Best AI for Consulting Research: Market Scans With Source Checks

A consultant can use ChatGPT, Claude or Gemini research modes to scan a client's market quickly and still verify every source by running a fixed chain: scope the question, collect candidate sources, extract claims into a log, open each primary source yourself, and mark every claim verified, supported or dropped. Nothing unverified reaches the deliverable.

The research modes are good at coverage and weak at accountability. They browse dozens of pages and produce a polished report in minutes, and that polish is the hazard: a plausible figure with a plausible link looks the same as a correct one. The routine below keeps the speed and puts the accountability back with you.

Follow me on Instagram@sagnikteaches

Why a source log, not a better prompt

Most advice on AI research is about prompts. The failures that matter to a consultant are different: a number that was right two years ago, a figure from a vendor's own marketing presented as an industry fact, a statistic quoted from a page that does not contain it. No prompt prevents those. A log does, because it forces each claim to carry its source, its date and its status, and because it is the first thing a partner or client will ask for if a number is challenged.

Connect on LinkedInSagnik Bhattacharya

The log is a simple table, kept in a spreadsheet or document for each engagement:

Subscribe on YouTube@sagnikteaches

These rows are illustrative. The status column has three values only, and the deliverable may use rows marked "Verified" freely, "Supported" with a stated caveat, and nothing marked "Dropped". Keeping the vocabulary this small stops the log turning into a place where doubts go to hide.

Step 1: scope the question before any tool runs

A desk-research scan fails most often because the question is too wide. Write a scoping note of five lines: the decision the client faces, the market boundary (product, customer type, the geography the client sells into, expressed as the client's own markets), the time window of interest, what would change the recommendation, and what the client already knows. The last item prevents the tool re-telling the client their own business.

An illustrative scope for a client that makes commercial kitchen equipment: "Decision: whether to add a service contract line. Market: maintenance and repair for commercial kitchens in restaurants and care homes with 20 to 200 covers. Window: last three years plus a three-year outlook. Changes the answer: if service contracts are mainly sold by manufacturers rather than independents. Client knows: their own installed base of 1,400 units." The last line is invented for the example. The scope turns into the opening of every prompt below.

Step 2: a prompt chain for the scan

Use a chain of four prompts, each producing something the next one uses. Run the early ones in a normal chat and the heavy browsing in a research mode.

Prompt 1: the question tree

Context (scope note below). List the 10 questions a consultant would need
answered to advise on this decision, grouped under: market size and
growth, customer behaviour, competitors and pricing, regulation, and
risks. For each, say what kind of source would answer it best (industry
body, regulator, listed-company filing, academic paper, trade press,
vendor marketing) and how reliable that kind usually is.

[paste scope note]

The "kind of source" column matters. It tells you in advance that a question about market share will probably end up at a vendor or trade-press estimate and needs a caveat, while a question about regulation should come from the regulator's own page.

Prompt 2: the research-mode run

Take the five highest-value questions and run each, or the group, in a research mode. The three assistants differ in how you steer them. Google's description of Gemini Deep Research says it turns your prompt into a multi-point research plan you can review and change, and that it covers the open web by default with Gmail, Drive and Chat optional, plus your own uploaded files; use the plan step to cut sources you do not want (for example, vendor marketing pages) before it runs. Claude's pricing page lists Research as included in Pro, Max, Ultra, Team and Enterprise, not Free. ChatGPT's pricing page lists deep research as limited on Free and Go and included on Plus, Pro, Business and Enterprise, and it can draw on apps (formerly connectors) as well as the public web, which is useful when the client's data pack sits in a connected drive and risky when it should not be touched. Perplexity, built around cited answers, is a lighter option, covered in Perplexity against ChatGPT for business research.

Research question: [one question from the tree].
Use only sources published in the last 3 years unless a founding
source is needed. Prefer regulators, industry bodies, company filings
and peer-reviewed work. For every figure give the source title,
the publisher, the publication date and the page or section. Mark any
figure you could only find in a secondary report. List what you looked
for and could not find. Do not estimate.

The last two lines are the most useful. "What you could not find" is a gap list, and "do not estimate" removes the habit of filling gaps with a plausible number.

