Amazon Marketing Data, Explained by Someone Who Reconciles It Weekly

What 'Amazon marketing data' actually covers

There is no single dataset called Amazon marketing data. There are four separate systems, built at different times for different jobs, and most confusion in this category comes from treating them as one thing.

• Sponsored ads reports — clicks, spend, and conversions from Sponsored Products, Sponsored Brands, and Sponsored Display, inside a click-attribution window you set.
• DSP reports — impressions, clicks, and conversions from programmatic display, video, and audio bought through Amazon's demand-side platform.
• Brand Analytics / Search Query Performance — search term and click share data, aggregated at the category level, with no cost or ROAS attached.
• Amazon Marketing Cloud (AMC) — a pseudonymized, event-level log that lets you join sponsored ads and DSP signals together and query the overlap directly.

Each one answers a narrower question than people assume. Sponsored ads reports tell you what happened inside search. DSP reports tell you what happened inside display and video. Neither tells you what happened when a shopper touched both — that's AMC's job, and it's the only one of the four built to answer it.

A worked example: where the double-counting hides

Say a shopper sees a display ad, doesn't click, browses later, sees a sponsored product ad, clicks, and buys. In the DSP report, that's a view-through conversion. In the sponsored ads report, that's a click-attributed sale. Add the two reports together and you've counted one purchase twice — and your blended ROAS now looks better than the business actually performed.

This is exactly what AMC is for: joining the event-level logs so the same purchase can't be claimed by two channels at once. We hold 109 live Amazon DSP advertiser seats and run that join as standard.

What it produces is worth showing as a chain rather than a headline, because the chain is the part that can be audited. Take 30 of those advertisers over July 2026. Impressions bought: 78.4 million, at a $4.00 CPM — multiply those out and the media behind them is about $313,600. Clicks off that inventory came in at a blended $1.42, which puts the click count near 220,000. Attributed purchases: 57,137, at a blended $5.49 each — multiply those back out and you land within a rounding error of the same $313,600 you started from. That reconciliation is the point. If a report's cost-per-acquisition and its purchase count don't multiply back to its spend, the two figures were pulled from different windows and one of them is wrong.

Return across that book came out at 6.04x, stated for the whole set of 30 rather than the strongest line item in it, and 20.1% of those purchases were from shoppers new to the brand. That last figure is the one that makes the first credible: a high return built entirely on shoppers who were already buying tells you the reporting found them, not that the media created them.

The mistake that ruins most of this data: treating last-click as proof

Last-click attribution tells you which ad was touched last before a sale. It does not tell you whether that ad caused the sale, or whether the shopper would have bought anyway. This distinction sounds academic until you're deciding whether to keep spending on display, and the click report says display is barely converting — because display's job is usually upper-funnel, and last-click structurally undercounts it.

We've made this mistake ourselves: early on, we reported DSP view-through conversions without checking whether the same shopper had already converted through sponsored ads. It inflated the blended ROAS and nobody caught it until spend kept climbing while overall brand sales didn't move at the same rate.

The fix isn't a better attribution model — no attribution model, no matter how sophisticated, can prove incrementality on its own. The fix is a holdout: hold a matched group of shoppers or geographies out of the campaign, run the campaign against everyone else, and compare the two. That's causal evidence. A cleaner dashboard is not.

When the numbers look wrong, check these first

If your Amazon marketing data doesn't add up, it's usually one of these, in order of how often we actually find them:

• Attribution window mismatch — sponsored ads and DSP reports can be set to different lookback windows (1-day, 7-day, 14-day). Compare like windows before you compare numbers.
• New-to-brand definition — this is based on a trailing purchase history window, not a single order. Confirm the window before quoting the percentage anywhere near a board deck.
• Channel-specific figures presented as blended — a CPC, CPM, or CPA from one placement type (video, audio, one ad format) quoted as if it applies to the whole account.
• UI vs. API discrepancy — the dashboard and the pulled report sometimes reflect different processing timestamps. Re-pull before assuming the dashboard is wrong.
• Double-counted conversions — the DSP-plus-sponsored-ads overlap above. If your blended ROAS jumped without a spend or creative change, check for this first.

If you've checked all five and the number still looks wrong, it probably is wrong — go back to the raw event logs in AMC rather than trusting either summary report.

Four sources, four different blind spots

The table below is the version of this we wish someone had handed us years ago — what each source actually measures, what it's genuinely good for, and where it quietly lies to you.