Reputational Damages Evidence

What the record carries on its own

This page is written for counsel deciding what a damages analysis in an internet defamation matter can be built from, and what the other side will do to it. The blunt version: the quantum is modeled, never measured. No instrument returns the amount by which a reputation fell, because reputation is not a quantity anything counts. Everything below measures a proxy — traffic, ratings, queries, revenue — and the link between the proxy and reputation is itself an assumption in every available method.

What the record carries without a model is narrower and genuinely useful: exposure and timing. That a URL existed at an address on stated dates and returned content. That the plaintiff's own property recorded certain impressions and clicks for named queries in named weeks. That a rating displayed a value on a date somebody captured it. The step from an observation to a figure is where the assumptions live.

Whether damages are recoverable, whether they are presumed on these facts, and how non-economic injury is treated are questions for counsel. What follows is what each method needs the fact-finder to accept.

The observations that need no model

Before a method is chosen there is a layer requiring no assumption at all, worth exhausting first because it is cheap and it is the part that survives.

• Presence and dates. What existed at which addresses, when it appeared, when it changed, when it stopped resolving.
• Impressions and clicks for the plaintiff's own property, by query, week by week, from the plaintiff's own search reporting. This is the closest thing to a demand signal a party owns rather than estimates, and a branded and non-branded query filter has been available in that reporting since March 2025.
• Documentary evidence of mechanism. A prospect's email saying they read the article, a call recording, a disqualification note in a procurement file. These evidence the mechanism rather than a correlation, and no opposing expert models them away.

The caveats belong with the data and come from the operator: queries are omitted to protect user privacy and that behavior changes once a filter is applied — including the brand filter this analysis needs — and the query table stores only the most significant rows. That does not make the data unusable. It makes it a floor with disclosed properties.

Method one: traffic and conversion, and its four assumptions

The most common approach compares the plaintiff's own traffic, leads or revenue before and after the publication date and attributes the difference. In a deposition the assumption is the question, and there are four.

• That the measurement instrument did not change across the break. Violated most often, checked least often. Analytics migrations, consent banners, tag changes, bot-filter changes and browser privacy releases all produce step changes in measured traffic with no change in real traffic. Nor is it hypothetical: the operator publishes a log of its own reporting errors, including one announced 3 April 2026 affecting impression reporting from 13 May 2025 through 27 April 2026.
• That the pre-period characterizes normal variation and seasonality. A single year cannot separate a seasonal trough from an event effect. Two full cycles is the minimum defensible ask, and analytics reaching back only eighteen months is a finding about the analysis rather than a detail.
• That the funnel is stable — same pricing, same offer, same landing pages, same media spend, same sales staffing. Any of those moving inside the window contaminates the estimate.
• That the observed conversion rate applies to the lost traffic. It usually does not. Traffic lost at the top of a branded funnel is not compositionally identical to traffic retained, so applying the retained rate to the lost population is an assumption and belongs in the report labeled as one.

Method two: rating movement, and the date it did not happen on

Where the plaintiff is reviewed publicly, the obvious series is the displayed rating and its distribution. Three assumptions sit under it, and the third is a trap.

That the displayed rating reflects the reviews submitted. It does not. Review platforms filter and moderate what is displayed and describe doing so in their own documentation. What an observer sees is produced by proprietary logic that changes without notice and operates retroactively.

That the historical series can be reconstructed later. It cannot. Platforms do not publish rating history, a review removed today was never there as far as the current page is concerned, and edited reviews leave no public trace. A rating series exists only if somebody captured it while it moved. An expert engaged afterward is reconstructing from archive captures — where review widgets are often script-rendered and absent — and whatever screenshots the client happened to take.

That the date the displayed rating dropped is the date something happened. Displayed ratings are rounded, usually to half stars, so the displayed value is a step function of the underlying mean. A real change can stay invisible for months and then appear as a sudden half-star drop on a day when nothing occurred. Treating that drop date as the event date mis-dates the event, and everything downstream is then wrong while looking rigorous.

