Attribution in Advertising: How It Works, and What It Can't Prove

What attribution in advertising actually means

Attribution is the method for deciding which ad touchpoint gets credit when a shopper converts. A shopper might see a display ad on Monday, click a sponsored product ad on Wednesday, and buy on Friday after searching your brand name directly. Attribution is the rule set that decides which of those three moments — or all three — gets the credit for the sale.

Every ad platform ships a default rule. Amazon Ads attributes sponsored ad sales to the last ad clicked within its attribution window. Meta attributes within its own click and view windows. None of these defaults talk to each other, which is the first thing that goes wrong: the same sale can get claimed, in full, by more than one dashboard at once.

Amazon Attribution and Amazon Marketing Cloud (AMC) exist to close part of that gap. Amazon Attribution tags non-Amazon marketing — social, search, email, display bought outside Amazon — so you can see its effect on Amazon shopping activity. AMC is the clean room where that data, plus DSP and sponsored ads event-level data, gets joined so you can check whether two of your own channels are claiming credit for the same shopper.

The attribution models, and what each one hides

There are two broad families: single-touch models, which give all the credit to one moment, and multi-touch models, which split it across several. Neither family proves causation — each describes a pattern, and every pattern-matching model has a specific blind spot.

The last row in the table below isn't a model in the same sense as the other six. It answers a different question — not "which touchpoint deserves credit" but "would this sale have happened anyway" — and that distinction is what the rest of this page turns on.

A worked example: the same purchase, three different answers

Say a shopper sees a display impression on Monday, clicks a sponsored product ad on Wednesday, and buys after a branded search on Friday. The order is $50.

• Last-touch: the branded search gets 100% of the credit. Display and sponsored ads show $0 attributed revenue for that sale, even though one of them likely started the path.
• First-touch: the display impression gets 100% of the credit. The sponsored click and the branded search — arguably the two moments closest to the purchase decision — get nothing.
• Linear: each of the three touchpoints gets roughly $16.67. Fair-looking, and just as much of a guess as the other two.

All three answers come from the exact same shopper, the exact same purchase, the exact same three touchpoints. None tell you whether the shopper would have searched your brand and bought anyway, with or without the display ad. That question needs a different tool — a holdout, where a matched group of shoppers never sees the display ad, so you can compare what happened to them against what happened to the group that did.

Why Facebook ads attribution and Amazon's numbers don't match

Facebook (Meta) attributes within its own click and view windows, using its own pixel or Conversions API signal. Amazon attributes within its own window, using its own event log. There is no shared identifier connecting the two, so a shopper who clicked a Facebook ad and later bought on Amazon can get counted as a Facebook conversion, an Amazon conversion, both, or neither — depending on the windows, the consent given, and whether Amazon Attribution tags were placed on the Facebook creative in the first place.

Mobile advertising attribution adds another layer of friction. App install attribution runs through mobile measurement partners using device-level SDK signals, not browser cookies. Since Apple's App Tracking Transparency change, a large share of iOS users decline device-level tracking, which pushed most mobile attribution toward probabilistic and aggregated reporting instead of one-to-one matching. The practical result: mobile numbers you see today are built to a different, generally lower-confidence standard than desktop web numbers from the same campaign, even inside one platform.

None of this means the numbers are wrong. It means they answer slightly different questions, measured with different rulers, and adding them up across platforms will overcount the truth every time.

Attribution proves correlation. Incrementality proves the ad worked.

This is the honest core of the problem: last-click and multi-touch attribution measure who was present near a sale. They cannot measure whether the ad caused the sale. A shopper who was already going to buy, and who also happened to see your display ad, shows up in attributed reporting looking identical to a shopper the ad genuinely persuaded. Only a holdout or matched-control test — comparing an exposed group against a similar unexposed group — separates the two.

Attributed numbers are still a useful starting point. Across 30 of Full Circle and reMKTR's Amazon DSP advertisers in July 2026, the book delivered 6.04x return on ad spend, 78.4 million impressions at a $4.00 CPM and a blended $1.42 cost-per-click, with a blended cost per acquisition of $5.49 across 57,137 attributed purchases, 20.1% of them from shoppers new to the brand. Those are attributed figures. The honest next step is reconciling them in AMC, where DSP and sponsored ads stop double-counting the same shopper, then running a holdout to confirm the display spend is adding sales rather than claiming credit for ones sponsored ads or organic search would have closed anyway.

The common mistakes — including one we've made

The most common mistake is treating a platform's default attribution window as ground truth and scaling budget against it. A campaign showing strong attributed ROAS on a seven-day click window can still be doing nothing incremental — the shoppers were converting anyway, on a different channel's dime.

The mistake we've made ourselves: reading a drop in attributed display conversions as proof the channel stopped working, and cutting it, when the real cause was a tracking or model change upstream — not an actual drop in sales. The fix is procedural, not clever: check whether the number moved because behavior changed or because measurement changed, before touching budget. That's why every change we run now carries three things before it goes live — the evidence behind it, a measurement plan, and a rollback trigger — so a bad attribution read doesn't get mistaken for a bad result.

The second common mistake is running sponsored ads and DSP without ever reconciling them in AMC. Both channels can independently report the same purchase as attributed revenue, which inflates total attributed sales above what the business actually sold — a problem that only surfaces once you join the event logs.

Where Dr. DSP fits

Dr. DSP fits at the point where attribution stops being enough. It's Amazon DSP — Amazon's demand-side platform for programmatic display, video and audio, not the Delivery Service Partner courier franchise — run as a managed product by Fable 5, out of Full Circle, which has managed more than $500M in revenue across 100+ brands. The argument isn't that attribution is wrong to use — it's that last-click and multi-touch models can't prove a display ad caused a sale, and holdouts and matched controls can. Every DSP campaign gets reconciled in AMC so DSP and sponsored ads stop double-counting each other, and every change carries a measurement plan and a rollback trigger before it runs. There's no published price — a demo, the first 30 days free, and pricing set on the call against real budget and scope, with Orbit included at no extra cost. If you're reading attribution off dashboard numbers alone, this is the gap it closes.