Can Your Brand Measure Black Friday Lift at All? The Floor

For a large number of advertisers, The Price of Being Found concludes, honest measurement of a channel's effect is unavailable at any price, by any method, because the effect is smaller than the noise the business itself generates. That is a property of arithmetic, not of tools, and it can be computed before anything is spent. Two tables and one multiplication decide whether your Black Friday test can produce an answer. If it cannot, running it anyway produces a confident report that means nothing, and the plan will act on it.

The loss to avoid is the eight weeks, the holdout revenue and the December meeting spent on a null result that was decided before launch. Octalysis's Loss and Avoidance works on people. It also works on plans, and this is the version that saves money.

Table one: individual holdouts

If you can randomise individuals, which mostly means email and on-site tests, two things drive everything: how often people buy and how small a change you want to see. Halving the effect you are looking for quadruples the sample. The book's table, total users across both arms:

Fifty thousand users a month at 1% conversion can detect a 25% lift and nothing smaller. Lewis and Rao, in the Quarterly Journal of Economics in 2015, put the general case in their abstract: informative advertising experiments can easily require more than 10 million person-weeks. The book adds a precision note: their two-million-user example is one campaign's calibration, not a universal threshold, and quoting it as one is wrong in both directions.

Table two: geographic holdouts

Most channels cannot be randomised by individual, so the test on the table is geographic, and two things decide it: how much weekly regional revenue wobbles (ν, the coefficient of variation) and how much history and live time you have. Smallest detectable lift:

Your ν is one spreadsheet formula: for each region, standard deviation of weekly revenue divided by its mean, then averaged across regions. One afternoon.

The rule

Two numbers you already have. A: the channel's spend divided by total revenue. iROAS: what you honestly believe a euro in that channel returns, not what the platform says. Multiply them; that is roughly how much total revenue would move if the channel were switched off. Compare it to your row. A channel at 5% of revenue with an honest return of 2 moves revenue 10%, above 8.3%, so the test can answer. At 2% of revenue, 4%, below the floor. The book's verdict on that second test: it will run for eight weeks, cost the holdout revenue, produce a confident-looking report, and tell you nothing, because it never could.

Holdout size barely enters the comparison; the book notes it drops out to a first approximation. Budget is the wrong question. Intensity is the right one.

If the channel does not qualify

Three options, all better than running it anyway: concentrate the spend so intensity rises, lengthen the test, or accept that the channel is not measurable at current scale and manage it on judgement, openly. The book calls the "not measurable" list the more useful of the two, because it is where measurement theatre stops being paid for. It also applies the same floor to its own publisher's product: a causal read on a GA4 export has a fit floor of roughly 5,000 euros a month in paid spend and 40 to 90 days of history, and reports which channels fall below it rather than inventing a number for them.

What to do this week

• If you have to defend the number: compute ν, find your row, qualify every channel, and bring both lists. The holdout for the qualifying channel starts by 2 October.
• If you own the budget: read the "not measurable" list first. Every channel on it is currently being managed on a number that could not exist.

The calendar has the dates. A Dutch province geo test cannot reach p below 0.05 is the second floor, and it catches tests that cleared this one.

As of 9 September 2026. Both tables, the A times iROAS rule and the Lewis and Rao quotation are from The Price of Being Found (Edition 2.10), Chapter 15 and Appendix D, and reproduce from the formulae given there. The fit floor for a causal read is this site's own stated threshold.