The Amazon DSP creative best practices that pay off come from measurement, not the spec sheet: compare creatives only within one objective bucket, judge them on rates rather than totals, and count every copy of an asset as one creative.
Creative performance on Amazon DSP is measurable, unequally distributed, and almost never examined, because everything about how DSP buying works points attention at orders and line items instead of the creatives running inside them. The console answers "did the flight deliver," budgets live on flights, pacing lives on flights, and the creative, the thing the viewer actually experiences, reports as an afterthought. This guide covers what creative-level performance means on DSP, which metrics exist for it, how to measure it manually from DSP reports, and the two structural problems (creative identity and cross-platform comparison) that make the manual version fight you.
The premise worth stating plainly: on a platform where targeting, bidding, and inventory access are increasingly automated, the creative is one of the few inputs you still control. Two video assets running through identical line items can deliver meaningfully different results, and if you never look, you keep paying for the difference without collecting the lesson.
Why does nobody measure DSP creative performance?
Three structural reasons, none of them good excuses.
The buying model hides it. DSP runs on delivery economics: flights, budgets, CPM goals, pacing. Success is defined at the order level, so reporting attention follows the order. A flight can pace perfectly at 100 percent while one of its three creatives quietly outperforms the other two on every engagement and conversion metric.
The volume is low, so testing instincts never engage. Paid social teams test dozens of creatives monthly and build measurement habits around the volume. A DSP advertiser might run three to ten creatives per quarter, which feels too few to "test." But fewer creatives make each one matter more, not less: a bad creative on DSP absorbs a bigger share of budget for longer than a bad creative in a social rotation ever could.
The reports make you work for it. Creative-level data exists in DSP reporting, but assembling a clean creative view means pulling reports in 31-day chunks against 95 days of retention on the older v3 reports, then reconciling creatives that run across multiple line items and orders. It is doable, which is the point of the next sections, and tedious, which is why it rarely happens.
What does creative performance mean on DSP?
Not what it means on Meta, and importing paid-social scoring wholesale is the classic mistake. The right frame depends on what the line item is buying:
For awareness and CTV buying (most DSP spend, including Prime Video and streaming placements), creative performance is delivery-quality performance: viewability (viewable over measurable impressions), video completion behavior, CTR where clicks are even possible, and effective CPM against comparable inventory. A CTV spot has no add-to-cart button; judging it on ROAS is a category error, the same one covered from the other side in creative-level ROAS.
For retail-endemic conversion lines, DSP reports attributed outcomes on its last-touch click and view model: purchases, detail page views, add-to-carts, sales, and new-to-brand purchases and sales, which is DSP's genuinely distinctive metric, since new-to-brand tells you whether a creative recruits new customers or reconverts existing ones. Two creatives with identical purchase counts and very different new-to-brand shares are doing different jobs for the business.
The single most important discipline: compare creatives within the same objective bucket only. An awareness CTV asset against a conversion display asset produces a meaningless comparison in both directions. Awareness spend judged on conversion metrics reads as waste when it is doing exactly its job.
What creative types does Amazon DSP run?
Four families: display, online video, streaming TV, and audio, across Amazon's own properties and third-party publishers. Amazon's ad specs list desktop and mobile web display, mobile app banners, online video (in-stream and out-stream, 16:9 or 9:16, 6 to 120 seconds), streaming TV, and audio, plus Amazon's e-commerce display creatives.
The e-commerce display format is responsive e-commerce creative (REC). You submit one creative and Amazon generates it into every supported size, for Amazon, third-party sites, and devices such as Echo Show and Fire TV. Each REC ad pairs your image or video with an e-commerce section that Amazon builds from your product listing: live price, deal badges, coupons, star ratings, and Prime eligibility. One creative can carry up to 20 ASINs, and it rotates products and skips out-of-stock items on its own. Video inside REC is US only, on third-party publisher sites.
That split matters for measurement. In a REC ad, Amazon writes the e-commerce half, so two REC creatives differ mainly in the image or video you supplied and the ASINs behind them. Judge them on that, within the conversion bucket, and expect product price and deal changes to move results as much as the creative does.
