Position-Based (U-Shaped) Attribution Explained

The logic behind the shape

The model encodes a specific, testable belief about a customer journey: that the moment someone first discovers a brand and the moment they decide to buy both matter more than the touches in between, which are treated as supporting the decision rather than driving it. Graphed as a bar chart of credit by touch position, the two ends are tall and the middle sags — hence "U-shaped." It's a genuine compromise between first-touch and last-touch rather than a rejection of either: it keeps both, and simply stops pretending the middle didn't happen.

There's also a variant worth knowing about, W-shaped attribution, which adds a third weighted point — typically the moment a lead or an equivalent mid-funnel milestone occurs — and is more common in B2B funnels with a defined qualification stage than in ecommerce, where no clean equivalent milestone usually exists between first exposure and purchase. For most Amazon-led brands, the standard two-anchor U-shape is the more natural fit, since the funnel rarely has a distinct, trackable middle event worth anchoring a third weight to.

A worked example

Take a $180 order with five touches: a streaming ad, a DSP display impression, a non-branded Sponsored Products click, a Sponsored Brands click, and a branded search click that closes the sale. Under the standard 40/20/40 split, the streaming ad (first touch) gets 40% of $180 = $72. The branded search click (last touch) gets another 40% = $72. The remaining 20%, $36, splits evenly across the three middle touches — the DSP display impression, the Sponsored Products click and the Sponsored Brands click — at $12 each.

Compare that to linear attribution's even $36 per touch across all five, and to last-touch's full $180 on branded search alone. Position-based sits between the two: it protects the first touch from being erased the way last-touch erases it, while still crediting the close more than the supporting middle touches — which is a genuinely different claim from either alternative, not just a blend of them.

When 40/20/40 is the wrong split

The 40/20/40 ratio is a convention, not a law — nothing about the model requires those exact numbers, and the right split depends on how your funnel actually behaves. A brand whose real decision-making happens almost entirely at the moment of comparison-shopping, with the initial discovery touch mattering comparatively little, might reasonably run something closer to 20/60/20, weighting the close far more heavily. A brand in a category with a long, deliberate consideration phase — where the middle touches are doing real work, not just filler between two decisive moments — might flatten the split toward something closer to 30/40/30, giving the middle more credit than the default convention allows.

Where this fits an Amazon-led business specifically

Position-based attribution maps unusually well onto a common Amazon pattern: a DSP display or streaming impression that starts awareness, several sponsored-ads touches in the middle that build consideration through search, and a final branded-search click that closes. That's a genuine three-act structure many Amazon paths follow, which is part of why this model, more than pure linear or pure time-decay, tends to match how Amazon marketers already intuitively think about their funnel — first exposure, mid-funnel consideration, closing search.

The common mistake, including ours

The mistake is applying the textbook 40/20/40 split without checking whether the first touch in your reported paths is actually the true first touch, or just the earliest one your attribution window happened to capture. Sponsored Products' click window runs 7 days for sellers; DSP's offsite exchange inventory has historically run on a 14-day click and view structure. If a customer's real journey started 20 days before purchase but your reporting window only reaches back 14, the touch your position-based model calls "first" is really the earliest one that survived the window — not the genuine start of the journey. We've built a position-based report before without checking this, and the "first touch" credit landed on a mid-journey Sponsored Brands click that only looked like the start because an earlier DSP impression had aged out of the report entirely.

That error is specifically dangerous with a U-shaped model, more so than with linear or time-decay, because 40% of every sale's credit rides on correctly identifying that one touch. Get the true first touch wrong systematically — by using too short a window — and you don't just misallocate a small slice of credit the way a linear model would; you hand a large, fixed share of every sale to whichever channel happens to survive the truncated window most often, which in practice tends to inflate mid-funnel sponsored ads at the direct expense of the upper-funnel media that actually deserves the credit.

How to set this up correctly

Before running a position-based model, confirm the lookback window you're querying is long enough to capture genuine first exposure for your category's typical path length — Amazon Marketing Cloud's extended traffic lookback now reaches considerably further back than any single ad product's native reporting window, which makes it the more reliable source for identifying a true first touch. And treat the 40/20/40 split as a starting point to test, not a fixed rule — recompute it against your own path data periodically, the same discipline that applies to choosing a time-decay half-life.

When position-based attribution gives you a result you don't trust

If the model credits a channel more than seems plausible given what you know about the account, work backward through the same two checks every time: is the reported first touch actually the first touch, or an artifact of a too-short lookback window, and has anything about the middle touches' role changed recently — a creative refresh, a new placement, a shift in who's actually seeing the middle-funnel ads. If both check out and the result still looks wrong, that's a legitimate reason to question the 40/20/40 convention itself for your specific funnel rather than the data behind it, and to test an alternative split against a holdout period before adopting it as the new standard.