Separating What Happened From What Caused It

The shape almost every one of these disputes arrives in

Nearly every marketing dispute arrives in the same shape. There are two series. One is something the defendant did — a campaign launched, a competitor's advertising claim, an agency's configuration change, a keyword bought, a tracking script deployed. The other is something that happened to the plaintiff — traffic down, conversions down, revenue down, rankings gone. Plotted on one axis they move together, and the complaint is built on that agreement.

The professional question is whether anything joins them. Correlation is what the record hands you for free; it is a set of time series, and series produced by the same market move alike. Causation is what the claim requires, and nothing in the record supplies it. The distance between the two is the entire analysis. An expert who does not treat it as a distance has narrated a coincidence.

This page is the orientation: what would actually measure cause, why that measurement is almost never available by the time counsel calls, and what an expert does instead.

What the claim requires, stated as an element rather than a chart

Causation in these cases is a legal element with a shape, and the shape does not match a chart. In Lexmark International, Inc. v. Static Control Components, Inc., 572 U.S. 118 (2014), the Supreme Court held that a false advertising plaintiff under the Lanham Act must show proximate causation — "economic or reputational injury flowing directly from the deception wrought by the defendant's advertising," which "occurs when deception of consumers causes them to withhold trade from the plaintiff." That is a claim about a specific mechanism running from a specific act through specific consumers to a specific loss.

The damages framework asks the same thing in economic language. The Reference Manual on Scientific Evidence, 4th ed. (2025), frames the exercise as the difference between the actual world and the but-for world, and states that the measurement must "isolate the change in the plaintiff's economic position caused by the harmful act and exclude any change … arising from other causes." Both formulations put the same burden in the same place: not on the coincidence, but on the exclusion of everything else.

The one family of methods that measures cause, and why it has usually expired

There is a method that answers the question directly. An incrementality test — a randomized experiment run on advertising — splits a population of users or geographies, exposes one arm to the campaign, withholds it from the other, and treats the difference as the causal effect. Google's Conversion Lift is described in its own documentation 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 same quantity by synthetic control, building an artificial comparison market out of untreated markets.

These are different in kind from everything else, because randomization removes the need to assume that every confounder was measured. They are also, in litigation, almost always gone. The holdout has to be created by withholding advertising while the campaign runs; once it ran everywhere, no untreated arm exists and none can be manufactured afterward. The test has to be powered in advance, and access was gated — Google states that Conversion Lift "isn't available for all Google Ads accounts" and requires an account representative. The granular data a retrospective reconstruction would need has usually aged out of the platform. The consequence is worth saying plainly. In most retrospective marketing disputes, no causal measurement of the campaign exists and none can now be created.

A ranking of what the available evidence can bear

It helps to have the hierarchy in front of you before anyone offers an exhibit, ranked by the strength of the causal claim each can support:

1. A randomized incrementality test run at the time — causal, contemporaneous, prospective by design.
2. A marketing mix model calibrated by experiment — Google builds experiment priors into Meridian; Meta calls experimental results the "ground truth" for calibration.
3. An uncalibrated marketing mix model — causal inference from observational data, under assumptions its own publisher says cannot be tested.
4. Data-driven attribution — a model over observed paths on one company's surfaces, which Google concedes can produce the same result as 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, and nothing else.

Most disputes are litigated at level six and argued as though they sat at level one.

An attribution number allocates credit; it never observed a cause

Attribution is the concept most often mistaken for causation, because the interface presents it as a measurement. It is not one. Google Ads defines attribution models as determining "how credit for conversions is assigned to different ads, clicks and factors along the user's path to completing a conversion." That is a rule for dividing credit among things that were observed. Change the rule and every per-channel number changes while not one thing that happened changes with it.

Three consequences follow, each a cross-examination line. First, Google's current documentation offers two models, last click and data-driven; first click, linear, time decay and position-based were withdrawn in 2023, and conversion actions still using them were switched to data-driven automatically. Second, Google states that "depending on data availability, the last click and data-driven attribution models can have the same results in certain situations," so a data-driven label on an exhibit does not settle what logic produced it. Third, each platform attributes conversions to itself, under its own window and its own model, which means summing conversions across Google, Meta and an analytics property produces a figure that is not a quantity. Underneath all of it, some conversions being allocated were never observed: Google models them for cross-device, consent-limited and device-restricted situations, and the report does not mark which is which.

