What happened when eBay turned off its own brand-keyword ads?

What happened when eBay turned off its own brand-keyword ads?

The question every paid-search team eventually faces: are the ads buying conversions, or buying credit for conversions that would have happened anyway?

The study. eBay's economics team (Blake, Nosko, Tadelis, published 2015 in Econometrica) ran a controlled geographic experiment — switching off paid search across a random subset of US designated market areas and measuring the difference against held-out controls. This is incrementality testing: comparing a treated group against an untreated one to isolate causal lift, not just correlation.

What they found. For branded keyword ads (people searching "eBay"), the return on ad spend was effectively negative. Organic search simply absorbed the traffic — users typing your name will click your free listing when the ad disappears. For non-branded keywords, returns were positive but far smaller for frequent users than last-click reports implied. Heavy buyers showed near-zero incremental response.

The nuance. Correlation looked spectacular: branded ads "converted" beautifully because they intercepted people already heading to checkout. Causation, measured by the holdout, told the opposite story. The bias scales with how well-known your brand is.

Bottom line for practitioners: last-click attribution systematically over-credits branded and high-intent search because those clicks ride on demand you already created. Before trusting a channel's reported return, run a geo holdout — even a two-week pause across matched regions exposes whether the spend is incremental or just intercepting. eBay's branded-search return wasn't a model artifact; the model was crediting interception as creation.