Cross-Examining a Search Expert

What you are actually looking for

You are not looking for a witness who cannot defend search optimization as a discipline. You will not find one, and the attempt makes you look unprepared. You are looking for the distance between the analysis performed and the sentence written at the end of it.

That distance is visible in the produced material rather than in the prose, so the preparation that matters is having someone work the other expert's files: re-run the crawl, pull the same Search Console range if your client has access, recompute the charts from the exports, and see whether the numbers come back.

Below are seven lines, the questions that open each, and what a good answer sounds like. That last part is not a courtesy — knowing which answers should satisfy you keeps a deposition from wandering, and tells you early whether you are heading for a motion or for a cross-examination.

Line one — where the data came from

Most quantified search opinions rest on numbers the witness never measured. This line establishes whether he knows that.

1. List every data source you used, and say who obtained each and when.
2. Did you use my client's own Search Console, analytics, or server log data? If not, why not?
3. This figure — a measurement of something that happened, or a vendor's estimate?
4. How does the vendor compute this traffic estimate? What inputs go into it?

What a good answer sounds like. The witness distinguishes measurement from model without prompting, says plainly that the third-party figure is an estimate from panel data and inference, explains why he used it — no analytics before a certain date, or a competitor's site he cannot measure — and identifies which conclusions rest on measured data.

What a bad answer sounds like. Treating a rank tracker's position as what users experienced, or not knowing how the estimate is produced. Best of all: conceding that Search Console data for the period existed, was available, and was never requested.

Line two — the confound

Search traffic moves for reasons unconnected to your client. Where the report does not address them, this is the line most likely to support a motion rather than merely a cross.

1. Name every broad core update that Google announced during your damage window. Give me the dates.
2. What did you do to test whether any of those updates explains part of this decline?
3. What was this site's traffic in the same calendar period in each of the three preceding years, and did other sites in the sector move in this one?
4. Was the analytics tracking altered during the period you measured — including in the deployment you are criticizing?

What a good answer sounds like. He names the updates and their dates without notes, describes what he examined to test each — a peer cohort, the site's own unaffected sections, the shape of the decline against the rollout — and concedes that one may account for some share of the loss, which is why his opinion is stated as it is. An expert who volunteers the confound he could not exclude is harder to damage than one who says nothing happened.

What a bad answer sounds like. Not being able to name an update in the window, or saying updates were considered without saying what was done. An assertion that the update did not affect this site because the decline started earlier is worth pursuing: it is testable, and often wrong on the dates.

Line three — how the affected pages were chosen

Nearly every search damages analysis rests on a cohort: URLs said to be affected, measured against URLs said to be unaffected. The expert chooses it, rarely explains it, and the result often does not survive a different choice.

1. How did you decide which pages were affected by the conduct?
2. Was that rule written down before you looked at the traffic data, or after?
3. What happens to your result if the cohort is defined by template instead of by directory? By date of change instead of by section?
4. Are the two groups comparable — similar traffic levels, similar query types, similar seasonality — before the period at issue?

What a good answer sounds like. A criterion stated in advance and applied mechanically — every URL under a stated path, served by a stated template, or appearing in the redirect map — with the lists produced, the exceptions named, and one alternative cut tested and reported.

What a bad answer sounds like. A cohort assembled by inspection, meaning the pages that fell were called affected. That is circular, and once on the record it is fatal to the comparison, because the control group has been defined as the pages that did not exhibit the effect being measured.

Line four — the date of the standard

Search guidance changes. An expert measuring conduct from one year against documentation published later is measuring against a standard that did not exist, and this line is badly under-used.

1. What document establishes the standard you say the defendant failed to meet?
2. What is the publication date of the edition you used?
3. Did you review the edition current at the time of the conduct? Did you compare them?
4. Identify every material difference between the two editions on the points relevant to your opinion.

What a good answer sounds like. The witness has the contemporaneous edition, has compared it with the later one, and can state whether the guidance on the point changed. That is the analysis the court performed in Campmor, Inc. v. Brulant, LLC, No. 2:09-cv-05465 (D.N.J. Apr. 23, 2013) (ECF No. 111), where the objection was rejected after the court reviewed both editions and found little substantive difference.

Read Campmor carefully before relying on it either way. It is not authority that using current guidance for old conduct is acceptable; its reasoning was comparative, and where guidance has materially changed between the conduct and the opinion the same reasoning cuts the other way. An expert who has not made the comparison has not established his standard.

