The Sunk Cost Prison of Last Year's Black Friday Winner

The channel that reported the best ROAS last Black Friday will get the largest share of this year's budget unless someone tests it, and the two best-known experiments in the field say that a reported winner is often the channel buying sales that were already on their way. The budget is not being allocated to what worked. It is being allocated to what was credited.

Octalysis names the trap: the Sunk Cost Prison, where past investment makes the next investment feel obligatory, and its quieter cousin Status Quo Sloth, where changing the plan costs more effort than repeating it. Both are Loss and Avoidance drives, and both are running in every Black Friday planning document right now.

The experiment that should be more famous

Blake, Nosko and Tadelis, in Econometrica in 2015, switched off eBay's own brand-keyword advertising and watched what happened. Their finding, as The Price of Being Found quotes it: brand-keyword ads had no measurable short-term benefit, because 99.5% of the forgone paid clicks came straight back through natural search. Same visitors, same purchases. The advertising had been buying traffic that was already coming, and in every dashboard it had been the best-performing channel in the account.

That is the shape of a sunk cost prison built from a report. The channel looks like the winner precisely because it sits closest to the purchase, which is where the credit lands and where the causal effect is smallest.

The finding that explains why the report cannot tell you

Gordon, Zettelmeyer, Bhargava and Chapsky, in Marketing Science in 2019, compared standard observational attribution methods against 15 large randomised experiments at Facebook. In the paper's own words, in half of the studies the estimated lift was off by a factor of three across all methods, and six of the fourteen checkout studies could not detect a significant lift at all. Their worked example: a true lift of 72.8%, a naive exposed-versus-unexposed comparison of 316%.

The book's reading is the useful one for a budget meeting: the naive number is not a bad fit to the data. It is an excellent fit. It faithfully reproduces the platform's choice of whom to show the ad, and the people it chose were the ones most likely to buy anyway. The better the model fits, the more it may be describing targeting rather than effect.

Which of last year's winners is actually in prison

You cannot tell from the report. You can tell from a test, and only if the test can work. The book's rule takes two numbers: the channel's spend as a share of revenue, and the incremental return you would defend to a CFO. Multiply them. If the product clears the smallest lift your revenue noise allows a holdout to see, about 8.3% for a typical DTC brand on the standard design, the question is answerable. If it does not, the channel is not measurable at your scale, and the honest plan says so instead of repeating last year's number.

For the biggest channel in the plan, the product usually clears the bar. That is exactly the channel worth holding out for eight weeks, and the last start date before Black Friday is 2 October.

Leaving the prison without burning the budget

A holdout is not switching the channel off. It is withholding it from a slice of regions, with the pre-period fixed and the detectable lift written down before launch. The spend on the treated regions continues. What you buy is the one thing the report cannot supply: an estimate of what the channel adds, with an interval, from a design that could have found nothing if nothing was there.

If the winner survives the test, scale it with a number behind it. If it does not, the budget it held becomes the most valuable line in the Q4 plan, because it is the only one that was ever going to be reallocated on evidence.

What to do this week

• If you own the budget: list every channel with last year's reported ROAS next to this year's planned spend, then mark which of those rows has a test behind it. The unmarked rows are the prison.
• If you have to defend it: pick the largest unmarked row and run the spend-share times return arithmetic on it. If it qualifies, propose the holdout with the MDE on the slide. If it does not, propose managing it on stated judgement, which is also a decision.

The Black Friday 2026 calendar has the dates. A causal read on your GA4 export is the observational counterpart: per-channel incremental ROAS with intervals, useful for choosing which channel to test first.

As of 9 September 2026. The eBay and Facebook experiments are quoted as The Price of Being Found (Edition 2.10) quotes them in Chapter 9, and carry the book's scope notes.