Attribution and ROI

The answer this page exists to give

There is no primary record of which advertising caused which sale. Every figure claiming to say so is a model output, of one of four kinds.

Data-driven attribution is a counterfactual estimate. It compares the paths of customers who converted to the paths of those who did not, and divides credit accordingly. It has no access to the world in which a touchpoint did not occur.

Marketing mix modeling is a regression. Google classifies it, in its own documentation, as "causal inference from observational data" — and observational is not experimental.

A platform-reported conversion is a platform's claim about its own performance — the seller's report of its own delivery, from inputs nobody else can inspect.

And a share of the conversions being allocated were never observed. Since the tracking and consent changes from 2021 onward, part of any reported count is a modeled estimate of events the platform could not see.

None of that makes the data useless. It makes it evidence of a particular kind, and the work is saying so rather than passing an estimate off as a measurement.

An attribution model is a rule for dividing credit, and the menu has nearly emptied

Google defines an attribution model as determining "how credit for conversions is assigned to different ads, clicks and factors along the user's path." Nothing about the underlying events changes when the rule changes: the same clicks and the same revenue can be reported as very different per-channel contributions under different models. That is the most common source of a misleading exhibit in these disputes.

The menu has narrowed to almost nothing. Google Ads now offers exactly two models, last click and data-driven; GA4 offers three. First click, linear, time decay and position-based were withdrawn from Google Ads over mid-July to September 2023 and from GA4 as of November 2023.

The mechanism matters. Google announced that from September 2023 it would switch conversion actions still using the withdrawn models to data-driven attribution automatically, so a series crossing that boundary may contain a model change the advertiser never chose. Which model reported each period is discoverable from the conversion-action configuration and change history, and establishing it precedes any before-and-after comparison.

What data-driven attribution claims, and the four things it does not

Google's current description is that data-driven attribution "uses your conversion data to calculate the actual contribution of each ad interaction across the conversion path," identifying patterns "by comparing the paths of customers who convert to the paths of customers who don't." Three qualifications belong with it, and Google supplies all three.

Its field of view is Google's own surfaces. Coverage is website, store visit and Analytics conversions from Search including Shopping, YouTube, Display and Demand Gen. It does not see Meta, a marketplace, email, affiliate, retail media or television, so an exhibit built from it describes part of the marketing as though it were the whole.

It can degenerate. Google recommends at least 200 conversions and 2,000 ad interactions in a 30-day period while confirming that the model runs regardless of volume — and concedes that "depending on data availability, the last click and data-driven attribution models can have the same results in certain situations." A data-driven label does not establish that anything other than last-click logic produced the number.

Its published description does not name a method. The Shapley-value and counterfactual language experts quote lives in Google's legacy Universal Analytics documentation; the current pages describe a machine-learned model without naming a technique. Asserting that today's model is a Shapley computation is an inference about a proprietary system.

One further point changes exhibits: conversions can be reattributed for up to seven days, so the extract date must be stated.

Underneath the model, conversions that were never observed

Attribution allocates credit among interactions. Separately, some of the conversions being allocated were never seen. Google's definition is that modeled conversions "use data that doesn't identify individual users to estimate conversions that Google is unable to observe directly," applied where cookies are limited, where consent requirements apply, where device identifiers are unavailable, and across devices. Modeled and observed conversions appear in the same column.

The sentence that matters most is Google's own, and it is regularly overstated in both directions: the modeling "determines whether a Google ad interaction led to the online conversion. It doesn't determine whether or not a conversion happened." Read carefully, that assigns attribution, not existence — a narrower claim than critics usually make, and still a causal assignment made by a model whose logic is not published.

Meta told developers in January 2021 that "statistical modeling will be used for certain attribution windows and/or metrics," alongside a limit of eight conversion events per domain and the removal of breakdowns for offsite conversion events — stripping age, gender, region and placement splits from exactly the events a performance dispute is about.

A conversions cell is therefore a modeled estimate passed through a credit-allocation model — two layers, neither disclosed on the export.

Marketing mix modeling is a regression under assumptions its authors say cannot be tested

Marketing mix modeling relates a business outcome to spend by channel plus control variables, over time and often across geographies. Meta's own analyst guide gives the form as a linear equation estimated with ridge regression.

Both major open-source implementations document their own limits. Google's Meridian, generally available since 29 January 2025, builds incrementality experiment results in as priors — Google's own model treats experiments as the thing that anchors it, not the reverse. Meta describes Robyn as "an experimental, AI/ML-powered and open sourced" package and states that it "strongly recommend[s] using experimental and causal results that are considered to be the ground truth to calibrate MMM."