Line five — the causal leap

The technical work is often sound and the causal sentence often is not. The most valuable question in this line is the one about disconfirmation.

1. What is the analytical step between your observation and your causal conclusion?
2. If your theory were wrong — if something else caused this — what would you have expected to see in the data?
3. What proportion of the decline do you attribute to the conduct? How did you compute the proportion?
4. Was there any point in your analysis at which the data could have told you that you were wrong?

What a good answer sounds like. He states a mechanism rather than a coincidence — these URLs were removed from the index on this date by this directive, and the traffic they lost is the traffic that disappeared — and he can describe what would have falsified it: a decline in sections the change never touched, a decline beginning before the deployment, or a peer set that fell equally.

What a bad answer sounds like. “In my experience, this is what causes that.” The formulation to have in mind is from Zenith Electronics Corp. v. WH-TV Broadcasting Corp., 395 F.3d 416 (7th Cir. 2005): a witness who invokes “my expertise” rather than analytic strategies widely used by specialists is not an expert as Rule 702 defines the term, and reliable inferences depend on more than say-so.

Line six — what was considered and never produced

Rule 26(a)(2)(B)(ii) requires the report to contain the facts or data considered — not relied on. For an expert who runs crawls, pulls exports, and reviews material from counsel, that is a large obligation, commonly under-satisfied, and the gap between what was considered and what was produced is where discarded analyses live.

1. List every crawl you ran on this site, including the ones you did not use.
2. What did counsel send you — spreadsheets, summaries, figures?
3. Did counsel supply any number that appears in your report, or any assumption you relied on?
4. Did you reach a preliminary conclusion different from your final one? What changed it?
5. Is everything you considered in the production? Point me to it.

What a good answer sounds like. A complete list matching the production, with any counsel-supplied figure identified in the report as a stated assumption. Note the limits before you press: Rule 26(b)(4)(B) protects drafts of the report regardless of form, and Rule 26(b)(4)(C) protects attorney–expert communications except as to compensation, attorney-supplied facts or data the expert considered, and attorney-supplied assumptions the expert relied on. Those carve-outs are the target; the drafts are not.

What a bad answer sounds like. Crawls or exports that exist and were never produced, or a report figure that originated in a spreadsheet from counsel and is identified as an assumption nowhere.

Line seven — the number the machine produced

Where an exhibit contains a score — a toxicity rating for a link, a similarity percentage, a visibility index, a classification of reviews as fabricated — the witness has adopted somebody else's model as part of his method. Very few can describe it.

1. Which system produced this number? Name it. What version?
2. What exactly does the score measure? Who defined that?
3. What inputs does it take, and what does the vendor publish about how it is computed?
4. If you ran it again today, would you get the same numbers? Have you tried?
5. Was the tool validated against known examples for this task, on material like this?
6. Where a language model was used: what was the prompt, verbatim? What settings? How many times did you run it, and did you select among outputs?

What a good answer sounds like. The model and version named, the vendor's published description given accurately, a validation described — a sample from the same population labeled by a human under a written rule, compared against the model's output, with false positives and negatives reported separately — and the raw outputs produced. That is a defensible use of an automated tool.

What a bad answer sounds like. “It is the industry standard tool.” That is not a description of a method, and where the score carries a conclusion, this line is the one most likely to justify a motion rather than a cross.

Reading the answers, and knowing when to stop

Three judgments come out of these lines, worth making explicitly.

Motion or cross? A missing step is motion material — no stated sources, a cohort defined by the effect, a causal opinion with no elimination work, a score with no validation. A contestable choice is cross material — a comparison period you would have picked differently, a confound weighted less than you would weight it. Moving on the second kind hands the other side a reliability ruling and tells them what to repair.

How much to take at the deposition. Answers headed for a motion must be complete and closed on the record, which means asking the follow-up that forecloses the later explanation. Answers headed for cross are better left partly untaken, since a full airing gives the witness months to prepare a better version.

When you have enough. A witness whose report labeled every figure as measured or modeled, named the updates in the window, stated the cohort rule in advance, compared the editions, described what would have falsified his theory, produced everything he considered, and validated any automated score has given you a strong report. Recognizing that early is worth more than a long examination; the remaining work is on scope. Whether any of it supports a motion is a judgment for counsel, better made with someone who can work the other expert's data than only read his prose.

Frequently Asked Questions

Published 2026-08-11