The newest generation of research tools does something the first generation did not: it goes and gets sources. A deep-research agent runs dozens or hundreds of searches, reads what it finds, and hands back a report with footnotes that link to real pages. The links open. That is a visible improvement over a chatbot inventing a case name, and it is easy to mistake for the problem being solved.
A study published in May 2026 measured exactly that gap, and the measurement is the most useful number a lawyer can carry into a vendor meeting this year [1].
What was measured
Six researchers built a framework that reads a deep-research report the way a clerk would: it extracts every inline citation with a parser, opens the cited page, and scores the pair on three questions. Does the link work? Is the page about the thing the sentence says? Does the page actually support the claim the sentence makes? They ran it over reports from fourteen models, closed and open [1].
The three scores come apart. For the strongest frontier models, links worked more than 94 percent of the time and the cited page was topically relevant more than 80 percent of the time. The cited page supported the claim it was attached to between 39 and 77 percent of the time, depending on the model [1].
So a reader who checks that the links open will find almost all of them do. A reader who checks that the page is on topic will usually find that it is. A reader who reads the page and asks whether it says what the footnote claims will find, for the best model in the study, roughly one claim in four unsupported, and for the weakest, more than half.
More searching made it worse
The second finding is the one that should end the “just give it more retrieval” argument. The authors scaled the number of tool calls an agent could make, from 2 up to 150, and watched the fact-check score. Across two frontier models it fell by about 42 percent on average as the research got deeper [1]. More retrieved material produced more citations, and a smaller share of them were true to their source.
The mechanism is not mysterious. The agent is still a language model writing a sentence and then attaching a source it saw. Give it more sources and it has more to attach; nothing in the loop compares the sentence to the page before the footnote is placed. The link is a locator. It tells you where the model was looking. It does not tell you the model read correctly.
This is the legal result, repeated
If the shape of this feels familiar, it should. In 2024 Stanford’s RegLab measured the leading retrieval-based legal research products and found them hallucinating between 17 and 33 percent of the time, and the authors drew the same distinction: a response can be “misgrounded,” meaning it “describes the law correctly, but cites a source which does not in fact support its claims” [3]. Two years and a generation of tools later, the deep-research study finds the same failure in the same place, with the citation now carrying a working hyperlink.
That is why the duty in ABA Formal Opinion 512 does not get lighter when the tool gets better at retrieval. The opinion asks the lawyer to verify the output, and verification means opening the source and reading it against the sentence, which is the one step none of these systems performs [4].
What a check would have to do
Read the study’s three questions again, because they are the specification for a real check. Link works: trivial, and the tools already pass it. Content relevant: a model can score this and the study used models to do so; a follow-up paper found that cheaper models score it about as well as frontier ones [2]. Fact supported: this is the one that matters, and it is the one the writing model cannot be trusted to grade for itself, because it is the same faculty that placed the footnote.
A citation gate does the third check as code, not as a judgment: the exact quoted words must appear in the resolved source, byte for byte, or the line does not render [see related]. A claim that paraphrases a page it did not quote is held, with the reason stated, rather than shipped with a link that opens. That is a stricter standard than any of the fourteen models met, and it is the standard a court applies when it reads the footnote.
The number to keep
Over 94 percent of links work. Between 39 and 77 percent of cited sources support the claim. The distance between those two figures is the work a lawyer is still doing by hand after the tool is done, and the work a check outside the model has to do before the tool’s output can be signed.