Cookieless & data clean rooms: real-world application

One advertiser, one retail media network, one room, one decision. Unilever puts something on the order of €8 billion a year into brand and marketing investment, and a growing share of it lands on Amazon: Sponsored Products, Sponsored Brands and Amazon DSP behind brands like Dove, Persil and Hellmann's. The question that follows is unglamorous and expensive to answer badly. When a shopper buys a bottle of Dove after being served four different Amazon placements in a fortnight, which of those placements earns the next euro? This lesson follows that measurement programme inside Amazon Marketing Cloud (AMC), Amazon Ads' own clean room, from the way the question was framed to the budget move it produced. The mechanics, thresholds and constraints described here are real. The output figures are illustrative round numbers, because no advertiser of that size publishes its queries. And note the conflict from the start: the room is built and operated by the company selling the media it measures.

Core concept: framing a question the room can answer

AMC holds pseudonymised, event-level records of your own Amazon campaigns (impressionsimpressionsThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.View full definition →, clicks, Amazon conversions) plus whatever first-party signals you upload, inside the no-raw-records arrangement the foundations lesson sets out. The interesting work at this stage is not the technology. It is drawing the boundary around what can be asked.

Four limits decide which questions are worth typing:

• History runs to roughly a year of events. A twelve-month seasonal comparison is possible; a three-year trend is not, unless you exported the aggregates yourself as you went.
• Outputs are aggregated with a floor. A result row built on fewer than about a hundred users comes back empty. Small cohorts, niche SKUs and short test windows disappear.
• You see your own campaigns. Competitive spend, competitor sales and category share are not in the room.
• Conversions are Amazon conversions. A purchase in a supermarket, on your D2C site or at a different retailer does not exist here.

Given those walls, "what was our Amazon ROASROASReturn on Ad Spend (ROAS) measures the revenue generated for every unit of currency spent on advertising, calculated as revenue divided by ad cost.View full definition → last quarter" is a question the room will answer and the answer will be worthless, because platform ROAS credits whatever placement sat closest to the click. The decision-shaped version is narrower: if a meaningful share of purchases credited to branded search were already reached by upper-funnelfunnelThe customer journey from awareness to purchase, typically Awareness, Interest, Consideration, Decision, Action, with prospects narrowing at each stage.View full definition → display, how much branded search budget should move to prospecting, and what would we need to see to move it?

Write the second question down before running the first query. A pre-registered threshold ("we shift budget if overlap exceeds 50% and the frequency curve flattens below eight impressions") is the only defence against a room that will happily produce a chart supporting last quarter's plan.

Step by step through the programme

1. NAME THE DECIDER, NOT JUST THE ANALYST

The reallocation touches the search budget an agency team is often compensated against. Put the brand lead and the e-commerce lead on the decision, keep the agency on the execution, and agree in advance what result triggers what action. Programmes that skip this produce a deck and no budget movement.

1. UPLOAD THE FIRST-PARTY FILE, THEN ACCEPT HOW LITTLE IT ADDS

A CPG can push hashed email and address records into AMC and match them against Amazon shoppers, using the same resolution logic the CDPCDPSoftware that unifies customer data from every source into one persistent profile that marketing, sales and service teams can act on.View full definition → foundations lesson covers. The uncomfortable arithmetic: a global household brand's opted-in CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → in one market often covers a low single-digit share of the retailer's buyers in that category. Consumer goods companies sell through retailers, so they have never held the purchase relationship. The signal that makes the room useful belongs to Amazon, not to you.

That changes what you negotiate for. You are the smaller data holder, so trade for query flexibility, longer windows and access to path-level templates rather than assuming your file buys you leverage.

1. RECONSTRUCT THE PATH INSTEAD OF THE ATTRIBUTIONATTRIBUTIONA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → REPORT

Standard Amazon reporting credits purchases inside a 14-day window to the ad that was clicked. Path analysis in the room de-duplicates that: for every purchaser, the full sequence of placements they were exposed to, ordered, with unique reachunique reachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → per placement and the overlap between them. The output that matters is not a ranking of channels. It is the proportion of "search-attributed" purchases where display got there first.

1. PLOT THE FREQUENCY CURVE BEFORE ARGUING ABOUT CHANNELS

Bucket exposed households by impression count over fourteen days, then plot purchase rate per bucket. You are looking for where the curve flattens and how much delivery sits beyond that point. Treat it as correlational: heavy category buyers see more ads because they browse more, so the curve tells you where to test a cap, not where the optimum is.

1. PUSH AN AUDIENCE BACK OUT, NOT JUST A CHART

AMC audiences can be activated into Amazon DSP and sponsored ads, which is what separates a measurement programme from a research project. Two builds pay for themselves quickly: suppress Subscribe & Save subscribers of the same SKU from prospecting, and suppress recent purchasers of a product with a long repurchase cycle. No planning cycle required.

