First-Touch vs Last-Touch Attribution: What Each One Hides

Two models, one shared flaw

Last-touch attribution is the default across almost every native ad platform report, including Amazon's — Sponsored Products, Sponsored Brands and Sponsored Display all credit the ad interaction closest to the sale. First-touch flips the credit to whichever ad or channel started the journey. Both models share the same underlying flaw: they collapse a multi-step decision into a single point and ignore everything else that happened along the way. The only difference is which single point they pick.

That shared flaw matters because most real Amazon paths involve more than one touch. A shopper sees a display ad, later searches the category, clicks a sponsored product, doesn't buy, comes back the next day through a branded search, and buys. Last-touch credits the final branded search. First-touch credits the original display impression. Neither credits the middle step — the non-branded sponsored product click that arguably did the actual persuading — at all.

The worked example

Take a $120 order with four attributed touches, in order: a DSP display impression, a non-branded Sponsored Products click, a Sponsored Brands click, and a branded search click that closes the sale. Under last-touch, the branded search campaign shows a $120 attributed sale on what may have been a near-guaranteed click from a shopper who'd already decided. Under first-touch, the DSP display line shows the full $120, and every campaign that ran in between — including the two clicks that arguably did more of the actual persuading — shows zero.

Run the same $120 through an equal-weight four-touch split instead: $30 to each. No single model here is correct in an absolute sense. What the comparison shows is how far apart the three answers land from the same underlying data — a genuinely useful signal on its own, because it tells you how sensitive your channel-mix decisions are to a modelling choice nobody in the account may have consciously made.

What last-touch structurally hides

Last-touch systematically undervalues upper-funnel and awareness activity — display, streaming, video, anything whose job is to start a journey rather than close it. In our own DSP book, streaming and connected-TV inventory returned a 0.77x ROAS on a last-touch basis — the worst return multiple on the whole report — while carrying the highest new-to-brand rate of any inventory type, 56.3%. Read on last-touch alone, that line item looks like the first one to cut. Read with the new-to-brand context attached, it's doing exactly the job it was bought for: introducing people to the brand who then convert somewhere else, later, through a touch that gets all the credit last-touch reporting hands it.

What first-touch structurally hides

First-touch has the opposite blind spot. It overvalues whatever is cheapest to serve early and most visible at the top of a funnel, and it structurally undervalues the touches that do the actual conversion work — retargeting, branded search, cart-abandonment campaigns — none of which get credit under a pure first-touch model no matter how directly they closed the sale. A brand that manages budget purely on first-touch data will tend to overfund broad-reach prospecting and underfund the retargeting and closing campaigns that convert the audience prospecting built, which is its own kind of expensive mistake.

The common mistake, including ours

The mistake is picking one model and treating its output as the account's true performance, rather than using both models — or better, several — as different lenses on the same underlying path. We've reported a first-touch DSP number to a client specifically because it made the display line look stronger during a budget-renewal conversation, without pairing it against the last-touch number the client's own finance team was already using internally. Both numbers were technically correct. Presenting only the one that favoured our own recommendation wasn't, and we don't do it as a single unpaired figure anymore — both go in the same table now, every time.

The same mistake shows up in reverse when a channel wants to be defunded. A stakeholder skeptical of DSP spend will often reach for the last-touch number specifically because it's the one that makes upper-funnel media look weakest, without asking whether that number was ever built to answer the question being asked of it. Neither direction is dishonest on its own — both models are real — but selecting whichever one supports a decision already made, rather than reporting both and letting the gap between them inform the decision, is the pattern worth catching in your own team as much as anyone else's.

A quick sanity check before you trust either number

Before making a budget call off either model, check the attribution window each touch fell inside. Sponsored Products uses a 7-day click window for sellers (14 for vendors); Sponsored Brands and Sponsored Display use 14 days; DSP's offsite exchange inventory has historically run on the same 14-day click and view structure, while Amazon's January 2026 update replaced the blanket 14-day view window for Sponsored Brands and Sponsored Display vCPM campaigns, and Amazon DSP's in-store placements, with a shorter, machine-judged model. A path that spans longer than the shortest window in play will have touches quietly drop out of one channel's report before they drop out of another's — which can make a first-touch or last-touch comparison look inconsistent for a reason that has nothing to do with the model itself.

What to actually do instead

Report both first-touch and last-touch side by side for any channel where the gap between them is large — that gap itself is the useful signal, since it tells you how much of a channel's real contribution is invisible under whichever model your dashboard defaults to. Where the two disagree sharply, that's exactly where an incrementality test — a holdout or a geo-lift split — earns its cost, because it answers the causal question neither single-touch model can: not who gets credit, but whether the ad changed the outcome at all.