A deep research agent will now read 160 web pages, think for fifteen minutes, and hand you a report with forty footnotes. A 2026 audit of these agents found the footnotes are the least trustworthy thing in the document. The links resolve. The pages are on topic. Whether the source actually supports the sentence it sits under is closer to a coin flip.
Deep research is no longer an occasional chat feature. It ships as a standard mode across ChatGPT, Gemini, Perplexity, Claude and Grok, and Google has turned it into a developer primitive: the Gemini Deep Research API lets an app kick off an agent that autonomously plans a task, runs up to 160 web searches, analyzes private files, and returns a fully cited report in under twenty minutes, all from a few lines of Python. That is the moment the reliability question stops being academic. A feature you invoke by hand a few times a week is one you eyeball. A feature wired into a pipeline that drafts client briefs or investment memos is one whose citations nobody reads before they ship.
The links work. The facts often don’t.
The sharpest measurement comes from a May 2026 preprint, Cited but Not Verified, which parsed the citations produced by 14 models across OpenAI, Anthropic, Google and several open-source labs. It scored every citation on three separate dimensions, and the three diverge in a way that should change how you read one of these reports.
The first dimension is whether the link works: across models, URL validity ran from 80.8% to 100%. The second is whether the linked page is topically relevant: 60.6% to 95.7%. The third is whether the source actually supports the factual claim it is attached to: 24.4% to 76.8%. The authors put the contrast plainly: even the strongest frontier models keep link validity above 94% and relevance above 80%, yet land at only 39% to 77% factual accuracy. (It is a preprint, not yet peer reviewed, so weigh it as one careful measurement rather than settled science.)
That gap is the whole problem. The parts of a citation you can check in a glance, the URL resolves, the page is about the right topic, are the parts the agents get right. The part that requires opening the page and reading it against the claim is the part they get wrong up to three-quarters of the time. A separate citation-accuracy roundup put the spread wider still, from roughly a third of citations hallucinated by the most reliable search model to the worst offenders near 90%, depending on the model and the domain.
This is the same failure shape this site documented in AI spreadsheet analysis: the wrong output is engineered to pass your review. A fabricated citation with a plausible URL and an on-topic landing page survives a skim precisely because it looks like every real citation around it. Checking that the footnotes exist proves nothing. Checking that they say what the report claims is the only test that matters, and it is the one nobody runs on a forty-source document.
Depth is the selling point and the failure mode
Here is the finding that should make anyone pause before turning the search budget up. The same preprint reports that fact-checking accuracy fell by roughly 42% on average as search depth rose from 2 tool calls to 150, while link validity and relevance held steady above 92% the whole way. The more the agent reads, the more confident and thorough the report looks, and the less its citations actually ground the claims.
Now reread Google’s pitch for its Deep Research API: up to 160 web searches per task. That is not a coincidence of numbers; it is the exact region where the audit says grounding degrades most. Depth is what these products advertise, because a report built on 160 sources reads as more authoritative than one built on 12. Depth is also what erodes the link between claim and evidence, because a model juggling 150 retrieved passages loses track of which passage justified which sentence. You are paying for the feature that breaks the thing you actually need.
The breadth gap between tiers is real, to be clear. The preprint found frontier models completed a cited report 83% to 100% of the time, while fewer than half of open-source models managed it at all, at 17% to 40% task success. If you self-host, this is a capability you mostly do not have yet. The point is not that deep research is broken. It is that its headline metric, sources consulted, is inversely correlated with the metric you care about, sources that hold up.
Two tools, two jobs
The deeper confusion is that “AI research” now covers two architectures that fail in opposite directions, and most buyers treat them as interchangeable.
Open-web deep research (ChatGPT, Gemini, Perplexity, Claude Research, Grok) is built for discovery. You point it at a question you cannot answer yourself and it goes and finds terrain you did not know existed. Independent comparisons sort them by exactly the traits that matter for discovery: Perplexity is fastest and leans hardest on transparent, numbered citations with confidence flags; ChatGPT writes the longest structured reports; Gemini reads the most sources and drops the result straight into Docs and Sheets; Claude pulls fewer pages but reasons more carefully over them; Grok is the only one pulling live posts for breaking news. Every one of them inherits the open web’s noise, and every one of them requires you to verify before you rely.
Closed-corpus grounding is the other tool, and it answers a different question. NotebookLM, renamed Gemini Notebook on July 16, 2026, along with Claude Projects and ChatGPT with your files uploaded, restricts its answers to the documents you provided. It cannot tell you anything outside that corpus, which is the entire point: every claim traces back to a source you chose and can open. It will not discover the paper you missed. It also will not invent one.
The decision rule is simple once the two are separated. For provenance-critical work, a board brief, a regulatory filing, a due-diligence memo, anything where a single fabricated figure is a liability, start from a closed corpus you assembled and verified. For a landscape scan where you want to know what is out there and you will fact-check the shortlist anyway, open-web deep research earns its keep. Using an open-web agent for the provenance job is how a made-up statistic ends up in a document with your name on it.
A verification pass that takes ten minutes
You do not need to re-research the report. You need to check the sentences the report is standing on.
Start with the load-bearing citations, the three or four that support the central claim, not the full reference list. Open each linked page and read the claim against the source, not the URL against the internet. The gap the preprint names is precisely “cited but not verified”: a genuine page attached to a sentence it does not actually support. Then prefer the agents that quote the supporting passage inline and let you jump to it in one click; a citation you cannot trace that fast is a citation to re-check by hand. Anything that leaves your desk with your name attached gets moved to a closed corpus or re-cited manually before it ships.
Twenty years in IT operations, and more recently fractional COO work assembling diligence and board material, taught me a rule that transfers cleanly here: a wrong number with a confident footnote is more dangerous than a missing one, because the missing number gets flagged in review and the footnoted one clears it. When I was reconciling vendor claims at a large telecom, the discipline that caught the errors was never “does a source exist.” It was “can I open the source and does it say this.” A deep research agent is a very fast analyst who has learned to produce the footnote without doing the second half. This is the same governance instinct behind content provenance and verifiable records: a claim you cannot trace is a claim you cannot defend.
Deep research agents are genuinely strong at the first part of the job, mapping a question, surfacing sources you would have missed, drafting a structure in minutes instead of hours. They are not yet a set of citations you can publish unread. Treat the report as a research lead rather than a finished document, verify the claims it rests on, and for the work that has to be defensible, build from sources you picked yourself. The version that looks finished is the one to check hardest, because looking finished is the part it has already mastered. For anyone still choosing tools on benchmark scores alone, the deeper lesson is the one the site keeps returning to: test on your own work, not the leaderboard.