Key Takeaways
- Between 2019 and mid-2024, Amazon DSP systematically overcounted attributed revenue by 20 to 40 percent, according to industry estimates, giving itself full credit for organic sales that would have happened without any ad exposure.
- Amazon rebuilt the DSP attribution framework in 2025, introducing 3 major upgrades: multi-touch attribution with AI modeling, extended lookback windows of up to 5 years in Amazon Marketing Cloud, and randomized controlled trials that prove incremental lift against a control group.
- The new framework passes the pause test: brands now see measurable drops in incremental sales when DSP campaigns go off and measurable lifts when they come back on, which did not happen under the old attribution model.
- H&R Block ran the new attribution model in late 2025 and recorded 144 percent conversion lift on campaigns introducing new customers, while campaigns retargeting existing customers showed significantly lower incremental lift.
- A pet supplement brand running a 60-day holdout test in Q4 2025 recorded $47,000 in proven incremental revenue on $12,000 in DSP spend, a real 4x ROAS verified against a matched control group.
- Brands using the new attribution framework in late 2025 report 30 to 40 percent efficiency gains and 35 percent reductions in cost per acquisition, all measured against control groups rather than dashboard claims.
General Summary
Between 2019 and mid-2024, Amazon DSP developed a reputation for inflated attribution, with industry estimates suggesting 20 to 40 percent of credited sales were non-incremental. Brands would run campaigns for 60 to 90 days, see impressive ROAS figures, pause everything, and watch revenue stay exactly flat. The cause was last-click attribution with short lookback windows that gave DSP full credit for purchases that would have happened organically. Amazon responded with a complete attribution rebuild in 2025, introducing multi-touch attribution with AI modeling, lookback windows of up to 5 years in Amazon Marketing Cloud, and randomized controlled trials that measure actual causal lift against a suppressed control group. Early results from the new framework are measurable: H&R Block recorded 144 percent conversion lift on new-customer campaigns, a pet supplement brand proved $47,000 in incremental revenue on $12,000 in spend, and a kitchen gadget brand improved DSP efficiency by 35 percent by reallocating budget away from high-propensity audiences. Amazon DSP ad revenue grew 24 percent year-over-year in 2025. For brands that abandoned DSP based on 2023 experiences, the specific attribution failures that drove that decision have been rebuilt.
Extractive Summary
Brands spent tens of thousands on DSP, watched dashboards claim four-figure ROAS, paused everything, and saw zero change in actual sales. Amazon rebuilt the entire DSP attribution framework with 3 major upgrades: multi-touch attribution with AI modeling, extended lookback windows of up to 5 years, and randomized controlled trials proving incremental lift. The new framework compares performance with DSP on against performance with DSP off and credits only the measurable difference. The rebuilt system passes the pause test: incremental sales drop when DSP goes off and rise when it comes back on. Brands that quit DSP in 2023 or 2024 because the attribution was inflated made a rational decision. Avoiding DSP in 2026 for the same reason means avoiding a problem that no longer exists.
Abstractive Summary
Amazon DSP’s credibility collapse between 2019 and 2024 was a measurement failure, not a product failure. The platform’s reliance on last-click attribution with short lookback windows produced systematic overcounting that industry analysts estimated at 20 to 40 percent of attributed revenue. Amazon’s response, rolled out through Accelerate 2025 and subsequent updates, represents one of the most significant attribution infrastructure overhauls in programmatic advertising. The shift from deterministic last-touch measurement to probabilistic multi-touch modeling with randomized control trial capability mirrors the broader industry movement toward incrementality-based advertising. For established Amazon sellers, this transition creates a narrow window. Competitors who recognise the attribution rebuild will capture incremental market share from upper-funnel advertising while those operating on 2023 assumptions continue to underinvest in channels that now prove their value.
What Happened to Amazon DSP Between 2023 and 2024?
Brands spent tens of thousands on DSP, watched dashboards claim four-figure ROAS, paused everything, and saw zero change in actual sales. The attribution was overcounting. This pattern repeated across dozens of accounts between 2019 and mid-2024. A brand would run DSP for 60 or 90 days, see impressive dashboard numbers, pause the campaigns, and their revenue stayed exactly flat.
One brand spent $25,000 on DSP over 3 months. The dashboard reported $120,000 in attributed revenue. They paused in month 4. Revenue that month: $41,000, exactly the same as their previous 3-month average. DSP was not driving $40,000 per month. It was crediting itself for $40,000 that was happening organically regardless.
