Before-and-After Damages Analysis in Search

What the model asserts, and the assumption underneath it

A before-and-after damages model compares a defined pre-event period with the damage period and treats the difference as the loss. The arithmetic is trivial. Everything that matters sits in one assumption: that the pre-event level would have continued but for the defendant's conduct. That assumption is not background to the opinion. It is the opinion, and it is the part that gets tested.

The measured quantity is usually organic search traffic — visits arriving from unpaid search results, as distinct from paid advertising, direct visits, email, and referrals. Whichever unit is used, the model needs a period before the event that fairly represents what would have happened next.

Lost profits must generally be proven to reasonable certainty: evidence providing sufficient data from which the court or jury may properly estimate damages. Note the asymmetry that governs this discipline. The fact of damage is held to a stricter standard than the amount, and in search matters that cuts against the claimant: the fact of a decline is usually plain in the data, and the attribution of it to the defendant is not.

Why organic search has no stable counterfactual

Most damages settings have a reasonably quiet background. Organic search does not. Four independent forces move the same series the model is measuring, and all four are moving during the damage period.

Algorithmic change. Google ships broad core updates several times a year, plus spam updates and continuous unannounced ranking changes that carry no date at all.
Seasonality. Search demand is strongly seasonal and the seasonality is category-specific. Tax services, gift retail, and travel do not share a calendar.
Competitors. Other sites publish, rebuild, and migrate on their own schedule, and a competitor's improvement is indistinguishable from a claimant's decline in a ranking series.
Query demand. The volume of the underlying searches moves independently of anyone's website, and a collapse in interest looks exactly like a collapse in ranking if you measure only sessions.

Two further movers are avoidable and cause an enormous share of the disputes I see: a change in measurement, and the party's own site changes in the same window.

A naive before-and-after on organic sessions therefore attributes to the defendant everything Google, the calendar, and the market did in that window. This is the model most commonly attempted in search matters and the model most likely to be excluded. Both are true at once, which is why this page exists.

The exclusion authority, and why this is now admissibility

The controlling analogy is not a search case. In Concord Boat Corp. v. Brunswick Corp., 207 F.3d 1039 (8th Cir. 2000), the Eighth Circuit held that the plaintiffs' econometric damages model should not have been admitted because it "did not incorporate all aspects of the economic reality of the stern drive engine market." The model ignored inconvenient evidence and failed to account for market events both sides agreed were unrelated to any anticompetitive conduct, including a product recall and problems arising from a merger. The court reversed. The opinion is available from Justia.

The search analogue is exact. A before-and-after model that ignores a core update running inside the damage window, or a migration the claimant performed itself, omits market events both sides agree the defendant did not cause.

What changed the stakes is the amendment to Federal Rule of Evidence 702, effective 1 December 2023. The proponent must now demonstrate to the court that it is more likely than not that the testimony rests on sufficient facts or data under (b) and that the opinion reflects a reliable application of the principles and methods under (d). The Advisory Committee said why: many courts had held that the sufficiency of an expert's basis and the application of the methodology were questions of weight rather than admissibility, and "[t]hese rulings are an incorrect application of Rules 702 and 104(a)." A method that compares before to after does not support an opinion that the defendant caused it, and (d) is aimed at exactly that gap.

What a clean baseline actually requires

A defensible pre-event window is not simply the twelve months before the complaint. It has to satisfy six conditions at once, and one of them usually fails.

Long enough to contain a full seasonal cycle. A year at minimum, so the same calendar months are compared, and ideally longer so one year's seasonality can be checked against the year before.
Free of the party's own discontinuities. No migration, redesign, content purge, paywall, or change of hosting or domain inside the baseline.
Consistent in measurement throughout. Search Console clicks and analytics sessions are different quantities, and a switch between them mid-series produces an artifact that looks like a loss.
Still in existence. Search Console Performance data covers a rolling 16 months; where the loss began earlier, Google's own record is gone unless it was exported beforehand. Analytics event-level retention in a standard property is a setting of two or fourteen months, with longer tiers in the paid edition; that setting does not affect standard aggregated reports, so monthly sessions by channel usually survive when user-level detail does not.
Anchored to an event date fixed to the day. If the date of the conduct is contested, the baseline boundary is contested too.
Checked against what Google shipped. A documented review of confirmed ranking updates overlapping the window belongs in the report, whether or not it helps the retaining party.

