Deep Research explained

Every major AI app now has a research feature. OpenAI calls it Deep Research. Google calls it Deep Research too. Anthropic calls it Research. The names line up because the idea is the same: you give it a question, it searches the web many times, reads what it finds, cross-checks, and hands you a report with citations. This lesson is about using that well, whichever app you have.

What it actually does

A research run has four jobs, and you can watch it do each one in the progress log:

• Search. It turns your brief into a set of sub-questions and runs a web search for each. Then it runs more searches based on what the first batch turned up. A single run can fire dozens of searches.
• Read. It opens the promising results and pulls out the parts that answer a sub-question. It skips duplicates and pages that turn out to be thin.
• Synthesize. It compares what the sources say, notices where they disagree, and writes a report organized around your question rather than around the pages it found.
• Cite. It attaches a source to the claims it pulled from pages, usually as numbered links, with a source list at the end.

The word synthesize just means combining several sources into one coherent answer. It is the step that separates a research run from a list of search results, and it is also the step where the most errors creep in, because the model is deciding what the sources "mean" together.

Where to find it, as of August 2026

Product names shift, so treat this as a snapshot and look for the equivalent button if yours has moved.

• ChatGPT. Deep Research lives in the tools menu of the message box. It is a paid feature; as of August 2026 you need Plus or above, and each plan comes with a monthly allowance of runs rather than unlimited use. Runs typically take somewhere between five and thirty minutes depending on the question.
• Gemini. Deep Research is in the Gemini app and works without a paid plan, though Google says it may be unavailable during high demand. Paid tiers get more usage; as of mid-2026 Google counts Deep Research against a shared usage allowance rather than a fixed number of reports. Gemini shows you its research plan before it starts and lets you edit it, which is the single best feature in any of these tools.
• Claude. The feature is called Research, and it is available on paid plans (Pro, Max, Team, Enterprise) with web search switched on. Research counts against your normal usage rather than a separate quota, but a run eats through that allowance faster than chatting. Runs usually finish in minutes, and harder questions can go much longer.

If you only have a free plan, you can still get part of the benefit: turn on web search in a normal chat and ask for sources. You lose the loop, but you keep the citations. Lesson 6 has the full cost comparison.

How to write a research brief

A research brief is the prompt you hand to a research feature. It should read like the note you would leave for a capable assistant before leaving for the day, because that is exactly what it is. Five parts:

• Who you are and why. "I own a window tint shop in St. George and I am deciding whether to add paint protection film." The why steers every judgment call in the run.
• The exact question. One question, sharply put. Two questions get two shallow answers. "What does it cost to get started, and what do shops my size typically earn from it in year one?"
• Include and exclude. Time range, geography, what counts. "Sources from 2024 or later. US only. Ignore franchise opportunities."
• What you already know. Saves the run from re-discovering it, and signals your level. "I already know the two main film brands. I do not need a brand comparison."
• Delivery. Length, structure, and what to do when it is unsure. "Under 800 words. A cost table first. When sources disagree, say so instead of picking one."

That last instruction, "say so instead of picking one," does more for accuracy than anything else in the brief. Left alone, the synthesis step tends to smooth over disagreement. You want the seams showing.

Three briefs that waste a run

The failures are predictable, which means they are avoidable:

• The one-liner. "Research gutter guards." Ten minutes later you have a general encyclopedia article. The run did its job. You did not give it one.
• The two-in-one. "Compare skid steer leasing options and also tell me how to market to HOAs." Two unrelated questions split the run's attention and its word count. Run them separately, or pick the one you need this week.
• The already-answered. A question a plain chat with web search would have settled in twenty seconds. Lesson 6 has a decision table for this. The short version: if you would be satisfied with one good source, you do not need twenty.

One more, specific to tools that show you a plan before running: read the plan. If it lists a sub-question you do not care about, delete it. If it is missing the one you do, add it. Thirty seconds there shapes every minute that follows.

Reading what comes back

A research report arrives looking finished: headings, a summary, numbered citations, maybe a table. Resist reading it top to bottom like an article. Read it like a draft from a new employee, which is the approach lesson 5 builds out in full. For now, three habits:

• Look at the source list first. How many sources? What kind? If it is mostly one company's marketing pages, the report is that company's opinion with extra steps.
• Tap three citations. Pick the three numbers you care about most and open the page behind each. You are checking that the page exists and says what the report says it says.
• Find the hedges. Phrases like "sources vary," "reportedly," or "as of" mark the places where the run was unsure. Those are where you spend your attention.

And keep the rule of stakes from AI Basics lesson 9. A research report about which patio heater to buy needs a glance. A research report you will quote to a customer, a lender, or a lawyer needs every load-bearing claim opened.

Try this yourself

Run one real research task this week using the full brief format. Here is one you can adapt. Change the business and the question, keep the structure.

I own a small landscaping company in Washington County, Utah, with two crews. I am deciding whether to buy or lease our next skid steer.

Question: for a machine in the 60 to 75 horsepower range used about 600 hours a year, what is the realistic five-year cost of buying new, buying two to three years used, and leasing?

Include: purchase price ranges, typical lease terms, maintenance and resale assumptions, and Utah sales tax treatment if it differs. Sources from 2024 or later, US only.
I already know the main brands. Skip brand comparisons.

Deliver: a cost table for the three options first, then a one-paragraph recommendation. Where sources disagree on a number, show both and say which you trust more and why. Cite every number.

When it finishes, before you read the recommendation, open the three citations behind the biggest numbers in the table. Then read the rest. You will have a feel for how much to trust that run, and you will have practiced the review habit that makes the whole feature safe to use.