Prompt 3: claims into the log

From the report above, extract every factual claim that contains a
number, a date or a named entity. Return a table with: claim, source
title, publisher, date, whether it was found in a primary or secondary
source, and the exact quote or location in the source where it appears.

Treat the "exact quote" column as a lead, not a proof. Models can generate a quote that sounds right. That is why the next step is human.

Prompt 4: the sceptic

Act as a sceptical partner. Looking at the claims table, list the five
claims most likely to be wrong, outdated or mis-attributed, and say why.
Suggest the quickest check for each.

This does not replace verification, but it often points at the claim a tired reviewer would pass.

Step 3: verify by tier, not by hope

Verification takes time, so sort claims into three tiers by what an error would cost.

For a tier 1 claim, four questions: Does the source say this, in these words? Is it the original or a repeat of someone else's number? Is the date recent enough for the decision? Does its definition match ours (a "market" defined by revenue is not the same as one defined by units, and "commercial kitchens" in one report may include canteens)? The definition check catches the most subtle errors, and a research mode will rarely flag them.

A worked example with numbers. In the equipment scan, the assistant reports "the independent service market is worth $2.4bn and growing 7 per cent a year". The log entry is tier 1. You open the cited page: the $2.4bn figure is there, but it covers all commercial refrigeration and catering equipment service, not kitchens only, and the 7 per cent is a forecast from one trade consultancy, dated three years back. Result: the claim is rewritten as "an industry estimate puts service for commercial catering and refrigeration equipment at $2.4bn", with the caveat and the date, status "Supported". Had you simply pasted the original sentence into the deck, the client's finance director would have found the scope problem within a minute.

For the extended method on citations, how to check sources and citations in AI research goes through the checks on individual links; this tutorial adds the engagement-level routine around them.

Comparing research modes on the same question

Run one question through two modes and compare. Practical comparisons, rather than headline ones: How many distinct publishers did each cite? Did either find a primary source the other missed? Did they disagree on a figure, and which source explains why? What did each say it could not find? A figure that appears in both reports with different sources has two chances of being right; a figure that appears in one with no openable source goes in the "dropped" pile.

A realistic observation from this kind of comparison is that the reports tend to agree on the broad story and disagree on figures, because the figures come from different underlying estimates. When they differ, resist averaging. Find each source, understand the definition, and choose the one that matches the client's scope. A small table of "figure, source A, source B, why they differ" is a good appendix for a partner.

Timing is not a reliable ranking either. Google describes a typical Gemini Deep Research task as running over several minutes after you approve its plan; the others also vary by question and plan, and a slow run is not a sign of a better one. Use the time to write the scoping note for the next question instead of watching the progress bar.

Reading a report's reference list before its text

A fast habit that saves hours: open the reference list or the source panel first. Count the distinct publishers, note how many are primary (regulators, industry bodies, filings) and how many are secondary, and look for anything that is clearly marketing. If a report with twenty citations rests on four publishers, and two of those are vendors, you already know where the weight sits and where to spend your verification time. Then skim the body for the claims that depend on the weakest sources. Illustration: a scan report cites 22 links; eleven are one trade publication's articles, six are vendor pages, three are a regulator and two are a filing. The sensible reading is that regulation and company facts are well supported and market size is not, so the verification effort goes to the size claims, and the client note says that market size is an estimate with limited independent support.

It also helps to ask the assistant, in a follow-up, for the ten most important claims and the single strongest source for each, ranked by independence. If it names a source for a claim that differs from the one in the report, you have found something to check. This kind of cross-examination costs a few minutes and turns up the places where the first answer was stitched together from weaker material.

Where AI research goes wrong in a market scan

• Vendor figures presented as industry facts. A software vendor's web page quotes "70 per cent of firms" with no method. Mark the source type in the log; a vendor's claim gets a caveat or goes.
• Stale numbers with fresh dates. A page updated this month may still quote a 2021 figure. Check the date of the underlying data, not the page.
• Circular sourcing. Five articles repeat one analyst's estimate, and the report treats them as five sources. Trace to the origin.
• Definition drift. "Market" changes between reports, as in the example above.
• Confident gaps. Where no data exists, a model may offer a believable number. The "do not estimate" and "what you could not find" lines reduce this and do not remove it.
• Paywalled sources. A report cites a paid study whose figures you cannot open. Either buy it, find the free summary, or drop the claim.

