Identification of the Plaintiff

Why no expert settles this element

This page is written for counsel weighing whether a technical analyst is worth retaining on whether content that does not name a client refers to the client. An analyst produces evidence going to that question without answering it, and the budget should be set on that basis.

Whether a reasonable reader would understand a statement to refer to this plaintiff is an element, and in most formulations a question for the finder of fact. It is not a measurement: no instrument returns it, no dataset contains it, and no method has an error rate against it. What follows is what the record can put in front of the people who decide it.

What is technical here is narrower and, unusually for this subject, tractable: machine-observable linkage between the content and the plaintiff's identity, and observed reader behavior around it. I can show that a page carries a person's name in its markup, that third parties linked to it using that name, that commenters underneath named the person, that the page appeared in position three for that name on stated dates, and that an image on the page matches a headshot published elsewhere. Each is a fact about the record. None is a fact about a reader's mind, and an expert who closes that gap offers an opinion the record does not support.

When the content names the plaintiff: two traps in a string search

Where the content uses the name, identification is mostly a search-and-capture exercise. Two failure modes are worth naming, because both are common and both are avoidable.

Near-miss naming. Misspellings, initials, nicknames, transliterations, former names, married names, handle-only references — and deliberate obfuscation, where zero-width characters or homoglyphs are inserted inside a name so it reads normally to a human while defeating exact-string search and platform detection alike. An analyst running a plain string search misses all of it, and so does opposing counsel's paralegal. Normalizing text before searching — Unicode normalization, homoglyph folding, stripping zero-width and formatting characters — is a real and defensible methodological step, and its absence is an equally real criticism of any content inventory produced without it.

Identification without a name at all. A photograph identifies. Perceptual hashing and reverse image search can support that the image on the disputed page is the same image as a headshot published on an employer's site, and that is a technical finding with a reproducible method behind it. What it establishes is image identity. Whether the image identified the plaintiff to a reader is the same legal question as before, restated in pixels.

The strongest linkage artifact is the publisher's own markup

The most under-used identification artifact in this field sits in the defendant's own HTML. Structured data — schema.org properties such as about, mentions and sameAs, an author or subject entity, or an identifier field — is a machine-readable declaration by the publisher of who the page concerns. Where sameAs points at the plaintiff's own professional profiles, the publisher has stated the subject in a form requiring no inference from anybody.

That is a different quality of evidence from search behavior. It is not a correlation, a ranking or an estimate. It is the publisher's own statement, present in the delivered source, capturable and hashable at the moment of collection. It is also frequently generated automatically by a plugin, which the other side will point out and which counsel should know before a deposition rather than during one.

The same logic runs through the rest of the delivered page. Title tags, meta descriptions, image alt attributes, the URL slug, category and tag assignments and open-graph properties are all authored fields describing the subject. Preserve them with the page source, because a rendered screenshot contains none of them.

What third parties said the page was about

Three further classes of material go to whether readers connected the content to the person, and all are stronger than anything derived from a search engine.

Anchor text on inbound links. The words a third party chose when linking are that party's own statement of what the content is about. Inbound links whose anchor text is the plaintiff's name are contemporaneous with the publication rather than reconstructed for litigation, which is exactly what a survey cannot be. The caveat is the source: every major backlink index is a vendor crawl with undisclosed coverage, so a count is a floor rather than a total and should be reported as one.

On-page context. Employer, role, license number, practice area, a linked profile, a sentence lifted from the plaintiff's own site. This is close reading rather than technology, and it is often the most persuasive material in the file. It does not need an expert to find it; it needs somebody to sit down with the page.

Reader statements in situ. Comments, replies, quote-posts and forum threads in which readers name the plaintiff are the most direct available evidence that readers made the connection — and the most fragile. They are edited, deleted and moderated continuously, and are frequently script-loaded and therefore absent from archive captures. Capturing the comment thread rather than only the article is the difference between having this evidence and not having it.

What a search result capture actually establishes

The standard offer here is some version of "search his name and this is the second result." It is worth doing, it is measurable, and it is routinely presented in a form that does not survive cross-examination.

A search engine result page is personalized, localized, time-varying, device-varying and session-varying. A capture establishes exactly this: for this query string, from this location parameter, on this device and viewport, in this session state, at this timestamp, the engine returned these results. To be reproducible the evidence has to record all of those, and most result-page screenshots in circulation record none. The query is retyped from memory, the session is logged in to a personal account, personalization was never suppressed, and the capture is of the visible viewport rather than the full page.

