In one sentence
The new deep research module does not merely name its sources — it actually retrieves every single one, archives it and looks for the asserted quotation on the page; whatever fails is removed from the opinion instead of remaining as a footnote.
The problem with named sources
A source reference beneath a legal text creates an impression it often does not honor. It looks like proof, but to begin with it is only an assertion about proof — someone states that something specific is to be found at that address.
Three things about that can be wrong independently of one another. The address may lead nowhere. The page may exist but say something other than what is asserted. And even if everything was accurate at the time of writing, the page may have moved, been rewritten or vanished six months later — at which point the reference is worthless, precisely at the moment when someone wants to verify it.
For a piece of continuous prose that may be tolerable. For a document that goes into a pleading or out to a client, it is not.
Four checks per source
Deep research therefore treats every source found on the web as something that first has to prove itself:
- Actually retrieve it. Not name it, but call it up — with a check of the response status and the content type. A page that does not respond, or that delivers something else entirely, is not a source.
- Archive it. Every reachable source is preserved, where possible, in the Internet Archive; the archived version stands next to the original address in the list of sources. That keeps the reference verifiable even if the original later disappears.
- Look for the quotation on the page. A verbatim quotation counts as substantiated only if it was in fact found there.
- Sort it out. Whatever fails does not remain as an inconspicuous footnote. It is removed — and reappears instead in the section on open points.
Point 2 is the one that carries the most weight in legal practice and is done the least often. An archived state is not proof that a statement was correct; it is proof of what stood there at the time of retrieval. That is exactly what is needed a year later.
Why this requires a mode of its own
This chain of checks takes time, which is why it is not a mode that can simply be placed in front of a chat reply. A deep research run takes roughly twelve to fifty-five minutes and runs in the background; at the end there is a written opinion as a PDF and a DOCX in the case file, and a notification arrives — in the portal and by email.
The process has seven stages, and two of them are unusual enough to be worth naming.
The first is clarifying the brief. The question as posed is first turned into a structured research brief: the issue, the area of law, the legal system, the period under consideration, express exclusions. This brief is submitted to the lawyer for confirmation and editing before anything is consumed. That is not merely a question of cost: most unusable research results arise not because the searching was poor, but because the wrong question was searched.
The fourth is a cross-check pass. A separate, deliberately adversarial pass reads the interim result against the grain and looks for contradictions, missing counter-positions and assertions without support. For every objection that would require a source, a targeted counter-search is run. Whatever is still open afterwards is not smoothed over but named.
In between and after that: the outline with five to sixteen sub-questions, the gathering per sub-question (first the platform's own body of statutes and case law, then the case file, only then the web), the section-by-section writing, the closing citation check and the output.
The order of gathering is deliberate
That the platform's own holdings come first is not a cost-saving measure but a quality decision: the official statutory wording is the better source than a page that quotes it. Only when the platform's own library and the case file yield nothing does the research go out onto the open web — to the place where the four checks above become necessary.
What this does not replace
The fast Research mode in the assistant remains exactly what it is: a web answer with links, in seconds rather than minutes. Anyone who only wants to know whether there is recent case law on a topic will continue to use it. Deep research is the expensive route for the questions where a paragraph is not enough — it is a paid add-on module with its own monthly quota and requires a paid plan.
And it does not replace review by a lawyer. An opinion is a basis for research; that is stated on its cover page, and it is not there as a platitude. What is checked is the wording of a source — not whether it fits the case at hand. That question stays where it belongs.
→ For law firms · Pleadings & evidence · Agents & watchers · Legal library
