What Is Deep Research in AI? How It Works, What It Costs and When to Use It in 2026 (ChatGPT, Gemini, Claude, Perplexity, Jobbit)

A florist in Leeds asked a chatbot whether there was a market for dried-flower subscriptions in the UK and got three confident paragraphs with no sources, two invented statistics and a competitor that closed in 2023. She asked the same question as a deep research task and, 14 minutes later, had a 9-page report with 38 cited pages, the prices of six live competitors, a Royal Mail parcel-size table and a list of the things it could not find. Deep research is the name the AI companies gave to that second mode, and it is the single most useful feature a small business can learn to use this year.

It arrived in a rush: Google added Deep Research to Gemini in December 2024, OpenAI launched its version in ChatGPT in February 2025, Perplexity followed within the fortnight, and by 2026 every general assistant has one, under names like Research, Researcher or Super Agent. BrightLocal's 2026 consumer survey found 45% of people now use ChatGPT or another generative AI tool to find local businesses, up from 6% a year earlier, which is one measure of how fast the habit is spreading. This guide explains what deep research actually is, how it works, what each tool allows for free, the jobs it does well and badly, and how to read a report without being misled.

What deep research means

Deep research is an AI agent working on a question for minutes rather than seconds. Instead of answering from what the model remembers, it plans the question into parts, runs dozens of web searches, opens and reads the pages, checks what it finds against other pages, and writes a structured report with citations you can click. A chatbot answer is a recollection; a deep research report is a piece of work.

The five steps are the same in every product, whatever it is called:

1. Plan. The agent breaks "is there a market for dried-flower subscriptions in the UK" into sub-questions: market size, competitors, prices, delivery costs, regulations, demand signals.
2. Search. It runs searches for each sub-question, often 20 to 80 in total, and chooses which results to open.
3. Read. It reads the pages, including PDFs and tables, and extracts facts with their source.
4. Check. It compares sources, notices disagreements and dates, and decides what to trust.
5. Write. It produces a report with headings, tables and numbered citations, and says what it could not establish.

The reports take 2 to 30 minutes depending on the tool and the question, and they are long: 2,000 to 10,000 words is normal. That is the point. You are paying time and credits for breadth and sources, not for speed.

Deep research vs a chatbot answer vs a Google search

• A chatbot answer is fast and fluent and comes from the model's training, which has a cutoff date and no obligation to be right about your niche. It is the right tool for drafting, explaining and rewriting, and the wrong tool for facts you will act on.
• A Google search gives you live pages but leaves the reading, comparing and summarising to you. Thirty tabs later you have the answer and an afternoon gone.
• Deep research does the thirty tabs for you and shows its working. The output is slower and longer, and every number has a page behind it that you can open.

The practical rule: if a wrong answer would cost you money, time or embarrassment, use deep research or a human; if it would cost you nothing, use the chatbot.

10 deep research tools and what each allows for free in 2026

Prices are as published on each company's pages on 3 October 2026 and change often; the free allowances in particular are adjusted without announcement, so check the plan page before relying on a number. The pattern is stable, though: every tool gives you a taste for free and sells depth by the month.

What a deep research report costs per question

Most tools meter deep research separately from chat, because each report costs them real computing: dozens of searches, hundreds of pages read, and a long piece of writing. OpenAI describes its allowance as limited on Free and Go, expanded on Plus and maximum on Pro, rather than publishing one number; Perplexity publishes five a day for free users and 500 for Pro; Gemini's free tier allows a handful of reports a month, with far more on the paid plans. On credit-based agents such as Manus, Genspark and Jobbit, a report costs a slice of the monthly credits, with a bigger slice for a longer question.

For a small business the useful arithmetic is per decision, not per month. A question that decides whether to spend £3,000 on stock deserves a 20-minute report and a second one from a different tool to compare. A question about the opening hours of a supplier deserves a search.

7 jobs deep research is good for, and 4 it is bad for

Good for:

• Market and competitor research. Who sells what, at what price, with what delivery terms, with the pages to prove it.
• Supplier and product comparisons. Thirty options, five criteria, one table.
• Regulation and compliance questions. What the rules actually say, with the source, before you ask a professional to confirm.
• Grant, tender and funding searches. What is open, who qualifies, what the deadline is.
• Technical explanations. How a thing works, with primary sources instead of forum folklore.
• Content research. The facts, dates and figures behind a blog post or a pitch, each with a citation to check.
• Monitoring. The same question every Monday, as a digest, so you notice a change the week it happens.

Bad for:

• Anything behind a login. Trade prices, your own bank data, a paywalled report: the agent cannot see them unless you give them to it.
• Very recent events. Pages written in the last few hours may not be indexed; the report will be a day behind.
• Judgement calls. It can lay out the options; it cannot know your customers, your stock or your appetite for risk.
• Precise legal, tax or medical advice. It finds the rules; a professional applies them to you.

