WHERE event_type = 'add_to_cart'
  AND tracked_asin IN ('B00EXAMPLE1', 'B00EXAMPLE2')
  AND event_date >= CURRENT_DATE - INTERVAL '14' DAY

EXCEPT

SELECT DISTINCT 
  user_id
FROM amazon_attributed_events_by_conversion_time
WHERE tracked_asin IN ('B00EXAMPLE1', 'B00EXAMPLE2')
  AND conversion_event_date >= CURRENT_DATE - INTERVAL '14' DAY;

AMC at this point in time is now mandatory for any mature advertiser or agency to get the most out of Amazon advertising, especially as an endemic brand. It does also provide value for non-endemic brands, however it does also suffer from the ability to truly understand users when Amazon cannot tie ad exposure back to a login, which is especially apparent on Amazon DSP 3P Exchanges alongside a conversion event on a website / app that isn't Amazon. So like any technology, it will never be perfect.

Platform Deep-Dive: Navigating the Ecosystem

This section deepdives into the different type of data clean room categories and the individual solutions that exist beyond AMC.

1. Walled Garden DCRs

Google Ads Data Hub (ADH)

GCP / BigQuery
  • Core Architecture: Hosted on Google Cloud Platform (GCP) and powered by BigQuery. Combines Google log-level ad event data (YouTube, DV360, Google Ads) with advertiser BigQuery datasets.
  • Key Strengths: Unrivalled depth for YouTube reach/frequency, impression pathing and Google-owned inventory analysis. It is still the strongest way to measure the impact of YouTube.
  • Trade-offs & Limitations: Enforces strict 50-user aggregation checks and has heavy dependency on 3P cookies for ad impressions served on non-Google inventory.

Meta Advanced Analytics (AA)

Meta Analytics
  • Core Architecture: Privacy-enforced SQL environment joining advertiser offsite conversion data from the pixel / CAPI directly with Meta impression/click logs.
  • Key Strengths: By far the most powerful dataset available to Meta with the ability to pull reports in aggregation beyond Ads Manager e.g. conversion parameter level.
  • Trade-offs & Limitations: Still technically a beta and not available to most brands / agencies. All queries required manual approval by Meta reps to run.

2. Neutral Cloud Clean Rooms

Snowflake Clean Rooms / Databricks

Zero-Copy Architecture
  • Core Architecture: Native enterprise data sharing without physically copying or moving data files outside the host cloud region.
  • Key Strengths: Easy alignment with existing data infrastructure for advertisers; native support for Python, SQL and advanced machine learning models.
  • Trade-offs & Limitations: Requires data engineering talent and active cloud warehouse instances to leverage.

GCP BigQuery Clean Rooms & AWS Clean Rooms

Cloud Infrastructure
  • Core Architecture: Infrastructure-level clean room capabilities built directly on top of BigQuery data sharing and AWS S3 storage tiers.
  • Key Strengths: Flexible pay-as-you-go compute pricing with native support for SQL, differential privacy rules and custom machine learning pipelines.
  • Trade-offs & Limitations: Requires dedicated engineering resource to set up schema mappings and connect outputs to external ad activation channels.

3. Agency & Independent Solutions

InfoSum (WPP / Choreograph)

Federated Bunkers
  • Core Architecture: Patented non-relational "Bunkered" architecture where datasets are never combined or centralised into a single database.
  • Key Strengths: Highest-grade privacy posture; InfoSum are considered the holy grail of DCRs prior to WPP's acquisition.
  • Trade-offs & Limitations: With a hold-co oversight, the future state of it remains grey vs previous neutrality.

LiveRamp / Habu (Publicis)

RampID & UI Engine
  • Core Architecture: Merges Habu’s intuitive multi-cloud UI query builder with LiveRamp’s global RampID identity graph.
  • Key Strengths: Scalability for activation with RampID; core asset within Publicis/Epsilon stack. Habu were well known for sitting on top of the Walled Garden DCRs also.
  • Trade-offs & Limitations: Similar to InfoSum, the future state of this DCR looks more like a Publicis agency cog.

Decentriq

Confidential Enclaves
  • Core Architecture: Built on hardware-level Confidential Computing.
  • Key Strengths: European centric similar to InfoSum roots, ideal for European banking and healthcare.
  • Trade-offs & Limitations: Not as well known and threat of the cloud infrastructure providers looms over it.

