Marketplace management
Amazon DSP: Getting Beyond Last-Touch Reporting
The distinction matters because brands pay for incremental contribution—not for credit inside a reporting window.
Amazon describes DSP as a demand-side platform for buying display, video and audio inventory programmatically across multiple exchanges. Its own measurement documentation also distinguishes how impressions, clicks, viewability, traffic validation and reporting are handled. Those are important measurement controls, but they do not on their own answer the causal question of incrementality. Amazon’s DSP methodology is available here. Amazon’s DSP overview is here.
Attribution answers “what happened”; incrementality asks “what changed because of it”
Attribution assigns credit under a set of rules. Last-touch reporting gives full credit to the final recorded interaction. Multi-touch reporting distributes or presents credit across interactions. Both are useful for understanding paths and optimizing delivery.
Incrementality asks a different counterfactual: How many outcomes would not have happened if the DSP activity had not run? No path report can directly observe that alternate reality. It must be estimated through experimental design or through explicit, conservative assumptions.
Read path data as evidence, not a verdict
Path analysis can still improve decisions when it is interpreted correctly. Categorize conversions into a small number of path types:
| Path type | What it can suggest | Incrementality confidence |
|---|---|---|
| DSP only | DSP was the recorded exposure path | Higher, but not automatically causal |
| DSP final touch | DSP was close to the recorded conversion | Moderate to higher; assess prior demand signals |
| DSP assist with paid search | DSP may have influenced consideration | Moderate; search may capture existing intent |
| DSP assist with organic/direct | DSP may have built or reactivated demand | Moderate; use controls where possible |
| Non-DSP conversion | No DSP path was recorded | Not proof that DSP had no effect |
A conservative model can apply different estimated incremental rates to each path type. For example, a brand might assign a higher rate to DSP-only conversions, a lower rate to mixed paid-search paths, and zero incremental credit to paths that cannot be reliably linked. These are assumptions, not observed facts; they should be documented, challenged and updated when tested.
Build a contribution-margin model
Revenue is not the right success measure. Start with path-level sales, apply the incrementality assumption, then apply the contribution margin available after variable costs.
Estimated incremental contribution = attributed sales × assumed incremental rate × contribution margin
Net DSP contribution = estimated incremental contribution − DSP spend − incremental creative/technology cost
Illustrative example only: if a campaign reports $300,000 in attributed sales, a path-weighting model estimates that 25% is incremental, and the applicable contribution margin is 30%, estimated incremental contribution is $22,500. If spend is $20,000, the estimated net contribution before incremental operating costs is $2,500. The same campaign can look excellent on attributed ROAS yet be only marginally profitable under an incremental-contribution view.
The margin input should reflect the actual SKU mix and include variable Amazon economics. Amazon separates selling-plan fees, referral fees and optional service costs; its FBA calculator includes common fees such as referral fees, fulfillment costs, monthly storage and variable closing fees, while noting that other items may need to be added. Amazon’s seller pricing overview and calculator guidance are useful starting points.
The strongest answer: controlled measurement
When spend is material enough to justify the effort, design a holdout or matched-market test. The principle is simple: create a comparable group that is intentionally not exposed to the DSP tactic, then compare outcomes over a sufficient window while controlling other major changes.
Possible approaches include:
- Audience holdouts: Randomly suppress a share of an eligible audience from DSP exposure.
- Geo tests: Hold out comparable geographic areas where targeting and demand conditions allow.
- Time-based tests: Alternate defined on/off periods only when seasonality and carryover effects are manageable.
- Matched cohorts: Compare carefully matched audience cohorts when randomization is not possible; treat results as directional.
The design should state the primary outcome (for example, new-to-brand purchases or total contribution), the measurement window, exclusions, budget stability, inventory requirements and the decision rule before launch. Do not change pricing, promotions, creative and targeting simultaneously if the goal is to learn causality.
A practical DSP measurement cadence
- Define the job of DSP. Prospecting, retargeting, launch support, retail-event reach, audience reactivation or off-Amazon reach require different expectations.
- Set guardrails. Establish spend, frequency, inventory, margin and brand-safety limits.
- Read delivery and path data weekly. Check reach, frequency, supply, audience overlap, path composition and cost trends.
- Recalculate the economic view monthly. Use contribution margin and a documented incrementality assumption.
- Run a controlled test periodically. Use the result to revise the assumption and funding decision.
Common measurement mistakes
- Treating all DSP-attributed sales as incremental.
- Using revenue instead of contribution margin.
- Comparing one campaign’s ROAS to another without considering each tactic’s role.
- Calling a before/after change causal while price, inventory, promotions or search spend also changed.
- Expanding spend before the measurement plan is agreed.
DSP can be a valuable part of marketplace growth. The disciplined approach is not to reject attribution—it is to put attribution in its proper place: as operational evidence that informs a causal measurement plan and an economic decision.
