What a good attribution reporting API returns

A reporting API is worth automating only when its response carries four things: the estimate, an interval around it, the share of your orders the estimate covers, and the label saying how it was produced. Automate a response missing any of those and you have built a schedule around a guess.

Most teams reach for automation at the point where the manual report has become a chore. That is the wrong trigger. The right trigger is when the number in the report is one you would defend in a meeting, and you want it to arrive without a person assembling it. Automation multiplies whatever is already there.

The four fields, and why each one matters

The estimate on its own is the field every vendor returns. The other three are the ones that let a person downstream decide whether to act. A confidence interval that straddles break-even is a different instruction from one that sits entirely below it, and a dashboard that reports only the midpoint erases that difference every time it refreshes.

Coverage is the field teams discover late. If the read explains sixty percent of the orders in the window, the remaining forty percent is not attributed to nothing, it is simply unexplained, and any budget decision taken on the sixty needs to know that. We wrote about how that gap shows up in a GA4 window in the coverage problem in unassigned traffic.

The design label is the field vendors skip

Design is the one field that costs a vendor something to publish, because publishing it forces the admission that most reads are observational. Observational is not a failure. It is the honest description of an estimate produced from variation that already existed in your data, rather than from variation you created with a holdout test. It ranks below a randomised design and above a rules-based model, and a report that says so is more useful than one that implies more than it did.

We treat this as a standing requirement rather than a nice-to-have, and the reasoning is written up in why vendors should explain methodology openly.

What automation should not do

Automation should not smooth. A common pattern is a scheduled report that averages several windows to make the line look stable, which hides exactly the volatility a decision-maker needs to see. A second pattern is silent imputation, where a channel with too little spend to measure is assigned a number anyway rather than being named as below the measurable floor.

Both make the report look more finished than the underlying data supports. If a channel cannot be measured at its current spend, the honest response is a row that says so. We publish the same discipline in our own case work, including the two brands we declined because the spend was not yet large enough to read.

Where Causality Engine sits

Causality Engine reads a Google Analytics CSV export and returns a per-channel causal estimate with its interval, its coverage and its design label. The first read is a manual upload at €99, once, with a full refund if it does not move a budget decision. Developer API keys and the MCP server sit on Pro at €299 a month, alongside unlimited uploads and the direct integrations, which is the tier where scheduled reporting becomes possible at all.

Two paths worth walking before you wire anything: the interactive demo runs the real model on a sample store with no signup, and how it works sets out the method in plain language. If you want the automation-specific detail, attribution tools with automated GA4 ingestion and API access covers the questions to put to any vendor.

The test to run before you automate

Take last month's report. Cover the estimate column with your hand. Ask whether the remaining columns would let a colleague reach the same decision you reached. If the answer is no, the report is not ready to be scheduled, and the honest next step is fixing the read rather than the cadence.

Related answers

• Automate attribution reporting in one afternoon
• Own the pipe: attribution API keys and exports
• Automate attribution before the Black Friday window
• Attribution that adds nothing to your storefront
• Attribution belongs where the decision happens