Why Did Last-Click Attribution Fail for DSP?
Last-click attribution works for bottom-funnel tactics like Sponsored Products, where a shopper searches a term, clicks an ad, and buys immediately. The causal chain is direct. DSP operates at different funnel stages: awareness, consideration, and retargeting people who are not yet ready to buy. Last-click attribution cannot capture that value accurately, and for DSP, it systematically distorted it upward.
The old model gave DSP full credit for any sale following an ad exposure, even when that exposure had no influence on the purchase. A shopper watches a streaming TV ad on Fire TV. They do not click anything. Three weeks later they search a generic category keyword, find the listing in organic results, and buy. Old DSP attribution counted the full order value as DSP-attributed revenue. Whether the ad influenced the decision was unknowable under that model.
How Bad Was the Double-Counting Problem?
Industry estimates from 2024 put non-incremental DSP-attributed sales at 20 to 40 percent of total credited revenue. Turn DSP off, and those sales would still happen organically. Retargeting amplified the problem further. Retargeting audiences of listing visitors captured buyers who were already in-market and likely to convert. DSP took credit for 100 percent of those conversions, even though the retargeting ad had changed almost nothing.
Sellers posted screenshots in 2024 showing hundreds of thousands in DSP-attributed sales where it was the only paid touchpoint. Holdout tests comparing an exposed group against a suppressed group showed the actual difference was a fraction of what the dashboard claimed. DSP was getting credit for being present, not for causing the outcome.
What Changed in Amazon DSP Attribution for 2025 and 2026?
Amazon rebuilt the entire DSP attribution framework from the ground up in 2025, not an adjustment to the existing model but a full rebuild introducing 3 major upgrades: multi-touch attribution with AI modeling, extended lookback windows of up to 5 years, and randomized controlled trials that prove incremental lift against a matched control group.
How Does Multi-Touch Attribution Fix Phantom Conversions?
Multi-touch attribution launched at Amazon Accelerate 2025 and distributes credit across every touchpoint in the customer journey rather than awarding it all to the last interaction. If a shopper saw a DSP ad, then clicked a Sponsored Product, then returned through organic search and purchased, each touchpoint receives partial credit weighted by its actual influence on the sale.
Phantom conversions disappear under this model because DSP cannot claim 100 percent credit. It receives credit proportional to its measured contribution to the journey. A DSP impression that preceded an organic conversion by 3 weeks gets weighted differently from a DSP click that preceded a purchase by 10 minutes.
Why Do Extended Lookback Windows Matter?
Amazon Marketing Cloud now supports up to 5 years of data for long purchase cycles. For replenishment products, 25 months of ad traffic data reveals year-over-year patterns that short lookback windows missed entirely. The extended window makes it possible to distinguish between DSP driving a genuinely new customer and DSP intercepting someone who had already bought from the brand multiple times and would have repurchased regardless.
A first interaction with a DSP ad followed by a conversion 90 days later represents real top-of-funnel value. A 6th purchase from an existing customer who happened to see a retargeting ad is something different. The new model tracks that distinction across the full customer history.
What Are AI-Driven Incrementality Tests?
Amazon Marketing Cloud now includes custom machine learning models that run propensity scoring and churn prediction, enabling audience segmentation by likelihood to convert. Brands can build separate audiences of high-propensity and low-propensity shoppers and measure whether DSP moves the needle differently across those groups.
Randomized controlled trials are now built into the framework. Group A sees DSP ads. Group B does not. After 30 or 60 days, the conversion rate difference between the 2 groups is measured. That is not attribution modeling. That is causal proof of lift. H&R Block ran the new attribution model in late 2025 and recorded 144 percent conversion lift on campaigns that were genuinely introducing new customers. Campaigns retargeting existing customers showed significantly lower incremental lift. They reallocated budget accordingly. The old model would have credited both campaign types equally.
How Does the New System Prove Incremental Lift?
The new framework measures what happens with DSP running against what happens without it, then credits only the difference. Two matched audiences are built in AMC with similar demographics, purchase history, and behaviour. Group A sees DSP ads. Group B does not. After 30 to 60 days, if Group A converts at 8 percent and Group B converts at 5 percent, DSP’s incremental lift is 3 percentage points. That figure is measurable, reproducible, and directly attributable to the ads.
What Is the Difference Between Old and New Attribution Logic?
