This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
A chatbot answer is something you read and move on from. A report is something you save, paste into a deck and forward to your boss. That difference is why OpenAI’s new deep research agent deserves more attention from communications teams than the average model launch.
What OpenAI launched
On Sunday, OpenAI introduced deep research in ChatGPT. You give it a prompt, and over 5 to 30 minutes it finds, analyzes and synthesizes “hundreds of online sources” into what OpenAI calls a report “at the level of a research analyst,” with citations and a summary of its reasoning. It is aimed at people doing work in finance, science, policy and engineering, and OpenAI’s own example prompts include competitive analysis.
For now it is limited to Pro users, with up to 100 queries a month, and not yet available in the United Kingdom, Switzerland or the European Economic Area. Google has offered a similar Deep Research feature in Gemini Advanced since December.
OpenAI is unusually direct about the limitations. It says the agent can “hallucinate facts,” that it “may struggle with distinguishing authoritative information from rumors,” and that it shows weakness in confidence calibration, “often failing to convey uncertainty accurately.”
Put those caveats next to the intended use cases and the reputation issue is obvious. Analysts, investors and consultants will commission reports about companies. Those reports will look like finished work. Some of them will contain rumors presented with confidence. The company described will almost never see them.
A checklist for the reports you will never see
Run the reports your stakeholders would run. Write the prompts an analyst, a journalist, a prospective customer or a job candidate might use about your company and your category. If you lack Pro access, Gemini’s version is a reasonable proxy for the genre. The goal is to see the shape of what gets produced, not to chase one output.
Read the sources before the prose. The citation list tells you which pages the agent treats as authoritative. If a forum thread, a scraped directory or an outdated article is doing heavy lifting, that is the real finding.
Look for confident sentences about uncertain things. Given OpenAI’s own calibration warning, pay attention to claims stated flatly that should be hedged: pending litigation described as settled, rumored deals described as done, former executives described as current.
Fix sources, not reports. There is no way to correct a report someone else generated. The leverage is upstream. Keep official pages current and dated, state basic facts plainly, and pursue corrections with third-party publishers through normal channels. I made a version of this point in 2023 about answers that still need corroboration; a longer answer needs more of it, not less.
Prepare for the forwarded paragraph. Sooner or later an executive, board member or customer will arrive with a report that says something wrong. Have a short, sourced fact sheet ready so the response takes minutes, not a week of drafting.
Why the artifact matters
Last year, the argument was that cited answers could become the new homepage. Deep research takes that a step further. The answer is no longer a screen someone glances at. It is a document with a life of its own, shared inside organizations where the company being described has no voice.
The practical response is not alarm. It is the old discipline of making the true version easy to find, applied to a reader who now writes the memo itself.