Diagnosing a Traffic Drop for Litigation

Fix what fell before asking why it fell

The first failure in this work happens before any analysis: nobody defines the series. "Traffic dropped" can mean rankings moved, sessions fell, users fell, a channel fell, conversions fell, or revenue fell. Those are four different quantities on different instruments, and each step between them is a separate inferential leap. A ranking change can coexist with flat revenue. A traffic decline can coexist with rising profit if the traffic that left never converted. An expert who moves from a rank change to a revenue figure without evidencing each link has built a chain with three unsupported joints.

So the process starts by pinning the series precisely: which metric, from which system, at which grain, over which date range, extracted on which date, under which filters, in which time zone. That sounds fussy until the first time an opposing report turns out to have compared a metric in the account time zone against the same metric in the reporting time zone, and the entire dispute over a two-day shift dissolves.

Pin the definition first, in writing, and the rest of the process has something stable to work against.

Confirm the measurement was the same on both sides of the line

The second step is to establish that the two periods being compared were measured the same way. This is not a refinement. A period comparison that spans a definitional change is measuring the change.

The list of candidates is short and checkable. On the plaintiff's own side: a migration from Universal Analytics to Analytics 4, a tag change, a bot-filtering or view-filter change, a consent banner deployment, a change to which conversion actions were counted. On the platform side: Google Ads switched conversion actions still using the withdrawn rules-based attribution models to data-driven attribution in September 2023, and Analytics 4 dropped those same four models in November 2023; Meta removed 28-day click, 28-day view and 7-day view windows on 19 January 2021; Apple's tracking permission requirements took effect for all apps on 26 April 2021, after which a declining user's advertising identifier is returned as all zeros.

Every one of those is discoverable. The attribution model and conversion window in force in each period sit in the account's conversion-action configuration and its change history. Establishing them is cheap. Failing to establish them leaves the entire comparison open to a single question on cross.

Date the event precisely, then date everything else within reach of it

Only after the series and its definition are fixed does the date of the decline become useful — and the date is rarely a date. Declines have onsets, and onsets have widths.

Google publishes a dated history of its announced ranking updates with rollout durations, and the durations are the point. The March 2024 core update ran 45 days. The August 2024 core update ran 19 days and 4 hours; November 2024, 23 days and 13 hours; March 2025, 13 days and 21 hours; June 2025, 16 days and 18 hours; December 2025, 18 days and 2 hours; March 2026, 12 days and 4 hours; May 2026, 11 days and 21 hours. Spam updates run shorter — the June 2026 spam update completed in 2 days and 1 hour — but they still occupy a window rather than a point. The status dashboard is a primary source and is the first thing to lay against an organic timeline.

Two things follow. An argument that "the decline started on date X" has to contend with an update that was rolling out across a range of dates rather than landing on one. And the list is only what Google announces; the absence of a listed update on a given date does not establish the absence of a ranking change.

The alternatives, in the order they get worked

A competent analysis enumerates the competing explanations and states, for each, whether the available data excludes it, is consistent with it, or cannot address it. The order matters because the cheap and decisive tests come first, and because a report that lands on the defendant's conduct without visibly clearing the rest is doing post-hoc reasoning.

1. Seasonality — is the decline inside the range of the same weeks in prior years?
2. Market-wide movement — did the whole category move?
3. Search ranking updates — does the onset sit inside a documented rollout window, and did the whole site move rather than the pages at issue?
4. Platform policy and product changes — attribution model sunsets, window changes, consent changes, ad-format deprecations.
5. Competitor entry, exit or spend change — visible in auction insights, impression share, ad libraries and share-of-voice data.
6. Macroeconomic and category conditions — consumer spending and sector demand.
7. Own-price and promotion changes — a price rise or an ended promotion moves conversion rate without touching traffic.
8. Product, inventory and fulfillment changes — stockouts, discontinued items, shipping cost, checkout changes.
9. Site and technical changes — redesigns, migrations, URL changes, tag breakage, an accidental noindex, downtime.
10. Measurement changes on the plaintiff's own side — the items in the previous section.
11. Other channels — a cut in email, affiliate, retail media or television spend outside the disputed channel.
12. Sales-side and operational causes — lead response time, headcount, call center hours, credit policy.
13. The disputed conduct.

A report that does not visibly work through the first twelve before landing on the thirteenth should expect to be challenged as post-hoc, and the challenge will be fair.

Seasonality is the cheapest test and the one most often skipped

Seasonality gets skipped because it is unglamorous and because the answer is sometimes inconvenient. It is also the fastest way to end a weak case on either side.

