Algorithm Updates as a Damages Confound

Why this is an admissibility question after December 2023

For years, the argument that an expert failed to account for a Google algorithm update was treated as a matter of weight, for cross-examination. That framing is no longer safe.

Federal Rule of Evidence 702, as amended effective 1 December 2023, requires the proponent to demonstrate that it is more likely than not that the testimony rests on sufficient facts or data and that the expert's opinion reflects a reliable application of the principles and methods to the facts. The rewrite of subsection (d) shifted the focus to whether the opinion tracks what the method can support; it targets overstatement. The Advisory Committee was explicit: many courts had held that the sufficiency of an expert's basis and the application of the methodology were questions of weight and not admissibility, and "[t]hese rulings are an incorrect application of Rules 702 and 104(a)."

Apply that to the standard search causation opinion. The method is a comparison of timelines; the opinion is that the defendant caused the loss; the gap between them is everything else that moved the series in that window, the largest of which is usually a change Google made.

The closest authority is Concord Boat Corp. v. Brunswick Corp., 207 F.3d 1039 (8th Cir. 2000), where a damages model was held inadmissible because it "failed to account for market events that both sides agreed were not related to any anticompetitive conduct." The opinion is available from Justia. A core update is such an event: dated, market-wide, and not the defendant's doing.

How update dates are established

The dates are not a matter of opinion. Google publishes a ranking-updates history on its Search Status Dashboard, with its own start dates and rollout durations, and that is the neutral artifact both sides work from. Recent entries:

March 2024 core update — 5 March 2024, 45 days
March 2024 spam update — 5 March 2024, 14 days 21 hours
August 2024 core update — 15 August 2024, 19 days 4 hours
November 2024 core update — 11 November 2024, 23 days 13 hours
March 2025 core update — 13 March 2025, 13 days 21 hours
June 2025 core update — 30 June 2025, 16 days 18 hours
August 2025 spam update — 26 August 2025, 26 days 15 hours
December 2025 core update — 11 December 2025, 18 days 2 hours
March 2026 core update — 27 March 2026, 12 days 4 hours
May 2026 core update — 21 May 2026, 11 days 21 hours

Two features of that record do real work. First, rollouts have duration: a 45-day rollout means an inflection anywhere inside a six-week band is consistent with the update, and placing the conduct's effect on one day inside that band claims a precision the data does not carry. Second, updates overlap: the March 2024 core update began on the same day as the March 2024 spam update, so any decline in that window is confounded by two simultaneous Google changes before anyone reaches the defendant.

Confirmed updates versus unannounced change

The dashboard records the changes Google chose to confirm. It is not a changelog of the ranking system: Google ships continuous change it does not announce, and no public record dates it.

That distinction is where careless testimony gets made.

A confirmed, dated update is a fact. It has a name, a start date, a duration, and a publisher, and it can go in an exhibit.
An observed, undated market-wide movement is an inference from data. It can be shown, but it should be described as what it is rather than given a name it was never assigned.
A rumored update is neither. Industry chatter about an unconfirmed change is not evidence, and a witness who calls it an update on the record has handed the other side an easy cross.

The consequence runs both ways. The absence of a dashboard entry is not evidence that nothing changed; it is the absence of a confirmation. Neither side gets that asymmetry for free. A defendant cannot establish an algorithmic cause by pointing at the possibility of unannounced change, and a plaintiff cannot foreclose one by pointing at an empty dashboard. It has to be shown in data.

Volatility trackers are corroboration, not proof

Several vendors publish daily ranking-volatility indices, tracking positions for a fixed keyword set from fixed locations and devices and reducing the day's movement to one number. When that number spikes, the industry says an update happened.

Such an index supports one thing: that the vendor's panel recorded unusual movement on particular dates, which corroborates a dated update or suggests an undated one. The limits are the list to have ready:

the keyword set is the vendor's, not the party's, and often not in the party's category;
the locations and devices sampled are not the party's users', and the sampling frequency determines what movement is visible at all;
the weighting is proprietary and not auditable by the opposing expert or the court;
the index measures the panel, and says nothing about whether this site moved, or why.

Third-party tracking data as the sole quantitative basis for a damages opinion is a real weakness, and Rule 702(b) is where it is felt. Alongside first-party data and a control cohort it is useful. As proof that an update caused a claimant's loss, it is an argument in a chart.

Separating the conduct from the update

This is the analytic core. Six moves, in the order I would run them.

Fix both dates first. The conduct's date to the day from the documentary record, the update's window from the dashboard. Sometimes they are far enough apart that the question resolves itself; sometimes they overlap so completely that the data cannot separate them.
Use the footprint. Conduct-caused losses are usually confined to what the mechanism touched — the redirected URLs, the disallowed directory, the changed template — while a broad core update tends to move a site by topic across the whole domain. Comparing the affected cohort against an untouched one is the strongest single test available, because the update hits both and the conduct hits one.
Separate the metrics. A core update typically moves average position while impressions persist; a noindex removes impressions entirely; an interface change can leave both intact while clicks fall. Reporting sessions alone destroys the information that distinguishes these cases.
Bring in the market. If comparable competitors moved on the update dates and the claimant moved further or earlier, the excess is the quantity in dispute.
Read the party's own change history. Deployment logs, release notes, revision history, crawl comparisons. In my experience the most common thing pleaded as an algorithm update turns out to be a deployment.
Test the reversal. Where a defect was remediated, did the affected cohort recover on a schedule consistent with recrawling while the control did nothing? A reversal outside any update window is a stronger inference than the decline.

How the defense raises it, and what makes it land

The confound is a defense argument before it is anything else, and which version is made matters.