When the model is defensible, and it sometimes is

None of the above means before-and-after is always wrong. It is defensible in one recognizable fact pattern, and I have relied on it there without hesitation.

The pattern is a discrete, dated, mechanical event whose effect is confined to a knowable set of URLs. A noindex directive left in place after a launch, telling search engines not to include those pages in their index. A redirect set that returns 404 instead of 301. A robots.txt file disallowing an entire directory. In each, the mechanism is documented, the affected pages are enumerable, and the pages it did not touch act as a control.

Three conditions make it stronger still. A short window, because slow-moving confounds have less room across six weeks than across two years. No confirmed update overlapping the inflection, established from the dated record rather than asserted. And best of all, a reversal on remediation — the affected pages recovering on a schedule consistent with recrawling while the control pages do nothing, which is the mechanism switched off and on again.

The short version: before-and-after is defensible where a documented mechanism does the causal work and the comparison merely measures it, and indefensible where the comparison does the causal work by itself.

The structural break every pre-2024 model assumes away

One confound is fixed by no amount of careful baseline selection. Every damages model built for search before 2024 assumed the relationship between ranking position and traffic was roughly stationary — that holding position constant held clicks roughly constant. That is no longer true. A model spanning 2023 to 2026 that does not control for it attributes an industry-wide structural change to the defendant, which is the Concord Boat failure in modern dress. Three measurements of that shift have published methodologies and are worth stating precisely.

Zero-click share. SparkToro's analysis of Similarweb clickstream panel data found that 68.01% of US Google searches ended without a click between January and April 2026. The caveats travel with the number. It is a US desktop and mobile web browser panel for that four-month window, weighted on an assumed two-to-one mobile-to-desktop ratio, and the sample size is undisclosed. Most importantly, the panel excludes the Google mobile app, covering only mobile searches made in a browser, so SparkToro's own view is that the true rate is likely higher rather than lower. SparkToro also states that long-term comparisons against earlier years are not directly equivalent because the underlying data providers changed — which means the figure belongs in a report as a level for a stated window, never as a measured change over time.

Click-through impact. Ahrefs reports that AI Overviews are associated with a 34.5% lower click-through rate for the top-ranking organic result, from 300,000 informational-intent keywords, comparing Search Console desktop click-through at position one in March 2024 with March 2025. Ahrefs states the analysis is correlational rather than causal.

Corroboration. Pew Research Center found users clicked a traditional result link in 8% of Google visits where an AI summary appeared, against 15% where none did, from 900 US adults sharing browsing activity in March 2025. Google disputes the study; a report is stronger for saying so.

A baseline drawn from 2023 and a damage period running into 2026 are not measuring the same world, and the difference includes a market-wide shift no defendant caused.

How the model is attacked, and how it is defended

The cross-examination writes itself, which is a reason to run it against your own work first.

Why that pre-event window, and what does the number become one quarter longer or shorter?
What confirmed ranking updates ran inside the damage period, and what did you do about each?
What else changed on the site in that window, and who changed it?
What did the other channels and comparable competitors do over the same dates?
Did demand for the underlying queries move, and how do you know?
Was there any change to tracking, tagging, filtering, or property configuration?
Run your model on a period containing no alleged conduct. What loss does it report?

That last one is a placebo test, and it is the question I would ask first from the other side. A model reporting a large "loss" for a quiet period is measuring noise, and the witness usually learns this in a deposition.