Method three: branded search volume, and the operator's disqualifier

The third approach argues that demand for the brand fell by showing that searches for the brand name fell. The usual instrument is Google Trends, and Google's documentation disqualifies it for most matters of this size. From the FAQ about Google Trends data (read 15 August 2026): each data point "is divided by the total searches of the geography and time range it represents" and is then scaled "on a range of 0 to 100 based on a topic's proportion to all searches." It is a sample rather than a census, and it excludes "searches made by very few people," which display as zero, along with queries containing "apostrophes and other special characters." Then the sentence that decides most of these matters: the data is "not a perfect mirror of search activity," and "Trends data will sometimes reflect statistical noise rather than actual search interest" — a caveat tied specifically to low-volume queries.

A single business name is, in those terms, a low-volume query, and for most plaintiffs the tool returns zeroes or noise. The plaintiff's own search reporting is the better instrument, returning counts for a property the plaintiff controls rather than a normalized index. Third-party keyword tools report modeled figures from proprietary methodologies; in its annual report filed with the SEC, Similarweb calls what it produces "estimated insights." Neither instrument measures demand. Both measure an operator's record of query activity, and that matters precisely when the theory of harm is that people stopped looking.

The three methods side by side

Set out together, each method produces a number that depends on something the record does not contain.

Read down the third column and what the record establishes on its own becomes concrete: exposure and timing. Every figure in this subject is built on the second column.

The confounding problem, stated the way counsel needs it

Sequence is not cause, and for a litigator that does not go far enough: the expert has to name the alternatives and say what was done about each. An opinion that does not is the easiest one in this field to take apart.

The reason is mundane: a business whose traffic fell after a post also changed six other things. In the same two quarters it redesigned the site, moved hosts, paused a campaign, lost a large account, raised prices, and a competitor entered. Each is dateable from the plaintiff's own records, and each will be dated by the other side if it is not dated first.

• Ranking changes originating with the search engine rather than the content. Google publishes a dated log of ranking updates on its Search Status Dashboard (read 15 August 2026). Checking whether the traffic break coincides with a published update is cheap and primary-source, and its omission is a straightforward criticism of an opposing analysis.
• Site changes by the plaintiff — migrations, redesigns, URL changes, indexing directives set wrongly, outages.
• Marketing changes — media spend, campaign pauses, partners.
• Market changes — competitor entry, price movement, category demand, regulation.
• Other reputational events. These cluster, and where a disputed article, a regulatory complaint, a senior departure and a bad quarter land inside one quarter, no method separates them.

Counterfactual models and the control set

The formal toolkit here is difference-in-differences, synthetic control, interrupted time series and Bayesian structural time-series modeling. The last appears most often because an implementation was published with it — Brodersen, Gallusser, Koehler, Remy and Scott, Inferring causal impact using Bayesian structural time-series models, Annals of Applied Statistics 9(1):247–274 (2015), read 15 August 2026 — which predicts "the counterfactual market response in a synthetic control that would have occurred had no intervention taken place."

Its central requirement is a set of control series unaffected by the intervention, and in a defamation matter that requirement fails predictably:

• Competitors: if the content moved demand to them, the controls rose because of the event and the estimate is inflated.
• The plaintiff's other locations or product lines: if the event was firm-wide, the controls fell too and the estimate is biased toward zero.
• Category-level search demand: if the matter was newsworthy enough to move the category, the control is contaminated.

There is no purely statistical repair. The control set has to be justified substantively — argued in the report, with reasons each control was and remained unaffected — rather than selected because it fits the pre-period. Two things should accompany any such model and are usually missing: a placebo test, run on a pre-event date where nothing happened, and a sensitivity analysis across alternative control sets. A point estimate offered with neither has no demonstrated error behavior.

The conclusion nobody sells

The most valuable and least-supplied output in this market is a written conclusion that the data does not support an estimate. It is a legitimate finding, frequently the correct one, and far more robust under examination than a number resting on stacked assumptions.

That conclusion is honestly available in three common situations: where the analytics history is too short to characterize seasonality, where no rating series was captured while it moved, and where several reputational events landed inside one window. The report then states what the record shows — exposure and timing — and what it does not carry.

Where an estimate is supportable the report should look modest: a narrow series on the plaintiff's own property, a documented instrument, a long pre-period, a justified control set, a placebo run, a sensitivity analysis, and a range rather than a point. Whether a range of that kind is enough for the claim is for counsel.

What I will not produce is a spreadsheet multiplying a click-through curve by a modeled search volume by an assumed conversion rate by an order value by twelve months. Four uncertainties multiplied together do not become one confident figure.

Frequently Asked Questions

Published 2026-08-15