How to read a deep research report without being fooled

A report with 40 citations looks authoritative, and most of it will be right. The errors cluster in predictable places, so check those first:

• Click five citations, including the two behind the numbers that matter most. A citation that does not say what the report says it says is the most common failure.
• Check the dates. A price from 2024, a regulation that changed in April, a competitor that no longer trades.
• Watch the country. UK questions attract US sources; "small business" statistics in particular are often American, in dollars, under different rules.
• Separate found from inferred. Good reports say "no public figure exists; estimate based on X". Bad ones present the estimate as a fact.
• Ask for the gaps. "What could you not find?" is the most useful follow-up question there is.
• Run it twice. The same question on two tools, then compare. Where they agree, you can act; where they differ, you have found the thing to check by hand.

The step-by-step method for running a deep research task well, from the first prompt to the finished report, is in How to Run Deep Research With AI and Turn It Into a Report.

How an AI agent takes research further

ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse will all produce a deep research report, and for the report alone they are good. Where Jobbit differs is what happens after the report:

• A browser of its own. The agent opens pages in its workspace, so it can read the pricing table a chatbot only guesses at, and you can watch and take over.
• The report as a file. A formatted document with a contents page, tables and the citation list, ready to send, not a chat window to scroll.
• A digest on a schedule. "Every Monday, check these six competitors' prices and tell me what changed" runs as a scheduled task on paid plans.
• The next job in the same chat. The research becomes the price list, the landing page, the supplier email or the pitch deck without starting again.
• A person when it needs one. Phone the suppliers, visit the shop, read the paywalled report: Jobbit can bring in a UK freelancer from the same place, with payment held in escrow until the work is done.

Frank, 68, a retired bus driver, wanted to run guided walks in the Peak District and had no idea whether anyone would pay. One deep research task found 14 operators, their prices, their group sizes and their reviews, plus the insurance and first-aid requirements. The second message turned the findings into a one-page site with a booking form on his own domain, on the £7 a month Go plan. The third set up a monthly digest of competitor prices. The research was the smallest thing the agent did. If you are new to agents altogether, AI for Beginners starts from zero, and Chatbot vs AI Agent explains the difference this post leans on.

Mistakes people make with deep research

• Asking a vague question. "Research the coffee market" produces a vague report. "Independent coffee shops in Sheffield: how many, average prices for a flat white, which have closed since 2024, and what the three best-reviewed do differently" produces a useful one.
• Not saying the country. Say UK, say pounds, say the postcode area if it matters.
• Reading only the summary. The caveats live in the body.
• Trusting a number because it has a citation. Click it.
• Pasting confidential files into a tool whose policy you have not read. Check what the plan does with uploads.
• Using a chatbot for a decision and deep research for trivia. It is the other way round.
• Stopping at the report. The value is in what you do next: the price change, the email, the page, the order.

A first week with deep research

• Day 1: one real question. Something you were going to look up anyway. Give it the country, the numbers you need and the format you want back.
• Day 2: check it. Click five citations, note the dates, write down what it got wrong.
• Day 3: the competitor table. Every rival you know of, prices and reviews, in one table. Keep it.
• Day 4: the rule you were unsure about. Allergen labelling, the trading allowance, the licence you might need. Then confirm with the official page it cites.
• Day 5: the next job. Turn the best report into the thing it was for: a page, a price list, a pitch. On Jobbit it is the next message in the same chat.

Frequently asked questions

What is deep research in AI?

It is a mode where an AI agent spends minutes researching a question: planning it into parts, running dozens of searches, reading the pages, checking sources against each other and writing a long report with clickable citations. It is slower and longer than a chatbot answer and far more reliable for facts you will act on.

Is deep research free?

Partly. In October 2026, Perplexity allows five deep research tasks a day free, Gemini a handful of reports a month, and ChatGPT a limited allowance on its Free plan, with more on paid plans from about £4.50 to £20 a month. Credit-based agents such as Jobbit include research in a daily or monthly credit allowance; Jobbit's Free plan gives 100 credits a day with no card.

How long does deep research take?

Two to thirty minutes, depending on the tool and the question. Perplexity is the fastest at a few minutes; ChatGPT and Gemini often take 10 to 20 minutes for a broad question. You can carry on with other work while it runs.

How accurate is deep research?

Much better than a chatbot answer because every claim is tied to a page, but not perfect: citations can be misread, dates can be stale and US sources creep into UK questions. Check the citations behind the numbers that matter and run important questions twice.

What is the difference between deep research and an AI agent?

Deep research is one task an agent can do: read and report. A general AI agent such as Jobbit also acts on the result, building the page, sending the email, scheduling the digest or bringing in a human for the parts no software can do.

Can deep research see paywalled or private information?

No. It reads the public web and anything you give it. Trade prices behind a login, your accounts and paid reports are invisible to it unless you upload them, and you should check the tool's policy on uploads before you do.

Which deep research tool should a small business start with?

The one you already pay for, if you pay for one; if not, a free allowance on Perplexity or Gemini for a taste, then an agent that can act on the findings. On Jobbit the Free plan gives 100 credits a day with no card, the report comes back as a file, and the same chat builds whatever the research says you should build next.