Mix models infer from observation, and the central assumption cannot be tested

Marketing mix modeling — regression of a business outcome on spend by channel, with controls — is the method usually reached for when no experiment exists. Google classifies its own open-source implementation, Meridian, explicitly as "causal inference from observational data." The classification is the concession.

Google's required assumptions page states that the model needs conditional exchangeability, which requires that "all confounding variables are measured and included in the control array," and then says the quiet part: "it is difficult to know whether all of the confounding variables are measured because it is purely an assumption, and there is no statistical test to determine this from your data." Elsewhere Google adds that "multiple models can have good fit and predictive power yet provide different ROI and optimization results," and that "a model with 99% out-of-sample R-squared can still be a poor model for causal inference." That sentence answers the exhibit that offers goodness of fit as validation.

Two more limits matter. Google calls mix modeling "a macro tool that works well at the channel-level," which is not the resolution most disputes are litigated at, and describes its own worked example — twelve channels and six controls against two years of weekly data — as "too low to estimate the model reliably."

Dates that break a series before anyone touches the marketing

Before any comparison of two periods means anything, the periods have to be measuring the same thing. Several definitional changes fall inside the window of a great many live disputes, and each one moves reported numbers without moving a single sale.

• 19 January 2021 — Meta announced it would no longer support 28-day click, 28-day view and 7-day view attribution windows, moving to an ad-set setting defaulting to 7-day click and 1-day view.
• 26 April 2021 — Apple's App Tracking Transparency requirements took effect for all apps; where a user declines, Apple states the advertising identifier returns as all zeros.
• September 2023 — Google Ads switched conversion actions still on the withdrawn rules-based models to data-driven attribution.
• November 2023 — the same four models became unavailable in Google Analytics 4.
• 17 October 2025 — Google marked most Privacy Sandbox technologies, including the Attribution Reporting API, for removal, while keeping third-party cookies subject to user choice.

A year-over-year Meta comparison spanning January 2021 measures a window change; a Google Ads comparison spanning September 2023 may measure a forced model migration. An expert who has not identified which breaks fall inside a comparison has not established that the comparison means anything.

What an expert does instead, and what it should be called

What remains is observational: platform reports, analytics, server logs, invoices, revenue. A great deal can be done with those, provided nobody dresses them as an experiment. The defensible version of this work has a recognizable shape.

It states the question as a but-for question. It identifies the data available, its retention limits, its extract date, and which figures are modeled rather than counted. It establishes which attribution model, window and consent configuration were in force in each period compared. It enumerates the alternative explanations — seasonality, market-wide movement, ranking updates, competitor entry, pricing, site and tracking changes, other channels — and states, for each, whether the data excludes it, is consistent with it, or cannot reach it. It uses at least one comparison series and says what that series does not control for.

That is inference under constraint, not measurement. "Consistent with," "cannot be excluded," and "the data do not permit a determination" are legitimate findings and are often the correct ones. Rule FRE 702, as amended in December 2023, puts the burden of establishing reliability on the party offering the opinion and warns against overstating conclusions beyond what the method supports; that rule is analyzed in depth on the sibling search marketing site rather than here.

Where this work stops, and who picks it up

Three boundaries keep this kind of testimony inside its competence, and stating them is what makes the rest credible.

Quantum belongs to the damages expert. Impressions, clicks, sessions and conversions are not dollars. Margin, cost structure, mitigation, discounting and apportionment are the accounting expert's province. The seam is apportionment: the marketing expert says which explanations the data can and cannot exclude; the damages expert converts that into a number. A report that silently crosses the seam invites a Rule 702 motion — in Grasshopper House v. Clean & Sober Media (9th Cir. 2021, unpublished memorandum, not precedent under Circuit Rule 36-3), the panel affirmed exclusion of a damages expert whose "regression analysis was flawed as to the issue of causation."

State of mind is not in the data. A change history records that a setting changed, on a date, under a login. It does not record why, who instructed it, or whether it was deliberate, an error, a test, or an automated recommendation applied in bulk. Willfulness is relevant to remedy after Romag Fasteners, Inc. v. Fossil Group, Inc., 590 U.S. 212 (2020), but it is a conclusion for the trier of fact, not something read out of a spend chart.

Legal sufficiency is not the expert's call. An expert supplies the mechanism and the record. Whether that showing satisfies proximate cause is the court's question, and an opinion phrased as a legal conclusion invites exclusion.

Frequently Asked Questions

Published 2026-08-12