The data requirements are rarely met. Meta says an MMM needs at least two years of weekly data; Google, in a worked example of 26 parameters against 104 weekly observations, says outright that "this sample size is too low to estimate the model reliably."

The assumptions are where cross-examination lands, and Google states them against its own tool. All confounding variables must be measured and included in the control array — and "there is no statistical test to determine this from your data." Fit does not rescue it: "a model with 99% out-of-sample R-squared can still be a poor model for causal inference." Google also cautions that "MMM is a macro tool that works well at the channel-level" — which disposes of most litigation uses, because these disputes are about one campaign in one period.

What does establish causation, and why it is almost never in the file

One family of methods measures causal effect: randomized incrementality testing. A population of users or of geographies is split, one arm is exposed and one is held out, and the difference in outcome is the estimated causal effect. Because assignment is randomized and the holdout is contemporaneous, it does not depend on assuming every confounder was measured — the assumption a mix model cannot test.

Google describes Conversion Lift as "an incrementality tool that helps you measure the number of purchases, site visits, and any other conversions directly driven by people seeing your ads." Meta's GeoLift estimates the difference between observed results and "what would have happened in a world where it didn't take place," and documents its own weaknesses: bias from inexact matching, and a high chance of failing to find lift where a test was run without a prior power analysis.

In litigation about a past campaign it is usually absent, for four reasons:

1. The holdout has to exist before the fact. Once the campaign has run everywhere there is no untreated arm, and none can be manufactured afterward.
2. The design has to be powered in advance; an underpowered test produces a false null.
3. Access was gated — Google states Conversion Lift requires an account representative, so its absence is not itself evidence about anyone's competence.
4. The underlying data has usually expired under platform retention limits.

So, plainly: in the great majority of retrospective disputes, no causal measurement of the campaign exists and none can now be created. What exists is observational, honest work can be done with it, and presenting it as experimental is the failure mode that gets experts excluded.

The hierarchy, and the dates that break a time series

Ranked by the strength of the causal claim each supports:

1. A randomized incrementality test — causal, but requires prospective design
2. A mix model calibrated by experiment — both vendors treat experiments as the ground truth
3. An uncalibrated mix model — observational inference under untestable assumptions
4. Data-driven attribution — a model over Google-surface paths, sometimes degenerating to last click
5. Rules-based attribution — an arbitrary allocation rule, withdrawn by Google for that reason
6. A before-and-after chart with a vertical line on it — sequence only

Then there is the calendar, because an expert comparing two periods has to know whether a definitional change falls between them. Meta removed the 28-day click, 28-day view and 7-day view windows on 19 January 2021. Apple's App Tracking Transparency requirements applied to all apps from 26 April 2021. Google's rules-based models went in mid-July 2023, with remaining conversion actions auto-switched that September and GA4 following in November 2023. Chrome kept third-party cookies — Google abandoned deprecation on 22 July 2024 and the replacement prompt on 22 April 2025 — and on 17 October 2025 announced the retirement of most Privacy Sandbox technologies, including the Attribution Reporting API. Meta's Ads Insights API moved to tiered retention on 12 January 2026. A comparison spanning any of those dates measures a definitional change as much as performance.

What honest work on observational data looks like, and what it cannot claim

A defensible analysis states the question as a but-for question. It identifies the data available, its retention limits, its extract date, and which figures are modeled. It establishes which attribution model, window and consent configuration were in force in each period compared, and whether they changed. It enumerates the alternative explanations — seasonality, category movement, ranking updates, platform changes, competitor activity, price and promotion changes, site changes, measurement changes on the party's own side, other channels — and says for each whether the data excludes it, is consistent with it, or cannot address it. It uses a comparison series and states its limitations. It ends with a conclusion whose strength matches the method: "consistent with," "cannot be excluded" and "the data do not permit a determination" are legitimate findings and frequently the correct ones.

Courts have not been forgiving of the alternative. In Grasshopper House, LLC v. Clean & Sober Media, LLC — an unpublished Ninth Circuit memorandum disposition decided 20 August 2021, and not precedent under Rule 36-3 — the panel affirmed exclusion of a damages expert whose per-visit valuation rested on a regression the district court found "flawed as to the issue of causation." The inconvenient evidence in a marketing dispute is usually a ranking update, a price change or a tracking change sitting in the same window as the alleged wrong.

What none of this supports: that a chart with a vertical line establishes causation; that an attribution number measures cause; that platform conversion counts are counts of observed events; that figures from different platforms can be added; or that incrementality can be recovered after the fact.

Frequently Asked Questions

Published 2026-08-12