Traffic and Revenue Loss

Establishing that a decline happened at all, which is not automatic

Before anyone argues about cause, the decline has to be established as a fact about the world rather than a fact about a reporting system. That step gets skipped constantly.

A traffic or revenue series has a definition, and the definition has changed several times inside the window most live disputes cover. A Google Ads conversion series crossing September 2023 may contain a forced migration: Google withdrew first click, linear, time decay and position-based attribution from new conversion actions in mid-July 2023 and automatically switched the remainder to data-driven attribution in September. Google Analytics 4 lost the same four models as of November 2023. A Meta series crossing 19 January 2021 spans the removal of 28-day click, 28-day view and 7-day view windows. An app series crossing 26 April 2021 spans Apple's App Tracking Transparency enforcement. A Universal Analytics to GA4 migration changes what a session is.

So the first questions are mechanical. Which metric, from which property, on which date range, extracted when? Which attribution model and window were in force in each period compared, and did they change between them? Which numbers are modeled rather than observed? Data-driven attribution reattributes conversions for up to seven days after the fact, so an extract date belongs on every exhibit. A comparison spanning a definitional change measures the change, not the marketing.

The most common way this testimony fails

Say it plainly: post-hoc reasoning from a chart with two lines on it is the most common way this testimony fails. Traffic fell after the defendant did something, therefore the defendant caused the fall. That is not an analysis. It is a narrated coincidence, and it establishes temporal sequence and nothing else. Every alternative explanation stays live until separately addressed, and the burden of addressing them sits with the proponent.

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 should have been excluded because it "was not grounded in the economic reality of the stern drive engine market, for it ignored inconvenient evidence" — a competitor's recall, a merger, and a market position predating the challenged conduct. A model that omits the obvious competing causes in its own record is excludable however sophisticated it is. In In re Executive Telecard, Ltd. Securities Litigation, 979 F. Supp. 1021 (S.D.N.Y. 1997), an expert was excluded partly for benchmarking against an index with "no meaningful correlation" to the security at issue. Choosing the wrong comparison series is itself a methodological defect, and benchmarking a niche retailer against total national e-commerce is that error in another market.

The corollary is that a legitimate conclusion is often "the data do not permit a determination." An opinion whose strength matches its method survives cross-examination.

The competing explanations, enumerated before the conclusion

A defensible diagnosis works through the alternatives visibly and says, for each, whether the available data excludes it, is consistent with it, or cannot address it.

• Seasonality. Is the decline within the range of the same weeks in prior years? That needs two and preferably three prior comparable periods.
• Market-wide movement. Did the whole category move? That needs a category series.
• Search ranking updates. Google publishes dated updates with rollout durations.
• Platform product changes. Attribution model sunsets, window changes, consent changes, ad format deprecations.
• Competitor entry, exit or spend change. Visible in auction insights, impression share and ad libraries.
• Own price and promotion changes. A price rise or a lapsed promotion moves conversion rate without touching traffic.
• Product, inventory and fulfillment. Stockouts, discontinued items, shipping cost or checkout changes.
• Site and technical changes. Redesigns, migrations, URL changes, tag breakage, an accidental noindex, downtime.
• Measurement changes on the claimant's own side. A GA4 migration, a filter change, a bot-exclusion change, a consent banner deployment.
• Other channels. A cut in email, affiliate, retail media or television spend outside the disputed channel.
• Sales-side causes. Lead response time, headcount, call-center hours, credit policy.

The disputed conduct is the last item on that list, not the first. A report that does not visibly work through the others before landing on it should expect to be challenged as post hoc, because it is.

The published ranking-update history, and how to use it honestly

For organic traffic the first document to open is Google's own ranking updates history, which gives start dates and rollout durations. It is a primary source, it is dated, and it cuts both ways.

Recent entries show the shape of the problem. The March 2024 core update started 5 March 2024 and ran 45 days. The August 2024 core update started 15 August and ran 19 days. The March 2025 core update ran 13 days from 13 March; the June 2025 core update 16 days from 30 June; the December 2025 core update 18 days from 11 December; the March 2026 core update 12 days from 27 March; the May 2026 core update nearly 12 days from 21 May.

Two things follow. First, rollouts take weeks, so an argument that a decline "started on date X" has to contend with an update rolling out across a range of dates rather than landing on one, and a site-wide decline beginning inside a documented rollout window is a competing explanation that has to be addressed rather than ignored. Second, the list contains only what Google announces, and Google has long said it makes changes it does not announce. The absence of a listed update on a date does not establish the absence of a ranking change. That asymmetry is worth stating out loud, because it constrains both sides.

