Which Attribution Model Suits an Amazon-Led Business

Start from what Amazon already gives you by default

Every Amazon ad product reports last-touch or last-view by default — Sponsored Products, Sponsored Brands, Sponsored Display and DSP each credit the interaction closest to the sale, using their own separate attribution windows (7 days for Sponsored Products sellers, 14 days for vendors; 14 days for Sponsored Brands and Sponsored Display click attribution; Amazon's January 2026 update replaced the blanket 14-day view window for vCPM Sponsored Brands, Sponsored Display and DSP in-store placements with a shorter, machine-judged model). That's your baseline whether you choose it or not, and it's worth knowing precisely what it is before deciding whether to layer anything on top.

A decision framework, not a single answer

Match the model to what you actually know about your funnel. If you run a mostly single-channel account with short, simple paths, last-touch is close enough to the truth that adding complexity mostly adds noise. If you run DSP and sponsored ads together with genuinely multi-step paths, and you don't yet have strong evidence about which touches matter more, start with linear or equal-weight attribution as an honest baseline. If your category has a clear awareness-consideration-purchase structure — which many Amazon-led brands running streaming or DSP display alongside sponsored ads do — position-based attribution, calibrated with your own first-touch data, tends to fit the funnel shape best. Only move to data-driven attribution once you have the conversion volume to train it reliably, which for most single-brand accounts is a later-stage decision, not a starting one.

Why the model choice can't be the whole answer

Whichever model you pick still only describes how credit is split across touches that already happened — none of them prove a touch caused a sale rather than simply preceding it. That's the gap incrementality testing exists to close, and it's the reason the strongest measurement setups pair a rule-based or data-driven attribution model for routine reporting with a periodic holdout or geo-lift test to validate whether the attributed pattern actually reflects incremental lift. In our own DSP conversations, "is this incremental, or would they have bought anyway" is the single most common objection we hear — and the honest response is never a better attribution model; it's naming who designs and computes the test, and writing that design into the scope of work before spend starts.

The common mistake, including ours

The mistake is treating the choice of attribution model as if it were the whole measurement strategy, rather than one input alongside incrementality testing. We've walked into a client relationship where the prior agency's entire measurement pitch was "we use data-driven attribution, so our numbers are more accurate" — a claim that sounded rigorous and skipped past the actual question, which is whether any credit-splitting model, however sophisticated, tells you whether the spend caused the sale. It doesn't, on its own. We've made a milder version of the same mistake ourselves, leading a client report with an attribution-model comparison before we'd built the holdout test that would have actually answered their real question.

The pattern repeats often enough to be worth naming plainly: a more sophisticated-sounding attribution model is frequently offered as the answer to a client's incrementality question, because it's easier to build and demo than a genuine holdout test, and it lets both sides avoid the harder, slower conversation about test design, statistical power and what happens if the result comes back disappointing. A vendor who leads with model sophistication rather than test design is answering a question the client didn't quite ask.

A practical starting setup

For most Amazon-led businesses starting from scratch: report last-touch as the baseline everyone already understands, add a position-based or linear cross-channel view inside Amazon Marketing Cloud once you're running DSP and sponsored ads together, and schedule at least one incrementality test — a geo-lift or holdout — per year on your largest spend line, more often if budget allows. That combination costs less to set up than it sounds, and it answers both questions that matter: which touches get credit, and whether the spend is actually working.

Write the plan down before spend starts, not after the first report looks disappointing. Naming the model you'll use for routine reporting, the cadence for the incrementality check, and what you'll do if the two disagree — all agreed in advance — removes most of the argument that otherwise happens after the fact, when a bad number arrives and everyone reaches for whichever methodology defends their prior position.

When the setup above gives you conflicting signals

If your chosen attribution model and your incrementality test disagree — the model says a channel is contributing well, the holdout shows little to no lift — trust the holdout. A well-designed incrementality test answers the causal question directly; an attribution model only ever answers the credit-splitting question, however well-chosen. The right response isn't to discard the attribution model entirely, since it's still useful for day-to-day optimisation signal, but to treat its output for that specific channel with more scepticism, and to re-run the test before making a large budget change either way.

The same logic runs in reverse, and it's the direction people trust less instinctively: if the holdout shows real lift on a channel your attribution model has been consistently undercrediting — streaming and connected-TV inventory is the most common example we see, given how structurally it gets undervalued by last-touch and short-half-life time-decay models alike — that's a reason to protect the channel's budget even while its attributed numbers look weak, not a reason to wait for the dashboard to agree first.