The Answer Engine Recommends a Lawyer to People Who Only Asked a Question

Ask an AI answer engine what the statute of limitations is on a workers’ compensation claim, and in roughly one out of five cases it will hand you the name of a law firm you did not ask for. That number comes from a stratified random sample of 1,000 legal queries run by Herringbone in August 2026: 448 phrased as informational, 552 as commercial, spread across personal injury, workers’ compensation, Social Security disability, and veterans’ claims in shared Southeastern US metros.

Across the whole sample, 59% of answers recommended at least one specific provider. When the question carried commercial intent, that rose to 92%. When the user only wanted to learn something, it was 18%.

Two caveats before the analysis, because they change how much weight the numbers carry. Herringbone is a legal-marketing roll-up (BluShark Digital, Hennessey Digital, cj Advertising) and the study’s author, Dan Hinckley, is its Head of AI. The publisher sells into the exact market this study measures, and every finding here points toward “law firms need to care about answer engines.” Second, the study did not query AI Overviews directly. It queried a Gemini flash model with live Search grounding, which the author describes as “likely the same retrieval substrate behind AI Overviews.” That is the author’s own hedge, and it is the right one. This is a proxy measurement, not a measurement of the production surface.

With both caveats applied, the structural findings still hold, because they describe behavior any team can reproduce.

The 18% Is the Finding

Commercial queries producing commercial answers is not news. Somebody typing “best workers comp lawyer near me” has declared intent, and 92% recommendation rate on that population is the system doing what it was asked.

The informational 18% is different. A person asking how a process works receives, unprompted, the name of a business that sells services in that process. In a regulated advice domain. From a system that presents itself as an explainer.

The breakdown by answer style makes the mechanism clear. Informational queries produced 82% purely informational answers, 18% mixed answers, and zero pure-recommendation answers. Commercial queries produced 8% informational, 92% mixed, and again zero pure-recommendation. There is no separate selling mode in this system. The engine teaches, and it names providers while teaching. Recommendation is woven into explanation rather than partitioned from it, which means a user has no surface cue telling them the register just changed.

Every disclosure regime built for advertising assumes the two registers are separable. You label the ad. You mark the sponsored result. You cannot label a clause inside a paragraph that is otherwise an accurate explanation of a filing deadline.

Six Searches Nobody Sees

Behind each of those 1,000 questions, the system issued an average of 6.4 searches of its own. Herringbone counted 6,436 fan-out queries in total. None of them appear to the user, and none of them appear to the firm being evaluated.

The composition matters more than the count. Among the fan-outs generated from commercial-intent questions, 18% were informational research and 24% were navigational reputation checks, hitting named firms, directories, and review platforms. Roughly a quarter of the invisible work is the engine going to look up whether a business is any good, using sources the business does not control and cannot see being queried.

A firm that gets named has been screened. A firm that never gets named has probably also been screened and rejected, silently, on inputs nobody logged. There is no impression record, no rejection reason, no appeal. For a regulated profession where referral and advertising conduct is governed by bar rules, that is an audit trail with a hole in the middle. The profession has spent two years arguing about AI agents doing legal work at scale. The engine that routes clients to the firm arrived without the same scrutiny.

We have written before about adversarial manipulation of the answer layer, where an attacker poisons what the engine says about you. This is the other half. No adversary is required. The engine’s ordinary operation already makes commercial determinations about named businesses on evidence that is invisible to every party involved.

The Concentration Everyone Expected Did Not Happen

Here the study corrects an assumption we have argued from before. In the four separate games of answer engine optimization, we treated heavy citation concentration as the governing dynamic: a small number of sources capture most of the visibility, and winner-take-most is the shape to plan around.

The legal vertical does not look like that. The recommendation shortlist averages 7.6 firms per answer, roughly double a Google map pack. Across the sample, 936 firms competed for 3,670 recommendation slots. The leader holds 3.7%. The top five hold 11.9%. The top ten hold 18.4%.

The curve is a long tail with none of the power-law shape. Whether it is a property of this vertical, of the proxy model, or of the fact that legal services are intensely local, the study cannot say. What it does say is that the concentration thesis needs a scope condition attached to it. It described citation behavior in a class of informational queries. It does not describe recommendation behavior in a local services market, at least not here.

One more number breaks the marketing frame entirely. Of the 588 answers that recommended anything, 108 named only free public resources and no law firm at all. Government agencies and legal-aid nonprofits take 16% of all recommendation slots across the sample. Those are slots no budget can buy. An entity with no marketing function, no AEO strategy, and no commercial interest is occupying a sixth of the recommendation surface in a market where firms spend heavily to be found.

What the Method Is Worth Copying For

Set the findings aside for a moment. The measurement design is the part most teams should steal.

Herringbone ran three passes. Ask the question. Extract named entities using a second, independent model pass, with entities merely discussed in passing excluded from the count. Classify every fan-out search by intent type. Then anonymize the firms as Company A through L before computing slot share, so the analysis of concentration could not be steered by who the winners were.

The second-pass extraction is the discipline that matters. Asking one model to both generate and self-report which providers it recommended would have produced a number, and that number would have been unfalsifiable. Splitting generation from extraction is the difference between an audit and an anecdote.

Anonymization before the concentration analysis is the other one. A legal-marketing firm computing market share for a list that includes its own clients has an obvious pull toward a favorable read. Removing the names before the math removes the pull. It does not remove the conflict of interest, and it does not remove the incentive to publish a study whose conclusion is “you need what we sell.” It does make the specific concentration numbers harder to have massaged.

Run This Audit on Your Own Category This Month

What transfers here is the method, and the method is cheap: nobody in your category has run this audit yet.

Pick 100 questions a customer would actually type, split roughly evenly between informational phrasing and commercial phrasing. Run them through the grounded model you can access, and log every answer. Then measure four things.

How often does a specific provider get named on a purely informational question? That is your unrequested-recommendation rate, and it tells you whether your category’s answer layer is behaving as an explainer or as a broker.

How many providers appear per recommending answer, and what share does the leader hold? If it is 7.6 and 3.7%, presence beats dominance and the strategy is to be eligible everywhere. If it is 2 and 40%, you are in a different market and the calculus inverts.

What fraction of recommendation slots go to entities that are not competitors: government sites, nonprofits, trade bodies, free tools? That is the ceiling on what any commercial effort can win.

What did the engine search for on your behalf, and which reputation sources did it hit? Those sources are now part of your compliance surface whether or not anyone in your organization has ever looked at them.

For a regulated category, hand the log to whoever owns advertising compliance, not to whoever owns marketing. An unrequested commercial recommendation delivered inside an explanation, sourced from checks nobody can see, is a governance question first. The marketing question comes after somebody has decided the behavior is acceptable at all.

This analysis synthesizes We Asked Google’s AI Overviews 1,000 Legal Questions. Here’s How It Responded (Herringbone, August 2026).

Victorino Group helps regulated organizations audit how answer engines describe and recommend them, and turn the result into a control their compliance function can own. Let’s talk.