What the queries returned, and what they did not

Finding 1: the overlap that ate the ROAS story

Illustrative figures, real shape. Around six in ten purchases credited to branded Sponsored Products came from shoppers already served DSP display in the previous fortnight, and a large majority of those buyers had purchased the brand before. Branded search was posting the best reported ROAS in the platform interface while sitting downstream of demand created elsewhere and of habit created years earlier. Nothing here proves branded search is worthless: it defends the shelf against competitor conquesting, which the room cannot see. What it kills is the argument that branded search deserves the marginal euro because its ROAS number is highest.

Finding 2: the frequency ceiling

Purchase rate climbed steadily to roughly six impressions per household per fortnight, then flattened. Meanwhile a fifth of delivered impressions sat above that point. That is the cheapest finding in any retail media clean room programme, and the one most often left on the table, because acting on it reduces the impression volume the agency reports as reach.

Finding 3: the query that returned nothing

The team wanted a read on a new variant launched two months earlier. The cohort of exposed purchasers fell under the aggregation floor and the rows came back suppressed. Widening the window to a quarter did not fix it, because the SKU had not existed for a quarter. The workaround was to group the variant with its parent range and read at pack level, which answers a different and less useful question.

This is a structural bias worth naming out loud to your finance director: a clean room measures your biggest brands well and your newest, smallest ones badly, exactly inverting where measurement would change the most decisions. Innovation launches still need panel data, geogeoThe practice of making your brand and content visible and citable inside AI-generated answers from tools like ChatGPT, Gemini and Perplexity.View full definition → tests or a holdout, because the room will not talk about them.

The reallocation and the guardrail

The decision that came out of it: move a defined slice of branded search budget (illustrative: 15%) into DSP prospecting, apply a frequency cap in line with the curve, hold two markets unchanged as a comparison, and judge the quarter on total category sales rather than on platform-reported conversions.

The second-order consequence deserves the same slide. Amazon can only tell you that Amazon sales moved. For a Unilever-scale portfolio selling through grocery, discount and e-commerce at different margins, a lift on Amazon that comes from shoppers who would otherwise have bought in a supermarket is not growth, and may be worse than flat once trade terms and fulfilment costs are counted. The room has no view of that trade-off, so the commercial team has to supply it.

CMO action items

• Pre-register the decision rule in writing before the first query runs: the threshold, the amount of budget that moves if it is crossed, and who signs. A room without a pre-registered decision produces quarterly reading material.
• Fund the person, not the licence. Standard AMC access carries no separate fee for advertisers, so the real cost is a SQL-capable analyst who understands the media plan. One named owner beats a rota of agency requests.
• Export aggregate outputs to your own warehouse every month. Retention inside the room runs to about a year, so a multi-year baseline exists only if you built it deliberately, and it is the only asset that survives a change of retail partner.
• Validate one clean room finding a year outside the network, with a geo holdout or a matched-market test. Budget for it as measurement, not as a campaign.

Common mistakes that kill results

• Calling path analysis incrementality. Overlap tells you who was exposed before buying. It does not tell you who would have bought anyway. The honest sentence in the board deck is "branded search is downstream of display", not "display drove the sale".
• Letting the seller keep the only scoreboard. Amazon operates the room, sells the inventory, and defines the conversion. That does not make the outputs wrong, and the event-level detail is genuinely better than what the open web offers. It does mean an unaudited result should never be the sole basis for a nine-figure allocation.
• Optimising to the edge of the room. Every metric available inside it is an Amazon metric, so a team judged on those metrics will keep moving money toward Amazon regardless of total category performance. Set the success measure outside the room before the programme starts.
• Planning SKU-level reads the aggregation floor will suppress. Check expected cohort sizes against the threshold at the design stage; discovering it after three weeks of query writing is a waste of the only analyst you have.
• Running one study instead of building a pipeline. The overlap and frequency findings shift as creative, competition and category demand change. A query rerun monthly is worth ten times a one-off analysis with a slide title.

Key takeaways

• Frame the question as a budget decision with a threshold attached, then run the query. Reversing that order turns a clean room into an expensive confirmation machine.
• In a retail media room, the retailer holds nearly all the useful signal. A CPG's own CRM typically covers a small fraction of the retailer's category buyers, so negotiate for query flexibility rather than assuming your file gives you leverage.
• The recurring finding: branded search sits downstream of display and of existing habit, and its reported ROAS flatters it. The recurring quick win: a frequency cap where the purchase curve flattens.
• Aggregation floors and roughly one year of retained history mean small SKUs, new launches and long trends are invisible. Plan panel data, geo tests or holdouts for those, and export your aggregates monthly.
• A lift inside one retailer's walls is not a lift in demand. For a multi-channel CPG, the reallocation decision has to include the margin consequence of moving volume between trade channels, which no clean room will surface for you.