Attribution inflation is the ratio of the GMV a marketplace credits to your advertising to the GMV your advertising actually caused. It is at least one by construction — a platform’s attribution rule credits sales within a window that the ad may or may not have produced, and it has no mechanism to subtract the ones that would have happened anyway. The gap matters because it does not merely make reported returns look better; it changes the correct decision. Working through the profit arithmetic, attribution inflation enters as a haircut on your contribution rate, which means it multiplies the Target ROAS at which advertising stops paying. An inflation factor of 1.4 raises the correct target by 40% before any other adjustment.
Formula
attribution inflation a = attributed GMV / incremental GMV, a >= 1
effective contribution rate = contribution rate / a
break-even ROAS = a / contribution rate
Example: 20% contribution, a = 1.4
naive break-even = 1 / 0.20 = 5.00
true break-even = 1.4 / 0.20 = 7.00Example
A Shopee campaign reports ฿100,000 of attributed GMV on ฿16,000 of spend.
A holdout test shows ฿71,400 would have arrived without the ads running.
incremental GMV = 100,000 − 71,400 = ฿28,600
attribution inflation = 100,000 / 28,600 = 3.50
Reported ROAS = 6.25. Incremental ROAS = 28,600 / 16,000 = 1.79.
At a 20% contribution rate the campaign needed 5.00 incremental to
break even, so it lost money while reporting a 25% overshoot.How do you measure attribution inflation?
There are exactly two credible methods and neither uses platform reporting. A holdout randomises which users or which listings are eligible to see the ad and compares outcomes between arms. A geo test switches campaigns off in some regions and on in others, then compares the difference in sales against the difference in spend. Both are experiments, and both cost money in the arm that is switched off, which is why they are rare. What cannot substitute for either is a last-click report, an attribution-window comparison or a before-and-after read: none of them contains a counterfactual, so none of them can separate demand the ad created from demand it merely intercepted.
How large is attribution inflation in practice?
Large enough to invert conclusions, and largest exactly where campaigns look best. The eBay field experiment run by Blake, Nosko and Tadelis found that returns to branded paid search were statistically indistinguishable from zero once measured against a randomised control — customers searching the brand name converted whether or not the ad ran, so essentially the whole of the credited revenue was inflation. Lewis and Rao showed separately that detecting true advertising returns requires far more statistical power than typical campaigns can supply, which is why attributed figures dominate the record. On a marketplace the same mechanism operates through broad attribution: credit flows to the ad for purchases of other products from the same shop, and for buyers who were already in a checkout flow.
Why does attribution inflation raise the Target ROAS I should set?
Because the target is enforced on attributed GMV, and the profit is earned on incremental GMV. If a baht of credited revenue is only worth 1 ÷ a baht of real revenue to you, then your effective contribution rate is your true contribution rate divided by a, and break-even ROAS — which is one over the contribution rate — rises by a factor of a. That is a multiplicative correction and it compounds with the other two adjustments a profit-maximizing target requires: the markup for the fact that buying more clicks reprices the clicks you already had, and the markup for the gap between your average click and your marginal one.
Why it matters
Attribution inflation is the reason two well-run shops can follow identical ad advice and get opposite results. One sells a considered purchase to cold traffic and creates most of the demand it is credited with; the other sells a repeat-purchase staple to an audience already searching for it and is credited with demand it merely intercepted. Nothing on either dashboard distinguishes them, and the correct Target ROAS for the second can be several times the first.
In DataGlass
DataGlass keeps attributed and incrementality-adjusted views distinct rather than blending them, so a recommendation states which basis it was computed on. Where a shop has run a holdout or a geo test, the measured factor is applied to the targets that depend on it; where it has not, the attributed basis is used and labelled as such rather than presented as profit.