Deep research: several searches, one cited report

Ordinary web search looks something up and answers. Deep research treats your question as a small project: it breaks it into several angles, searches each, and then writes the whole thing up as one report with citations. It is built for questions such as comparing three tools for a team, summarising the current state of a topic, or working out what the rules are in a particular place.

The live research plan is the useful part for judging quality. You can see which angles it chose and notice when one is missing, then ask a follow-up that covers it.

• Comparisons: which of these options fits a five-person team that works in Google Workspace, and what are the trade-offs?
• Overviews: what changed in a regulation or standard over the last year, and who is affected?
• Background: a first pass on an unfamiliar market or technical topic before you talk to an expert.
• Planning: the main options, costs and risks for a decision you have to make.

It is a poor fit for a single fact, where a normal search is cheaper on your allowance, and for questions about private or paywalled material it cannot reach. For legal, medical or financial decisions, use it to prepare, not to decide. Our guide to AI copilots by profession shows how people in different jobs use research tools.

A long, confident report is easy to believe. Three habits keep you honest:

1. Open the two or three sources behind the claims your decision depends on and find the sentence that supports them. Reading about a source in other places, the habit called lateral reading in Stanford's Civic Online Reasoning materials, is quicker than judging a site by how it looks.
2. Look for what is missing. Ask: what is the strongest argument against this conclusion?
3. Check dates. A source from several years ago may describe a product or a law that has since changed.

Google's own advice on reliable, people-first content is a useful checklist for judging a page. Sources are only as good as the web, and the web has errors. If the topic is contested, ask the model to set out the positions side by side rather than pick one.

Each report is several searches plus a long answer, so it is limited per day:

• Free: 1 deep research report a day. After that, the chat says the answer skips the multi-source report.
• Pro: 30 reports a day.

Research reports come from their own daily allowance and not from your monthly credits. Requests on your own API key are not metered by Copilotly, although your provider will charge you for the usage, and a report can use a lot of it. See pricing for Pro: $6.99/week, $19.99/month or $149.99/year, with a 7-day free trial for first-time subscribers.

Deep research works best as the first half of a two-step job. Use it to map a topic, then narrow down with ordinary chat.

1. Write the brief. One paragraph: the decision, the audience, the constraints, the time frame. Add what you already know so it does not spend its searches rediscovering it.
2. Read the plan as it runs. The live plan lists the angles it is searching. If an obvious angle is missing, stop and restart with it added to the brief, rather than finding out after you have read the report.
3. Interrogate the report. Ask for the weakest part of the argument, for the claims that rest on a single source, and for what it could not find.
4. Close the gaps with search. Use web search for specific facts the report left open. They come cheaper than a second report.
5. Keep what you need. Export the chat as Markdown, or as PDF on Pro, or share it by link with a colleague.

With one free report a day on Free, it pays to spend it on a question that deserves it. Draft the brief in a note first, and if the topic is something you have a document about, attach the document to a normal chat instead, since file chats are built for questions about your own documents.

The model that writes the report matters as well as the searching. If you have a favourite, switch to it before you start. Pages such as Claude and Gemini describe how each tends to behave on long, structured writing.

Both read the live web and cite what they find. The difference is how much work happens before you see an answer, and how much of your daily allowance it uses.

When one search is enough

If the question has a single factual answer, such as a price, a date, an opening time or the latest version of a tool, use web search. It answers in seconds, cites its sources, and the free plan gives you 5 a day.

When a report is worth the allowance

Use deep research when the answer depends on several sub-questions that you would otherwise search one after another. Choosing between three project-management tools for a small team, for example, needs pricing, integrations, limits and reviews, and each of those is its own search.

Deep research versus asking a model directly

A model answering from its training has no sources and may be out of date. A research run reads current pages and links them. If you want to ask about a document you already own, a file chat is the better route, because the answer comes from your document rather than the open web.

You want	Use	Why
A single fact or a recent news item	Web search	Fast, and it costs one of your searches
A comparison with trade-offs	Deep research	Several angles, one structured report
A first pass on an unfamiliar topic	Deep research	The live plan shows you the shape of the topic
Answers from your own PDF or spreadsheet	File chat	Grounded in your document, not the web
A reworded paragraph	Neither	Use the paraphrasing tool

The report is only as focused as the question. A strong brief takes a minute to write and saves a wasted run, which matters when the free plan has one a day.

State the decision, not only the topic

"Project management tools" produces a survey. "Which project-management tool suits a five-person agency that bills by the hour and already uses Google Workspace?" produces a recommendation with reasons.

Set the boundaries

Give a region, a time frame and a budget where they matter. Rules, prices and availability differ by country, and a report that mixes them is hard to use.

Say what you already know

If you have already ruled out two options, say so and say why. It stops the run spending its searches on them.

Ask for the shape you want

A table of options, a short list of risks, a recommendation with the strongest counter-argument: say which you want. You can reshape the result afterwards in the same chat, and the rewrite tools help when you paste part of it into an email or document.

While it works, the chat shows Planning searches and then the plan itself. This is the best early signal of quality, because you can judge the approach before you read a word of the report.

What a good plan looks like

It covers the angles you would have covered yourself, and at least one you would not have thought of. For a comparison, expect separate lines for pricing, features, limits and independent reviews.

What to do when an angle is missing

Stop, add the missing angle to your brief and start again. This is cheaper than reading a report that ignores the part you cared about. If you are on Free and have used your report for the day, ask the missing question with web search instead.

What it cannot do

It reads what is publicly reachable. Paywalled articles, private documents, logged-in dashboards and anything behind a form are out of reach. If those matter, say so in the follow-up and give it the text yourself.

A report is a starting point for work, not the end of it. The steps below turn it into something you can use.

Pull out the decision and the evidence

Ask the chat for a one-paragraph recommendation, then a list of the facts it depends on, each with its source. That list is what you verify. Anything that is not on it was not important to the conclusion.

Turn it into a document or slides

Copilotly can generate documents, spreadsheets and slides as downloadable files, so a comparison table from your report can become a spreadsheet, and a summary can become a short deck. Check each figure before you share it, because a generated file is only as accurate as the report behind it.

Hand it to someone else

Share the chat by link, or export it as Markdown, or as PDF on Pro. Include the sources so your colleague can check them. If the report is the basis for a decision at work, write the decision and your reasoning in your own words next to it.

Keep the question for next time

Save the brief that worked. If you want similar reports regularly, a good brief is reusable: change the region, product or time frame and run it again. With memory switched on, you can also save standing context such as your industry in memory and custom instructions, so you write less each time.

Knowing the failure modes is how you use it well.

It reads what ranks, not what is right

Search results favour pages that are easy to find, which are not always the most accurate. For specialist topics, ask which sources are primary, such as a standard, a regulator, a vendor's own documentation or a paper, and which are commentary.

It can blend sources that disagree

When two sources conflict, a report may smooth the difference into one confident sentence. Ask directly: where do the sources disagree, and what would settle it?

It cannot weigh what you value

A recommendation depends on your priorities. If you care about privacy, cost, or speed of setup more than the average buyer, say so in the brief. Otherwise the report may weigh them the way a typical reader would.

It is no substitute for professional advice

For legal, medical, tax or investment decisions, use the report to prepare questions for a qualified person. Our comparison of AI and financial advisers sets out where that line sits.