Deep Research for Operators: AI-Assisted Market Research

A citation is not a verification. Write that on a sticky note and put it on your monitor.

The reason it needs saying is that AI research tools now produce reports that look like they were written by a careful analyst: headings, tables, footnotes, links. Footnotes feel like proof. They are only a claim about where the proof is.

On February 2, 2025, OpenAI launched deep research. It is genuinely useful. It is also easy to misuse in a way that makes you feel more informed while you become less so.

What is deep research, exactly?

It is an agent that searches, reads and synthesizes on its own, then hands you a report.

What OpenAI says it does

OpenAI describes deep research as an agentic capability that conducts multi-step research on the internet, finding, analyzing and synthesizing hundreds of online sources into a comprehensive report at the level of a research analyst. It is powered by a version of the upcoming OpenAI o3 model, a step on from the reasoning models I covered in OpenAI o1, optimized for web browsing and data analysis, and it can take from 5 to 30 minutes. Every output is documented with citations and a summary of its thinking.¹

How it launched

At launch OpenAI started with a version for Pro users, with up to 100 queries per month, and said Plus, Team and Enterprise users would follow.¹

What does OpenAI say it gets wrong?

More than most people read.

OpenAI's own limitations section says deep research can sometimes hallucinate facts in responses or make incorrect inferences, though at a notably lower rate than existing ChatGPT models, according to its internal evaluations. It says the tool currently shows weakness in confidence calibration, often failing to convey uncertainty accurately, and that there may be minor formatting errors in reports and citations at launch.¹

I cannot verify those internal evaluations. What I can tell you is that the failure modes OpenAI lists are exactly the ones that hurt a business owner: a confident wrong number, a rumor presented as a fact, and a link that does not say what the sentence claims.

Sam Altman said something similar about the models of the day in a Lex Fridman interview clip posted on March 21, 2024:

He was talking about the GPT-4 era and hallucinations in general. The rule has outlived the model, and a research report that looks finished is the easiest thing to skip checking.

Plays here, from YouTube

Where does deep research actually save time?

On the first draft of anything that would take a person many hours of reading.

Good jobs for it

• Competitor scans. Who sells something similar, how they position it, how they price it.
• Vendor comparisons. Features, pricing pages, reviews, with a skeptical eye.
• Regulation overviews. A readable map of what a rule covers, which you then confirm with the primary text and counsel.
• Niche background. Getting smart on an industry before a first call.

Jobs I would not hand it

Anything where being wrong is expensive and you cannot check: legal conclusions, medical claims, financial projections you will show to a lender. Use it to find the questions and the sources. Do the conclusion yourself.

How do you brief it so the report is useful?

A vague prompt gets a generic report. A brief gets something you can act on.

How do you audit the report before you use it?

This is the step that decides whether the tool helps or hurts. It takes less time than the research would have.

Grade every source

• Tier A, primary. Government and agency pages, court documents, company filings, official product documentation, original research.
• Tier B, reputable secondary. Established publications and respected trade sources that cite their evidence.
• Tier C, interested parties. Vendor blogs, agency posts and marketing pages. Useful, but they sell something.
• Tier D, unknown. Forums, anonymous posts, scraped content. Leads only.

Anything you plan to act on should rest on Tier A, or on Tier B with a second source.

Open the link and find the sentence

For each number, open the citation and find the sentence that supports it. If it is not there, the number is out. Check the date. A fact from 2019 is not a fact about 2025. Check who published it, and whether they had a reason to say it.

What does a finished, trustworthy output look like?

Not the raw report. A one-page memo in your own words that answers the decision, lists the five or six facts that carry it, shows where each fact came from, and says what remains unknown.

Keep an evidence log

For each fact that matters, record the claim, the source link, the date, the source tier and whether you checked a second source. A simple table does it. Six months later, when someone asks where a number came from, you will have the answer in seconds. It also makes the next report faster, because you can reuse what you already verified.

Why does this matter more as agents get better?

Because errors compound, and I made the same point about browser agents. Anthropic's guidance on building agents notes that the autonomous nature of agents means higher costs and the potential for compounding errors, which is why it recommends extensive testing with appropriate guardrails.² A research agent that reads a hundred sources can inherit the mistakes of ten of them and pass them on with confidence.

Your job is to be the guardrail. The cheaper the first draft becomes, the more valuable the person who can tell which sentences are true.

What is a good workflow for a one-person business?

Brief it in ten minutes. Let it run while you do something else. Spend thirty minutes auditing the sources that matter. Write the decision yourself, in your own words.

That turns a week of research into an afternoon, and it keeps you in charge of the part that carries the risk. I have no employees, so my time is the scarcest thing I have, and this is the shape of help I want from a tool: fast on the first draft, honest about its limits, and checked by me before it touches a client. It is the same standard behind the AI-assisted systems I build for clients.

Keep building.Ekai