Setting Attribution Baselines Before Black Friday Chaos

Decide what normal looks like before peak week rewrites every number. Baselines, holdouts, and rules for reading peak data.

Every Black Friday, attribution dashboards produce numbers nobody believes: channels doubling overnight, yesterday's loser suddenly carrying the business, payback windows colliding with the biggest traffic week of the year. Teams then make January budget decisions off data they themselves describe as chaos. The problem is not the week. It is the missing baseline.

A baseline set in calm weeks, with rules agreed in advance for how peak data will be read, turns the chaos into something measurable. Here is how to set one while there is still time.

Pick baseline weeks that resemble peak, minus the peak

Choose two or three recent full weeks with no major promotion, no stock incidents, and no tracking changes, and document why they qualify. From them, record the metrics you will compare against: channel share of first and last touch, assisted conversion patterns, new versus returning mix, and the lag between first visit and purchase.

Write the baseline down where the whole growth team can see it, with the dates and the exclusions. A baseline that lives in one analyst's spreadsheet becomes a rumour by November, quoted differently in every meeting it is needed in.

Agree the reading rules in advance

Peak data tempts everyone into whichever model flatters their channel. Prevent the argument by agreeing now: which view decides budget questions, which view is diagnostic only, how the peak window is defined, and when post peak data gets re read after returns and lagged conversions settle.

Include the known distortions by name: gift buyers polluting returning customer cohorts, discount driven conversion lag compression, and paid social prospecting that will be judged on January repeat behaviour rather than peak week returns. Naming a distortion in October defuses it in December.

Holdouts and sanity checks

Where you can, keep a small clean reference: a geographic or audience holdout on the biggest test, or at minimum a branded search and direct traffic series you trust as a demand thermometer. When every attributed number moves at once, an independent series tells you whether demand moved or measurement did.

Reconcile against money, not just models. Total revenue, orders, and new customer counts from the store platform are ground truth the attribution stack must roughly explain. If the models collectively claim more conversions than the store recorded, peak week is when you find out.

The January re read

Schedule the re read before the holiday, for a specific January date, when returns have landed and lagged conversions have matured. Peak decisions made in December on immature data should be labelled provisional, with the re read as the decision of record for budget setting.

Compare peak against baseline in the re read, channel by channel, and write down what actually changed: mix, lag, new customer share. That document becomes next year's baseline starting point, and the institutional memory that peak planning usually loses by February.

Related reading

If this topic is useful, these earlier guides go deeper on connected questions: View-Through vs Click-Through Attribution: What Each Number Is Really Claiming, Attribution During Holiday Promotions: Reading a Spike Without Fooling Yourself, and Attribution for Influencer and Creator Seeding Programs.

Frequently asked questions

What is an attribution baseline?

A documented set of normal period metrics, channel shares, touch patterns, conversion lag, and customer mix, from clean weeks chosen in advance. Peak results are read against it, so a doubling can be judged as real change, seasonal distortion, or measurement noise.

When should peak attribution decisions be finalised?

In a scheduled January re read, after returns and lagged conversions mature. December readings should be treated as provisional. Agreeing that rule before the holiday prevents budget decisions being locked to the least mature data of the year.

How do I tell demand change from measurement noise at peak?

Keep an independent reference series, such as branded search and direct demand plus store platform revenue and order totals. If attributed channels all surge while ground truth moves modestly, you are looking at mix and model effects, not a demand miracle.

What distortions should we name in advance?

Gift buying in returning cohorts, compressed conversion lag under discount pressure, and prospecting judged too early on peak returns instead of January repeat behaviour. Naming them in the baseline document stops them becoming arguments in December.