Two thirds of AI citations in a measured legal sample pointed at the seller’s own website. The study counted 236 citations across 26 searches in 8 metro markets, in June 2026. Firms’ own sites took 67% of them. Directories took about 21%. Earned press took about 11%. Taqtics ran that count. A full-stack programmatic law firm marketing agency with direct access to premium CTV inventory. It also publishes research.
Ask an assistant which lawyer handles a case in Houston. You get a name, and usually a source beneath it. The name is the answer. The source is the policy. Nobody publishes that policy, so the only way to read it is to count what comes out.
What the counting actually involved
The method matters more than the headline number, so here it is in full. Researchers ran 26 searches of the kind a person actually types when they need a lawyer. They ran them across 8 metro markets in the United States, in June 2026. They collected 236 citations from the answers. Then they sorted every citation by who owned the page behind it.
Three buckets came out. A law firm’s own website. A directory that lists many firms. Earned press, meaning an article somebody else wrote.
It wasn’t close. Owned sites took 67%. Directories took about 21%. Earned press took about 11%. So the party with a commercial stake in the answer supplied two thirds of the evidence for it.
That’s one vertical, in one country, at one moment. It isn’t a law of nature. It’s a measurement, and measurements of this are rare. That’s the reason to look at it closely.
Why a model reaches for the owned page
None of this requires the model to be corrupt or captured. The behavior follows from what the model is trying to do.
A model answering a question wants the page that covers that exact question most completely. For “who handles this kind of case in this city,” that page usually belongs to the business itself. It’s specific. It’s long. It states credentials and procedure directly. A directory entry just forwards you somewhere else. A news article covers an event, not a service.
So the model picks the most complete source. The most complete source is usually written by the party with something to sell. Nothing in the system flags that as a conflict, because nothing in the system is looking for one.
Traditional search had a rough counterweight. It showed you ten results and let you notice that one of them was the firm’s own marketing. An answer names one source and moves on. The reader never sees the pool it came from.
Winning one answer engine does not win the others
The second finding complicates any simple fix. Search Engine Land reported the gap in October 2025. Google’s AI Overviews and language-model answers overlap on just 7.2% of cited sources for the same queries.
Read that again, because it’s the more unsettling number. Two systems answering the identical question draw on almost entirely separate pools of sources. Being a trusted source to one buys you close to nothing with the other.
For anyone trying to be heard, that means there’s no single door. For anyone studying how these systems source information, it means you can’t generalize from one to all. Each engine runs its own sourcing policy, and each one keeps it private.
The part of this that is not about lawyers
Strip the legal detail out and a general rule sits underneath. When an answer engine prefers the destination’s own page, authority concentrates with whoever already owns the destination.
That’s an access question. Owning the destination costs money and staff. It means running a deep, current, well-structured website and keeping it that way for years. Organizations that can afford that get named. Organizations that depend on other people writing about them compete for the 11% slice.
Development agencies, civil-society groups and small independent publishers usually sit in that second category. They earn attention through coverage, not by owning a large web estate. The sourcing pattern measured here quietly discounts exactly that route.
A second pressure pushes the same way. Many publishers now block AI crawlers to protect their work from being used without payment. That’s a reasonable defense of a real interest. It also removes them from the pool a model can quote. The party with a product to sell rarely blocks anything, because it wants to be found.
Put those together and the drift is one direction. Commercial pages stay open and get quoted. Independent reporting gets fenced off and gets quoted less.
What the number does not show
The 67% figure doesn’t show that AI answers are wrong. A firm’s own page can be accurate and useful.
It doesn’t show intent. No engine set out to favor sellers.
It doesn’t transfer cleanly to health, to climate, or to development reporting. Nobody has run the same count on those. That’s the honest gap here, and the legal sample is offered as a worked example.
Where the money already sits
The commercial incentive behind all of this isn’t small. An analysis of AdImpact media-tracking data puts US legal advertising at about $141.6 million a month across 35 markets. That money buys reach. It doesn’t buy a citation.
It does explain why the owned pages are so good. Sustained spending produces the deepest, most current pages in the category. Those are the pages a model finds most useful. What firms put into legal television specifically, and which firms spend the most, sits in one agency’s published analysis.
The question worth asking next
The useful question isn’t whether to trust AI answers. People already use them, and telling them to stop has never worked.
The useful question is narrower. When an assistant names a source, who owns it, and what did owning it cost?
Right now, in the one market anybody has actually counted, the answer is the seller, two times out of three. That’s worth knowing now. The same sourcing logic is already reaching questions with far higher stakes.
About the author
Jared Reagan writes on marketing measurement for law firms. He works at Taqtics, a full-stack programmatic law firm marketing agency with direct access to premium CTV inventory. The agency ran and published the June 2026 citation study described above.