The test needs at least two and preferably three prior comparable periods — the same weeks, in the same series, on the same definition. If the current decline sits inside the band the same weeks produced in each prior year, the decline has not been shown to be unusual at all, and the burden shifts back to whoever asserted it was. If it sits well outside, seasonality has been excluded as a sufficient explanation and the analysis moves on with one fewer live alternative.

Two cautions. Prior-year comparisons import every definitional break between then and now, so the periods have to survive the measurement check first. And a business whose category itself shifted — a product that became seasonal, or stopped being — will show a seasonal pattern that is not stable across years. Where the pre-period is too short or too disturbed to support the test, the correct statement is that seasonality cannot be excluded on the available data, not that it has been.

What a comparison series is, and where one actually comes from

The point of a control is to answer one question: what did comparable things that were not exposed to the alleged cause do over the same period? Retrospectively, there is no true control, only a set of partial comparators with different weaknesses.

• Untreated geographies of the same business. The strongest available retrospective comparator where the conduct was geographically limited. It requires geo-level series that still exist and geos that were comparable in the pre-period. Meta's own synthetic-control documentation concedes the method carries "biases due to inexact matching" when the test unit cannot be reliably recreated from the controls.
• Untreated brands, product lines or site sections of the same business, where the dispute concerns one product or one section.
• The plaintiff's own non-disputed channels. If organic, email and direct fell in the same shape at the same time as paid, the cause is unlikely to be specific to paid. This is often the single most informative exhibit in the file and it costs nothing to build.
• Category or competitor series. Useful directionally, weak as controls. Third-party traffic estimates from panel and clickstream vendors are modeled estimates of somebody else's traffic, not the competitor's server logs, and a report citing one has to read and state that vendor's published methodology. Google Trends is normalized relative search interest on a 0–100 scale, not traffic and not revenue.
• Industry benchmarks. Context only; the benchmark population is rarely comparable to the plaintiff.
• The plaintiff's own pre-period. The weakest form, because it controls for nothing that changed over time. This is the before-and-after comparison wearing a different name.

Choosing the wrong comparator is itself a methodological defect, not a presentational one. In In re Executive Telecard Securities Litigation, 979 F. Supp. 1021 (S.D.N.Y. 1997), a damages expert was excluded in part for using a market index with "no meaningful correlation" to the security at issue. Benchmarking a niche retailer against total national e-commerce is the same error in a different market.

Post-hoc reasoning from a two-line chart is how this testimony fails

This is worth stating without hedging, because it is the single most common failure mode. Two lines on a chart — a date and a decline, or spend and revenue — establish temporal sequence. Nothing more. Every alternative explanation stays live until it is separately addressed, and under FRE 702 as amended in December 2023 the burden of addressing them sits with the party offering the opinion.

Courts have punished exactly this reasoning in adjacent fields. In Concord Boat Corp. v. Brunswick Corp., 207 F.3d 1039 (8th Cir. 2000), an economist's damages model was held excludable because it "was not grounded in the economic reality of the stern drive engine market, for it ignored inconvenient evidence" — purchasing behavior that did not fit, a competitor's recall, a merger, and a market position that predated the challenged conduct. In a marketing dispute the inconvenient evidence is usually a core update, a price change, a competitor launch, or a tracking change sitting in the same window as the alleged wrong. In Grasshopper House, LLC v. Clean & Sober Media, LLC (9th Cir. 20 Aug. 2021), an unpublished memorandum disposition and therefore not precedent under Circuit Rule 36-3, the panel affirmed exclusion of a damages expert on the ground that his "regression analysis was flawed as to the issue of causation."

There is a thin body of published decisions excluding digital marketing experts specifically; the field is young and most of these disputes settle. The general standard applies with full force regardless, and the methodological failures courts have punished elsewhere are all available here.

Writing a conclusion the method can carry

The last step is to say only what the work produced. A defensible write-up of a traffic decline states the question as a but-for question; identifies the data actually available, its retention limits and its extract date; flags which figures are modeled rather than counted; establishes the attribution model, window and consent configuration in force in each period; walks the alternatives and dispositions each; uses at least one comparison series and states what it does not control for; and distinguishes what the evidence shows from what it merely permits.

The conclusion then takes whatever strength the method left. "The decline began within the rollout window of a documented core update and affected the whole site rather than the pages at issue" is a finding. "The decline is confined to the disputed channel while every other channel is flat across the same weeks" is a finding. "The available data are consistent with the alleged cause and do not exclude two identified alternatives" is also a finding, and it is often the accurate one.

What the analysis cannot do is separate causes the record never distinguished. Where the conduct, a pricing change and a site migration all landed in the same fortnight, no amount of care recovers their separate contributions from aggregate series. The right answer there is that the data do not permit the separation — stated in the report, before opposing counsel states it for you.

Frequently Asked Questions

Published 2026-08-12