What a control series looks like, and where one actually comes from

A control answers one question: what did comparable things that were not exposed to the alleged cause do over the same period? The options, in descending order of strength, each carry a limitation that belongs in the same sentence as the option.

• Untreated geographies of the same business. The strongest retrospective control where the conduct was geographically limited. It requires geo-level series to still exist and the areas to have been comparable beforehand, and synthetic-control matching carries what Meta's GeoLift documentation calls "biases due to inexact matching."
• Untreated brands, product lines or site sections of the same business. Useful where the dispute concerns one product or one part of a site.
• The claimant's own non-disputed channels. If organic, email and direct all fell in the same shape at the same time as paid, the cause is unlikely to be specific to paid.
• Category or competitor series. Google Trends is normalized relative search interest on a 0 to 100 scale, not traffic. Third-party panel and clickstream figures are modeled estimates, not measurements of a competitor's traffic, and their methodology should be read and described rather than assumed.
• The claimant's own pre-period. The weakest form. This is the before-and-after comparison, and it controls for nothing that changed over time.

In most retrospective disputes a genuine control does not exist. What exists is a set of partial comparators, each excluding some alternatives and not others. Saying that is stronger than pretending otherwise.

Why incrementality cannot be measured after the fact

One family of methods actually establishes causation: randomized incrementality testing, where a population of users or geographies is split, one arm is exposed and one is held out, and the difference is the estimated causal effect. Google's Conversion Lift and Meta's Conversion Lift and GeoLift are the commercial forms.

None can be run retrospectively, for four reasons worth stating to counsel early, because they close off an expensive line of inquiry. The holdout has to be created by withholding ads during the campaign; once the campaign has run everywhere there is no untreated arm and none can be manufactured. The design has to be powered in advance — Meta documents that skipping a prior power analysis "leads to a high chance of failing to find lift, even if it actually happened." Access was gated: Google's Conversion Lift is not self-serve and requires an account representative, so its absence from an account says nothing about anyone's competence. And retention windows often mean the granular geo-level series a reconstruction would need no longer exists.

Where the parties did run a lift test in the relevant period, that is unusually strong evidence and worth asking for in the first document request. Where they did not, no expert can supply one now, and a claim to have "measured incrementality" from historical reports deserves a close look at what was actually done.

The record set I ask for first

Diagnosis is only as good as the inputs, and most of these have a clock on them. In rough order of what I request:

• Analytics exports at daily granularity with the extract date recorded, plus the property's configuration history — filters, bot exclusion settings, tag deployment, consent banner changes.
• Search Console performance data, which has its own limited retention window, exported rather than screenshotted.
• Ad account exports: campaign and ad-level performance, search terms, the conversion action configuration showing which attribution model was in force, and the change history, which logs who changed what and when.
• Server logs, where they exist, because they record requests the analytics tag never saw.
• The deploy and release log for the site, and the CDN or uptime record.
• The price and promotion calendar, and inventory or stockout records for the products at issue.
• Spend by channel across everything, not just the disputed channel.
• Auction insights, impression share and ad library captures for the competitors named.

Preservation matters more here than in most matter types, because several of these are held on rolling windows by third parties and are gone rather than merely inconvenient once the window passes.

What a diagnosis cannot settle

Rankings are not traffic, traffic is not conversions, and conversions are not profit. Each step is a separate inferential leap needing its own support. A ranking drop can coexist with flat revenue, and a traffic drop can coexist with rising profit if the lost traffic never converted. Moving from a rank change to a revenue figure without evidencing each link builds a chain with three unsupported joints.

Absence of evidence in a platform export is not evidence of absence. Retention windows, sampling, thresholding and modeled-versus-observed distinctions all remove information. That a query does not appear in a search terms report does not establish that nobody searched it.

In an agency or vendor dispute, no analysis establishes what a competent alternative would have achieved. The but-for world is a campaign that never ran. I can say whether the work conformed to the contract's scope and to documented practice, and identify specific failures and their mechanism. Asserting the revenue a different vendor would have produced is a counterfactual claim without a counterfactual.

And the diagnosis stops short of the number. Establishing that a decline occurred, when it began, and which explanations survive is where this work ends; converting that into lost profits requires margin, cost structure, mitigation and discounting, which is a damages expert's province. The two halves should meet at apportionment and not blur into each other.

Frequently Asked Questions

Published 2026-08-12