The attribution loop: change spend, then re-read

Automating an attribution read is worth doing because it makes re-reading cheap, and re-reading is only informative if each cycle carries one deliberate change. Five changes at once produce a report you cannot interpret, no matter how often it arrives.

This is the part teams skip. They automate the report, keep changing spend the way they always did, and then find the automated report no more useful than the manual one. The report was never the constraint. The experimental hygiene was.

One change per cycle

The discipline is unglamorous and it is the whole method. If you move Meta and cut Google in the same week, the next read gives you a joint effect. That is a real answer to a question nobody asked.

Structuring a cycle

Pick the channel with the most budget at stake, because that is where a wrong answer costs the most and where the estimate will be tightest. Change it by an amount large enough to show up above the noise. A five percent shift on a channel whose interval spans thirty points will not resolve; the read will come back honest and unhelpful.

Then hold everything else. Then re-read. Then write down what you expected before you look, because a prediction you record is evidence and a prediction you remember is a story.

The incremental ROAS guide covers how large a move needs to be for the estimate to separate, and how to measure incremental lift covers the same question from the test-design side.

When the loop says nothing

Sometimes the interval after a change still straddles break-even. That is a result, not a failure. It usually means one of three things: the change was too small, the window was too short, or the channel sits below the spend level at which any method can separate its effect from noise. The third case is the measurability floor, and the honest response is to name the channel as unmeasured rather than to report a number the data cannot support.

A vendor that always returns a confident number for every channel, regardless of spend, is telling you something about its willingness to say "we do not know" rather than something about your business.

What the API adds

Manually, a cycle costs a person an hour and therefore happens monthly. With a scheduled read it costs nothing and can happen weekly, which is the difference between four data points a year and forty. On Causality Engine, scheduled reads run on Pro at €299 a month with developer API keys, the MCP server, unlimited uploads and the direct integrations; the €99 one-time read is the right way to see the output shape first.

If you want the loop without the automation, run it manually for a quarter. Most brands discover the discipline matters more than the cadence, and settling attribution debates with your agency is often the real unlock.

The honest limit

A read on observed variation cannot tell you what would have happened under a change you never made. For that you need a holdout test, which costs the revenue forgone in the held-out group and takes weeks. The loop described here is the cheaper instrument, and knowing which questions it cannot answer is part of using it well.

Related answers

• Own the pipe: attribution API keys and exports
• Automate attribution before the Black Friday window
• What a hand-built attribution report really costs
• What you can still test without an attribution pixel
• Attribution in Slack: what to post and what to skip