The defenses are not rhetorical. Narrow the window. Restrict the comparison to the pages the mechanism actually touched. Quantify the confounds rather than denying them, and claim the residual. An opinion that says "this much of the decline aligns with a dated market-wide update and this much does not" is far harder to exclude than one that claims the whole decline and hopes.

The controls that make those answers possible sit in the same dataset — the party's other channels, unaffected page cohorts, competitor visibility, and query-level demand — and the design that combines them across the event date is a difference-in-differences comparison rather than a chart with an arrow on it.

Frequently Asked Questions

Is a before-and-after traffic comparison enough to prove damages in a search case?

Rarely on its own. The comparison establishes that traffic fell, which is usually not disputed. What it does not establish is that the defendant caused the fall, because organic search has no stable counterfactual: algorithm updates, seasonality, competitor activity, and shifts in query demand all move the same series in the same window. Under Rule 702 as amended in December 2023, the proponent must show the opinion reflects a reliable application of the method, and a bare before-and-after does not support a causal conclusion. It is a starting exhibit that needs controls layered on it.

How long does the baseline period need to be?

Long enough to contain a full seasonal cycle, which in practice means at least twelve months and preferably twenty-four, so the prior year's seasonality can be checked against the year before it. Length alone is not sufficient. The window must also be free of the party's own discontinuities — no migration, redesign, replatform, or tracking change inside it — and measured in a single consistent unit throughout. A long baseline that spans a change in how traffic was counted is worse than a shorter clean one.

What happens if the analytics data for the baseline period no longer exists?

Search Console Performance data covers a rolling sixteen months, so for a loss that began earlier, Google's own record is gone unless it was exported or a bulk export was configured in advance. Analytics event-level retention is a per-property setting of two or fourteen months in the standard edition, with longer tiers in the paid one. Importantly, the retention setting does not affect standard aggregated reports, so monthly sessions by channel often survive when user-level detail does not. What is left after that is server logs, contemporaneous exports, archived captures, and third-party estimates whose limits must be stated.

Can a before-and-after model survive a concurrent Google algorithm update?

Sometimes, but not by ignoring it. The update has to be dated from the published record, its effect on the party estimated using a control — unaffected page cohorts, a competitor set, or a comparable query group — and that effect separated from the conduct. Where the conduct and a broad core update land in the same week and move the same pages, the honest answer is sometimes that the available data cannot separate them. After the 2023 amendment to Rule 702(d), an unaddressed overlapping update is an admissibility problem rather than something to be argued about at trial.

Does the shift toward AI answers make older damages models unreliable?

It makes them unreliable across the boundary. Every search damages model built before 2024 assumed the relationship between ranking position and clicks was roughly stationary. It is not. SparkToro's analysis of Similarweb clickstream data put US zero-click searches at 68.01% for January to April 2026, on a panel that excludes the Google mobile app, and Ahrefs measured a 34.5% lower click-through rate at position one where AI Overviews appear. A model with a 2023 baseline and a 2026 damage period is comparing two different worlds and, uncorrected, charges the difference to the defendant.

What is a placebo test and why would an expert run one?

A placebo test runs the same model over a period in which no alleged conduct occurred and reports what loss it produces. If a model applied to a quiet window reports a substantial loss, the model is measuring ordinary variation rather than the defendant's conduct, and the result on the real window means much less. It is a cheap, reproducible check that goes directly to whether the method has an assessable error rate, which is what a reliability inquiry into technical testimony asks. Opposing counsel will run it if you do not.

Should the search expert convert the traffic loss into a dollar figure?

The defensible practice is not to. The search expert establishes causation and the counterfactual traffic or visibility series and stops there; a financial expert takes that series as a stated assumption and converts it into revenue and incremental profit, handling avoided costs, mitigation, and present value. A technical expert who multiplies sessions by an order value has produced a financial opinion without financial method, and a challenge aimed at that half of the work can reach the technical half with it.

Published 2026-08-11

Before-and-After Damages Analysis