The old model concluded that an ad worked because someone saw it and later bought. The new model concludes an ad worked because people who saw it bought at a measurably higher rate than people who did not. That distinction is the difference between correlation and causation.
With 5-year lookback data, the system can also prove whether DSP is acquiring new-to-brand customers or accelerating repeat purchases from existing ones. Suppress DSP for a control group for 90 days. If that group’s purchase rate stays flat while the exposed group’s rises, the long-term brand-building impact is quantified, not just asserted.
What Do Real Incrementality Test Results Look Like?
A pet supplement brand ran a holdout test in Q4 2025. Half their target audience saw DSP ads. Half did not. After 60 days, the exposed group showed 22 percent higher new-to-brand conversion rate and 18 percent higher repeat purchase rate than the control group. Total incremental revenue attributable to DSP: $47,000 over 2 months on $12,000 in spend. A real 4x ROAS, verified against a control, not reported by a dashboard.
When the brand paused DSP in month 3, the previously exposed group’s conversion rate declined toward the control group’s baseline over 30 days. That trajectory is the evidence. The ads were driving behaviour, not just claiming credit for it.
Do Brands See Real Drops When They Pause DSP Now?
The rebuilt framework passes the pause test. Turn DSP off and incremental sales drop. Turn it back on and incremental sales rise. Before mid-2024, pausing DSP changed nothing. Revenue held steady. That was the evidence attribution was broken. The pause test now shows a different result.
How Did the Same Brand Get Different Results in 2024 vs. 2025?
A brand ran DSP in early 2024: $15,000 over 2 months, dashboard showed $65,000 attributed. They paused it. Sales held at $32,000 per month, unchanged from before. They concluded DSP did not work.
They returned in late 2025 with the new multi-touch attribution and AMC incrementality testing active. Same brand, same category: $18,000 over 3 months, dashboard showed $52,000 attributed. Lower than the 2024 number, but more accurate. They ran a holdout test simultaneously, suppressing DSP for 30 percent of their target audience. Exposed group conversion rate: 11 percent. Suppressed group: 8.5 percent. That 2.5 percentage point gap translated to $13,000 in incremental monthly revenue directly caused by DSP. When they paused in month 4, the exposed group’s rate dropped 2 points over 30 days. The suppressed group stayed flat.
How Are Brands Using Propensity Scoring to Improve DSP Efficiency?
A kitchen gadget brand that quit DSP in 2023 after seeing no lift on pause returned in Q4 2025. Using AMC’s custom ML models, they scored their audience by propensity to convert and found DSP was driving 18 percent lift among low-propensity shoppers who were not already planning to buy, but only 3 percent lift among high-propensity shoppers who would likely have converted anyway.
The old attribution model credited both groups equally. The new model let them reallocate budget away from retargeting existing buyers and toward cold traffic acquisition. DSP efficiency improved 35 percent. When they paused, they recorded a measurable drop in new-to-brand conversions for the first time.
Is Amazon DSP Worth Reconsidering in 2026?
Abandoning DSP in 2023 or 2024 because the attribution was inflating results was a rational response to a real problem. Avoiding DSP in 2026 for the same reasons means avoiding a problem that no longer exists in the rebuilt system.
What Does a Proper DSP Test Look Like Now?
A 60-day test with a control group built in from day one gives a clear answer within 2 months. If DSP is not driving incremental sales, the exposed and control groups will perform the same. Pause it, lose nothing, move on. If it is driving incremental sales, the gap between the 2 groups will be visible by day 30. Scale what produces the lift. Suppress what does not.
Because the new framework connects to AMC, the full customer journey is visible: whether DSP is bringing in new-to-brand customers or accelerating repurchases, which creatives and placements are moving the needle, and which audiences are worth the spend. The information needed to make that decision is now available and verifiable.
What Are Brands Seeing With the New Attribution Framework?
Industry data from late 2025 shows brands using the new attribution framework recording 30 to 40 percent efficiency gains, 144 percent conversion lifts in some verticals, and 35 percent reductions in cost per acquisition. All figures measured against control groups, not self-reported by dashboards. Amazon DSP ad revenue grew 24 percent year-over-year in 2025, according to Amazon’s earnings data.
Sellers spending six figures monthly on Sponsored Products and Sponsored Brands without using DSP are leaving incremental growth on the table based on information that is now 2 years out of date. The attribution model that broke the platform’s credibility was rebuilt. The holdout test infrastructure that proves real lift now exists inside AMC. The data is available to verify the claim before committing meaningful budget.