4. Mobile Measurement Partner (MMP) Clean Rooms

AppsFlyer Data Clean Room

Mobile Attribution
  • Core Architecture: Mobile-first DCR tailored to aggregate in-app event data and ad exposure logs.
  • Key Strengths: SSafely matching device signals without exposing raw IDFA/IDFV data, including non-App touchpoints.
  • Trade-offs & Limitations: It is an additional cost in a typical Appsflyer contract, which is even on top of Data Locker (raw event data access).

Identity Resolution & Matching Mechanics

A clean room is only as effective as the join rate between datasets. Without third-party cookies, identity matching requires structured identity graphs using deterministic keys (hashed emails, phone numbers) or unified identity frameworks (UID2, RampID).

{
  "entity_id": "usr_98a7f23c",
  "match_keys": {
    "hashed_email_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
    "hashed_phone_no": "183a628886361a3554449a5b3a372eb0f171542f7c223c727be0f81d1ef7c0c1",
    "uid2": "AgAAAAR13Y...=="
  },
  "consent_flags": {
    "targeted_advertising": true,
    "measurement": true
  }
}

Advanced Diagnostics & Privacy Controls

Operating inside a clean room requires passing strict privacy checks that are programmatically enforced before query results are delivered.

1. Differential Privacy Injection

Clean room engines dynamically inject statistical noise into query outputs to prevent differential privacy attacks (reconstructing specific individual records through iterative query subtractions).

2. Minimum Aggregation Floor (K-Anonymity)

If an output cohort yields fewer than the platform's set threshold (typically k=50 or k=100), the query drops the row entirely or returns a NULL value to protect user identity.

Example of Query Execution & Noise Injection

How DCR privacy controls evaluate SQL queries before returning results

Step 1 ⚡

SQL Execution

Analyst submits restricted SQL join across encrypted 1P customer datasets.

Step 2 🛡️

K-Anonymity Check

Evaluates distinct cohort count. If aggregate users are less than or equal to K (e.g. 50 or 100 users), the query drops the row or yields NULL.

Step 3 🎲

Laplacian Noise

Applies mathematical noise (Differential Privacy) to result in aggregated output.

Step 4 📊

Cleansed Output

Safe, aggregate statistics are released for attribution, incrementality or campaign activation.

Privacy Formula: Final Output = Aggregate Value + Random Noise Scale (based on query sensitivity / privacy budget)

Are Data Clean Rooms Actually Successful?

The marketing industry has poured millions into DCR implementations over the past few years, but the reality of it proving value is mixed. Success depends entirely on the use case, data maturity and technical setup.

🟢 Where DCRs Are Winning

  • Commerce Media: Unquestionable success for CPG brands linking digital ads directly to in-store register receipts.
  • Walled Garden - Amazon: Amazon AMC power-users consistently drive innovation when AMC is involved with Amazon-run campaigns.
  • Privacy & Legal Compliance: Provides ironclad governance for legal teams uneasy about 1st-party data sharing.
  • Omnichannel Analysis: The ability to tie online & offline data is increasingly easier to do in a DCR.

🔴 Where DCRs Fall Short

  • Low Match Rates: Much like Customer Data Platforms, if an identifier cannot be matched, the end output is less trustworthy / accurate.
  • High Resource Barriers: SQL/Python skills was mandatory (Gen AI changing this), heavy compute costs and complex integration pipelines.
  • Platform Fragmentation: Managing 4-5 different DCR interfaces across Google, Meta, Amazon and Cloud partners creates operational tax.
  • The Agency Play: The acquisitions of InfoSum & LiveRamp / Habu shows the tech is powerful but becoming less accessible for known vendors.

The Verdict: DCRs are not a magic bullet for baseline media reporting or necessarily solve every single marketing use cases. They are a specialised strategic infrastructure best suited for enterprise brands with strong 1st-party data, high ad spend and dedicated data engineering / analytics resource. For advertisers heavily spending on Walled Gardens, it is a no brainer to use the Walled Garden specific DCRs. But for the more neutral / independent solutions, make sure you know what you want to get out of it before attempting to spin it up or test it.

Jonathan D'Souza-Rauto
Next

A deepdive into Meta Ads Tracking