The version that does not work is the general one: Google changes constantly, therefore causation is unknowable. It proves too much, would defeat every search claim ever brought, and experienced counsel treat it as noise.

The version that lands is specific and dated. A confirmed update whose rollout window contains the claimed inflection. Untouched pages that fell by a similar proportion. A competitor set that moved on the same dates. A pre-trend showing the decline began before the conduct.

The questions that carry it are narrow:

Which confirmed updates overlapped your damage period, and what did you do about each?
Did you consult the published update history, on what date, and is that in the report?
Did the pages the defendant did not touch move, and by how much?
Did you run your model across a date on which nothing happened?
Can you state the portion of the decline your analysis leaves unexplained?

An expert who cannot answer the last one has not done the analysis, and that single answer tells you most of what you need to know about the opinion.

How a claimant answers it

The answer is not denial, which is what makes the confound fatal instead of manageable.

The answer is decomposition: concede the update, quantify what it did to this site using a control, subtract it, and claim the residual. An expert who says "the update moved the control cohort by this much, and the incremental movement in the affected cohort is what the opinion attributes to the defendant" is offering a testable opinion with a stated error. One who says the update was not the cause and offers no measurement of it has, after the 2023 amendment, an opinion that overstates what the method supports.

Two further points a claimant's expert often misses. An update can be the mechanism rather than the confound: where the conduct left a site exposed to a spam update or a policy change, the update is the causal pathway, and the opinion is that the conduct made the site vulnerable to a foreseeable event.

And timing can run the claimant's way. If the affected cohort declined on the conduct's date, the update window opened three weeks later, and the untouched cohort did not move until then, the sequence itself separates the two causes.

The slow confound, and saying when the data cannot answer

One confound has no start date on any dashboard. The relationship between ranking position and traffic — the thing every search damages model implicitly holds fixed — has been changing as answers are increasingly delivered on the results page rather than through a click. A model with a 2023 baseline and a damage period running into 2026 spans that change, and treating the relationship as stationary charges an industry-wide shift to the defendant. It is the same Concord Boat error as an unaddressed core update, only slower.

Which leads to the last point, and the one I would most want a retaining attorney to hear. Sometimes the correct answer is that the available data does not separate the defendant's conduct from a concurrent algorithmic change. That is a result, not a failure — and under amended Rule 702(d) it is a safer position than an opinion claiming the separation without performing it, because a stated limitation is a limitation while an unstated one is an overstatement.

Frequently Asked Questions

How do you show a Google algorithm update, rather than the defendant, caused a traffic loss?

By dating both events and then comparing footprints. The update window comes from Google's published ranking-update history; the conduct's date comes from the documentary record. The test that does the work is a comparison between the pages the conduct touched and comparable pages it did not: an update moves both cohorts, while the conduct moves only one. Competitor movement over the same dates and the party's other channels supply further controls. A general observation that Google changes frequently establishes nothing on its own.

Where do the dates of Google algorithm updates come from?

Google publishes a ranking-updates history on its Search Status Dashboard, giving its own start dates and rollout durations for confirmed core, spam, and related updates. That is the neutral record and it is what belongs in an exhibit. Trade coverage and vendor commentary are secondary and often disagree about dates. The durations matter as much as the start dates: several recent core updates ran two to three weeks and one ran forty-five days, so an inflection anywhere inside that window is consistent with the update.

Are rank volatility trackers reliable evidence that an update occurred?

They are corroboration, not proof. A volatility index measures a vendor's own fixed keyword set, from fixed locations and devices, on a fixed sampling schedule, with proprietary weighting that no opposing expert can audit. It can show that unusual movement occurred on particular dates across that panel. It cannot show that a particular site moved for that reason, and it is not a substitute for the party's own first-party data. As the sole quantitative basis for a damages opinion, third-party tracking data is a genuine weakness under Rule 702(b).

What if Google never confirmed an update in the relevant window?

The absence of a published entry is the absence of a confirmation, not proof that nothing changed — Google ships continuous change it does not announce. Neither side gets to rely on that asymmetry rhetorically. A defendant asserting an unannounced algorithmic cause has to show it in data: comparable domains moving together on the same dates, and the claimant's own untouched pages moving too. A claimant cannot foreclose the argument merely by pointing at an empty dashboard, and should expect the data version of it.

Can a claim survive when a core update ran during the damage period?

Yes, if the update is measured rather than denied. The workable approach is decomposition: date the update from the published record, estimate its effect on this site using an unaffected page cohort or a competitor control, subtract that effect, and claim the residual. An opinion that quantifies the update and claims what remains is testable and states its own error. An opinion that ignores an overlapping update, or asserts without measurement that it had no effect, is the version most exposed after the 2023 amendment to Rule 702.

Is failing to account for an update a weight problem or an admissibility problem?

It is now properly an admissibility problem. Rule 702 as amended in December 2023 requires the proponent to show it is more likely than not that the opinion reflects a reliable application of the method to the facts, and the Advisory Committee stated expressly that treating the sufficiency of the basis and the application of the methodology as questions of weight rather than admissibility was an incorrect application of Rules 702 and 104(a). Concord Boat is the parallel: a damages model was excluded for omitting market events both sides agreed the defendant did not cause.

How long does a core update take to roll out, and why does that matter?

Google publishes the duration alongside the start date, and recent core updates have run from roughly six days to forty-five. It matters because duration sets the precision available. Where a rollout spans six weeks, an inflection anywhere inside those six weeks is consistent with the update, and an opinion attributing the change to a single day inside that band is claiming more precision than the record supports. Overlapping updates compound the problem: in March 2024 a core update and a spam update began on the same day.

Published 2026-08-11

Algorithm Updates as a Confound