A single capture is a single observation. Position is volatile. An opinion about visibility built on one screenshot is an opinion about one moment; a defensible one comes from a time series of repeated, scheduled, identically parameterized observations. Third-party rank-tracking data supplies that series at the cost of putting a vendor's undisclosed methodology into the record, and that trade has to be disclosed.

Autocomplete is the weakest form, and the operator says so

Query suggestions pairing a plaintiff's name with a damaging term make a vivid exhibit and a poor one. Google's own help page, How Google autocomplete predictions work (read 15 August 2026), states it in terms an opposing expert will read aloud: "Predictions aren't assertions of facts or opinions, but in some cases, they might be perceived as such." The same page describes predictions as generated from both real searches and word patterns found across the web, and describes an enforcement process that removes predictions violating its policies. The observed set is a filtered output of an undisclosed system rather than a record of what people searched.

So a prediction is evidence that the system produced that prediction at that moment. It is not a measurement of query volume, not evidence of what any reader believes, and by the operator's own description not an assertion of anything. It is removable and time-varying, so it has to be captured with the same rigor as a result page and re-captured on a schedule to be relied on at all. Offered as proof that the public associates the plaintiff with the allegation, it is an invitation to be corrected using the operator's own documentation.

The two datasets that only discovery produces

Two records would answer the linkage question far better than anything observable from outside, and both sit with the publisher.

Search query data for the disputed URL. A publisher's own search performance reporting shows the queries that produced impressions and clicks for a specific page. If a substantial share of them are the plaintiff's name, that is direct evidence that readers arriving at the page were looking for the plaintiff — an inference no external tool supplies, because no external tool sees the queries. The operator's published limitations travel with the data: queries are omitted to protect user privacy and that behavior changes once a filter is applied, the query table stores only the most significant rows, and reports cover a representative sample rather than a complete listing. The window over which the data stays available is limited, and I check the operator's current documentation rather than repeat a figure from memory.

Referrer data in the publisher's own server logs, showing which pages sent traffic to the disputed URL. Modern referrer policies strip most query strings, so this is weaker than it was, but internal referrals remain visible: a link from a competitor's site, a forum thread, a newsletter. Both datasets expire and neither is preserved by default.

Surveys, and the two limits that are rarely mentioned

Where linkage is contested, the formal instrument for measuring reader understanding is a survey, and the methodological reference in federal practice is the Federal Judicial Center and National Academies Reference Guide on Survey Research in the Reference Manual on Scientific Evidence, fourth edition (2025, read 15 August 2026). Four points carry directly.

• The population is a substantive decision. The guide directs attention to whether the sampling frame lists all members of the target population. In a defamation matter that population is not the public; it is the people whose opinion of the plaintiff matters, and defining it is an argument rather than a technicality.
• Probability sampling is preferred, and a fast, inexpensive survey uses an online opt-in panel, which is a nonprobability sample.
• Wording requires pilot testing wherever there is doubt a term will be clear, and administration should be double-blind, with interviewer and respondent both blind to the sponsor.
• A control group is what makes the design answer this question. Show one group the content and a control group a version with the identifying details altered, then measure the difference in whether respondents name the plaintiff.

Two limits are severe. Timing first: a survey run during litigation measures what people understand now, after the dispute has generated its own publicity, while the element concerns understanding at publication. There is no methodological repair; it can only be disclosed. Second, I could not establish whether US courts have accepted survey evidence on this element at all. The mature body of survey practice in US courts is trademark and false-advertising confusion, and whether it transfers is an open legal question rather than a technical one.

When not to retain anyone on this element

If the content names the plaintiff plainly and spells the name correctly, there is no technical question here. Someone has to capture the page properly and preserve the comment thread with it, and that is collection rather than analysis.

If the dispute is a group reference, a description fitting several people, or an implication drawn from juxtaposition, no technical method resolves it. Those turn on reading and on doctrine, and an expert retained to attack them produces a report stating what the page contains, which nobody disputed.

Where analysis earns its cost is narrower. The content is obfuscated and a normalized search finds material a string search missed. The publisher declared the subject in structured data and nobody has read the source. Third parties linked with the plaintiff's name as anchor text. Commenters named the plaintiff underneath and the thread is still live, which means it can still be preserved.

What the fact-finder will be weighing is whether readers connected the content to the plaintiff. They will not be weighing my opinion about it, because I do not have one the record supports.

Frequently Asked Questions

Published 2026-08-15