Key takeaway
Any estimate without a control group is confounded by construction. Seasonality, an algorithm update and a model refresh hit the treated and the untreated alike and cancel in the difference between them. Without that difference you have a chart, not a result.
The rule the engine enforces is that no verdict is issued without a control unit. It is the whole reason one design was rewritten and another was demoted.
Stop attributing a move to the thing you happened to change. A seasonal swing, a Google update or a model refresh is not a column in your data, and all of them land on the control arm too, which is what makes the difference readable.
Refuse the number that looks confident and is not. A univariate forecast of the series you changed has no control unit, so it cannot separate your intervention from whatever else moved that week, and that design was removed rather than kept behind a warning.
Ask any vendor whose test produced their lift figure. We publish no lift, no incremental ROAS and no p-value until a customer test concludes, and the 20 to 60% range everyone quotes is from the literature, which we keep saying.
Get a spend change, not a PDF. A concluded test writes a calibration factor that scales the response curve behind the budget optimizer, so the next allocation runs on the measured number.
Each returns the same result shape, so one verdict rule applies to all of them and a weak test reads as inconclusive rather than as a small effect.
Ads paused in test regions while matched untreated regions keep running, against a weighted synthetic control built over the candidates, reporting incremental revenue and percent lift. It needs at least 5 control regions and regional first-party conversions, and it is the only design that writes a calibration factor.
A channel paused everywhere it is not held out, measured against untreated control regions. Without those regions it returns inconclusive rather than a number. That is the change: this design used to require no regional data at all, and that is exactly what made it unsound.
Meta and Google randomise their own audiences and report the result. TrustData reads those group conversions against your deduplicated first-party outcomes. It is claim reconciliation, it is useful, and it is not an experiment we ran, so it writes no calibration factor.
Incremental revenue over observed revenue in the treated units, stored with a confidence and a measurement date. It fades toward no adjustment on a 90-day half-life, so an ageing measurement loses influence rather than acquiring a verdict.
FAQ
Not yet, and we are not going to imply otherwise while that is true. The designs are implemented, the estimator is one kernel shared across paid, SEO and AI visibility, and the plumbing that carries a concluded result into the budget optimizer is wired. What has not happened is a customer test running to a conclusion. Until one does, this page describes a method and its guarantees, and every number on it is either a design threshold or a figure from published literature with that stated. A vendor showing you lift percentages in a sales deck should be asked whose test produced them.
It measures whether a channel is causing conversions or merely appearing in the path of conversions that would have happened anyway. Attribution models, data-driven attribution included, distribute credit among the touchpoints that were present. None of them can tell you what would have happened if the ad had never run, because that path does not exist in your data. An experiment creates it by holding part of the population or the map untreated, so the untreated arm shows you the counterfactual instead of a model estimating it.
Because it was not a causal estimate. It paused a channel for a few weeks and fitted a time-series forecast to the paused channel's own history, then called that forecast the counterfactual. A forecast of the series you intervened on carries the intervention's confounders with it, so the gap between forecast and actual is your change plus everything else that moved. It now requires untreated control regions like every other design here, and returns inconclusive without them. Losing a test that anyone could run was the cost of not shipping a number nobody should trust.
The share of the revenue observed in the treated regions that the ads caused: incremental revenue over observed revenue. It is deliberately not the ratio of measured lift to the platform's claimed lift, because the ad tables carry no country dimension and there is no region-scoped platform revenue to divide by. Defining it that way would mean computing it from data that does not exist. A factor of 0.6 says 60% of the observed revenue was incremental. It decays toward 1.0 rather than toward 0, so a forgotten experiment fades out and the uncalibrated curve is what remains.
Attribution assigns credit among the channels that were present, and it is the daily working surface. Incrementality corrects it, and it is occasional and expensive because a real experiment costs you spend you deliberately withheld. Beacon is where the two meet: a concluded test rebuilds the response curve, the optimizer runs on the calibrated curve, and the resulting spend shift arrives as a recommendation with the test in its audit trail. That is the loop. Attribution alone tells you where the credit went; only the experiment tells you whether the credit was real.
14-day free trial
14-day free trial. Deduplicated attribution and first-party tracking ship on every tier. Experiments arrive with Beacon on Optimize.