Deep research tools, compared honestly: the report is where your work starts

You typed one paragraph, went for coffee, and came back to twelve pages: headings, findings, a confident executive summary, and footnotes all the way down. It looks like the output of a diligent analyst who worked through the night.

The feeling that report produces is the thing to be careful with. Length and structure read as diligence, and footnotes read as proof. But every claim in those twelve pages was assembled by the same kind of model that writes everything else, from sources it selected, read, and summarized without you. The report is not wrong because of that. It is unverified because of that, and those are different states.

As of early 2026, the practical differences are speed, depth, and ecosystem. Perplexity's deep research is the fastest, returning reports in a few minutes with citations on every claim, and its underlying engine had the best citation record in Columbia's Tow Center audit, at 37% incorrect on news sourcing. ChatGPT's deep research runs much longer per task, produces the longest and most structured reports, and rations runs monthly depending on plan. Gemini's leans on Google's index and scholarly integration.

Pick by the job: quick landscape scans favor the fast one, single deep dives favor the long-running one, academic sweeps favor the scholarly one. And notice what did not vary: every one of them writes fluent, confident prose whose reliability you cannot judge from the prose itself.

Three failure modes recur. Source quality laundering: the agent reads what the open web offers, and a well-formatted content-farm page weighs more than it should. Synthesis drift: ten accurate sources summarized into one paragraph can still yield a claim none of them makes, the same mischaracterization pattern Stanford measured in grounded legal tools at 17% to 33%. And coverage illusion: a report that reads exhaustive covered what the agent happened to retrieve, not what exists.

None of these show up in the report's tone. The twelve pages read identically whether the pipeline worked or failed, which by now you will recognize as the recurring theme of this whole series.

The honest setup treats a deep research report as a map of candidate evidence. First, let the agent do the sweep; breadth is what it is genuinely good at. Second, extract the claims that will actually carry weight in your decision, usually three to five, and verify each at the passage level: open the cited source, compare the claim, note what it does not prove. Third, discard or flag everything consequential you could not verify.

For the claims that survive, you now have something better than a report: a small set of checked facts with open sources behind them. That is the material decisions deserve, and it took the 90-second ritual per claim, not another twenty minutes of agent time.

One boundary matters more than any tool choice: deep research agents are built for the open web, and feeding them your contracts, client files, or internal reports means researching your confidential material with a tool shaped for public sources and public retention policies.

Questions about your own corpus belong in a grounded library: private by default, answering only from your documents and curated collections, every claim linked to its exact passage. That is Tatsulok's shape, and it means the verification step that costs real effort on a web report costs one click on your own material. Use the web agents for the world. Use a grounded library for what is yours.

The drill this week, no signup needed: take one deep research report you have received, pick its three most consequential claims, and run the passage check on each. Count how many survive intact. That number, not the page count, is the report's real value.