One naming note: Amazon now describes Amazon DSP as part of Amazon Ads Agent, with its programmatic settings living in DVA+ campaigns (currently for a limited set of US advertisers). Advertisers in other countries who used Amazon DSP still see the Amazon DSP experience, and the creative types above are the same either way.
How does Amazon DSP credit a conversion to a creative?
By last touch: Amazon credits each conversion to the most relevant click or viewable impression before it, 14 days by default for clicks and views. Since January 1, 2026, views on DSP ads served in Amazon's store use a shorter, signal-based window, and campaigns with flexible look-back use advertiser-selected windows. Views typically make up most DSP conversions, so a creative's attributed results lean heavily on viewable impressions. Amazon DSP attribution metrics covers the rules in full.
How do you measure it manually?
The workflow, from DSP reports to a defensible creative view:
- Pull creative-level reports for your window, and archive every pull immediately. The older v3 DSP reports serve 31 days per request and 95 days of history, so any history you do not save is history you lose. Label every conversion column with its window and model (standard or all views) and never blend them with other platforms' windows.
- Reconcile creatives across line items and orders. The same asset typically runs in several line items; sum its impressions, spend, and outcomes into one row per creative. This is the step that breaks, covered below.
- Bucket by objective before comparing. Awareness and CTV lines in one pool, conversion lines in another, per the framing above.
- Compute per-creative rates, not totals: viewability rate, completion behavior, CTR, CPM, and for conversion lines, cost per attributed purchase and new-to-brand share. Totals reward whichever creative got the most budget; rates reveal which earned it.
- Apply a readability floor (impressions for delivery metrics, conversions for outcome metrics) and mark everything under it undecided, the same discipline as any creative verdict.
Run monthly, this produces the thing almost no DSP advertiser has: a ranked answer to "which creatives should the next flight run," grounded in delivery history rather than whoever produced the most recent asset.
Why does the manual version break?
The identity problem. DSP creative reporting keys on creative IDs, and the same underlying asset frequently exists as multiple creatives: re-uploaded for a new campaign, resized for a different placement, renamed by a different team member. Your best-performing video can be split across three creative IDs, none of which individually clears your readability floor, so your actual winner never surfaces. The fix requires recognizing that two creative IDs contain the same bytes, which a spreadsheet cannot do.
The cross-platform blindness. The question that matters most to a brand running DSP alongside paid social ("does our hero video perform on Amazon the way it performs on Meta?") is unanswerable from inside either platform, because neither knows the other's creatives exist. DSP tells you its numbers; nothing connects them to the same asset's life elsewhere.
The persistence problem. Ninety-five days of retention on the older reports means the manual practice has no institutional memory unless someone maintains the archive forever, and the analysis quality is capped by the discipline of whoever pulls the reports.
The instrumented version
This is the gap Peachblue was built into: it is the only creative analytics platform for agencies at self-serve pricing that covers Amazon DSP. Enterprise creative analytics for DSP exists at enterprise prices; below that tier, the category simply had no product, which is also why no content about DSP creative performance exists beyond ad-spec pages. For a tool-by-tool comparison, from Amazon Marketing Cloud to Vidmob, see Amazon DSP creative analytics tools compared.
What the instrumented version does with the problems above: creative assets are mirrored from Amazon's Creative Asset Library and matched by content, with byte-identical assets collapsing to one creative regardless of how many creative IDs, line items, or re-uploads they run through, which dissolves the identity problem at the root. The same fingerprinting matches creatives across platforms, so the Meta-versus-DSP question about your hero asset becomes answerable, in both the app and conversationally through Claude. Every creative gets the 31-dimension analysis and scoring, with DSP and awareness delivery bucketed separately from conversion spend, so CTV assets are judged on delivery quality and never miscounted as waste. History accumulates from the day you connect and stays yours whatever Amazon's reports retain, which is the day-one argument for new DSP advertisers in miniature: your creative learnings compound only if something keeps them. And for agencies, all of it scopes per client.
Start with one manual pull either way. Rank last quarter's creatives within their objective buckets and look at the spread; on most accounts, the gap between the best and worst creative is the largest unexamined number in the whole program. The flights were pacing fine the entire time.