Research, search and the citation that does not say that
What connecting to search fixes, and what it does not
Tools that search the web before answering are a real improvement and they are the right default for anything about the current world — a rate, a rule, a company, a person, a date. A retrieval step has been added: pages are fetched, and the answer is written with those pages in front of the model. Fabricated sources largely disappear.
A different failure takes their place, and it is quieter.
The link is real and the claim is not in it. The page exists, it is on the right topic, it is from a credible body, and the sentence you have been given is a compression of several pages, or a plausible neighbour of what the page says, or accurate for last year. Because a footnote is present, nobody opens it. That is the whole trick, and it catches careful people.
It is also out of date in a specific way. Search results reflect what ranks, and what ranks is often a well-optimised summary of the rule rather than the rule. For a threshold, a fee or a deadline, the second-best source is somebody's blog post from 2024 explaining the current rate. It was current when written.
The verification loop
Three steps. Ninety seconds. Do it every time a claim will be relied on.
- Open the link. Not the summary of it.
- Find the sentence. Use your browser's find function with a distinctive phrase from the claim. If the page does not contain a sentence supporting it, the claim is unsupported, however reasonable it sounds.
- Check the date and the authority. When was the page updated, and is this body the one that sets the rule? A trade association explaining a regulator's rule is a secondary source. Go one step up and you are usually on the regulator's own page.
If the answer cites five sources and you cannot spare five minutes, then you cannot rely on the answer. That is not pedantry; it is the arithmetic of the verification-cost trap from module one.
Deep research modes
Most of the major tools now offer a longer-running mode that searches many pages and returns a structured report with dozens of citations, taking several minutes. These are genuinely useful for one specific job: mapping an unfamiliar area quickly. What the main positions are, who the main bodies are, what vocabulary the field uses, what you did not know to ask.
They are much weaker as a source of settled fact, for the reason above multiplied by scale. Forty citations is not forty verifications. Treat the output as a well-organised set of leads, and verify the four or five claims your work will actually rest on.
Two habits make them much better. Tell it explicitly what you already know and what you need, so it does not spend its effort on background. And ask it to flag where sources disagree, rather than producing a smooth consensus. Disagreement between sources is information, and a summary that hides it has thrown away the most useful thing it found.
What to search for, and what to ask a person
Some questions do not have a document behind them, and no amount of searching will produce one. Whether your regulator interprets a rule strictly in practice. Whether this supplier is reliable. What your committee will accept. Whether the previous head of department already tried this in 2019.
The tools are confident on all of these and have nothing to offer. Local, tacit, relational knowledge lives in colleagues, professional bodies and predecessors' files. It remains one of the strongest arguments for picking up a phone, and it has become more valuable, not less, now that generic information is free.
Where research meets the law
Two field-specific warnings, because both have already produced professional consequences.
Legal citations must be checked in a proper legal database, not by whether the case name looks right. Courts on several continents have now sanctioned lawyers for filed briefs containing fabricated or misdescribed authorities, and "the AI produced it" has been accepted nowhere as mitigation.
Clinical and safety guidance must come from the current guideline document, not from a summary of it. Guidance is versioned and dated for a reason, and a superseded threshold delivered fluently is more dangerous than no answer at all.
The one-line rule
A citation is a claim about a document. Until you have opened the document, you have two claims and no evidence.
The one thing to keep
A search-connected tool replaces invented sources with real links attached to claims the pages may not make, so the check moves from 'does this exist' to 'open it and find the sentence'.
Before you move on
A policy officer asks a search-connected assistant for the current threshold and gets an answer with three working links to official sites. What is the residual risk that the links do not remove?
Pick the one you would defend. Nobody sees your answer.