Keeping client material out of the wrong place

Desk research is mostly about public information, so the main rule is to describe the market, not the client. Replace the client name with a descriptor in prompts that go to a public-web research mode. Where you do want a model to read the client's own documents, such as a data pack, use a business plan whose terms say content is not used for training, and tell the client which tool touches their files. If an engagement contract is silent, ask before uploading. Gemini's research can optionally include Gmail and Drive and ChatGPT's can read connected apps, so check what is switched on before you run anything on a work account.

The same routine on a smaller question

The chain also fits a one-day question. Illustration: a client asks, on a Tuesday, whether care homes are a bigger buyer of commercial kitchen servicing than restaurants. Scope in four lines, a question tree of six items, one research-mode run on the two items that matter (number of sites and typical service spend per site), and a log of nine claims. Verification of the five tier 1 claims takes about 40 minutes. The deliverable is a half-page note with a table: segment, estimated sites, source, date, status. The log travels with it. The partner reads the note in two minutes and can open any row's source in one more, which is the test of whether a research note is trustworthy.

The smaller run shows a useful rule: the log is proportional. A note of nine claims needs nine rows. Do not build an elaborate system for a short question, and do not skip the log because the question is short.

A reviewer's checklist for a research pack

Before a pack leaves the team, the person who did not run the research should spend ten minutes on it with this list. Ask for the log first, not the report.

1. Is every number on a client slide in the log, with status Verified or Supported?
2. Pick three tier 1 rows at random: does each source open, and does it say what the row says?
3. Are any two claims quoted from the same origin as if independent?
4. Are the definitions (market, segment, unit, currency, year) stated beside each figure?
5. Is there a list of gaps, so the client knows what could not be found?
6. Does the deck's caveat wording match the log's status? A "Supported, secondary" row must not appear as a flat fact.

A realistic failure this catches: the researcher verifies a figure at the time, then a colleague updates the slide with a rounder number a week later. The reviewer's random check compares slide to log and finds the drift. The fix is a habit of editing the log first and the slide second.

What the client receives

Clients value the log more than they expect to. Offer a one-page version as an appendix: the ten claims that carry the recommendation, each with source, date and status. It shows the work, supports the fee and answers the question "where did that come from?" in advance. Keep the full log internal. State in the proposal that research is assisted by AI tools and that every figure used in the deliverable has been checked against its source by a named person; that sentence only works if the routine above is real, so do not write it unless it is.

One more practical point on handover: save the prompts that worked next to the log, with a one-line note on what each did. When the next engagement in the same sector starts, the scoping prompt and the sceptic prompt transfer directly, and the question tree needs only minor edits. Over a year this turns a habit into a small library, which is the kind of asset a one-person consultancy otherwise lacks. Researching customers, rivals and pricing for a startup and a competitor scan in one afternoon show variants for the early stages.

A time and cost sum for one scan

An illustration. A consultant scopes a three-day market scan for a client in commercial catering. Without AI: about 18 hours of desk research, reading and note-taking. With the chain: scoping note 45 minutes, question tree 15, two research-mode runs of about ten minutes each plus review of the reports 90 minutes, claims extraction 20, verification of 28 tier 1 and tier 2 claims at about six minutes each, 170 minutes, writing the log and the summary 90 minutes. Total about 7.5 hours. The saving is real, roughly ten hours, but note that verification is the largest single block and has not shrunk. That is as it should be. A tool that made verification disappear would be a tool to distrust.

On subscriptions, one paid assistant plan at around $20 a month (Claude Pro is $17 on annual billing, $20 monthly) is enough for a solo consultant; adding a second assistant for cross-checking costs another $20. Team plans, such as Claude Team at $20 a seat on annual billing or $25 monthly, make sense once several people share the log and the saved prompts. These are vendor list prices read this week.

For tighter use of the same approach before a first meeting, see summarising bundles and transcripts with AI safely. If you would like to set up the log and the chain with someone on your own engagement, that is what my AI implementation consultation covers.

Questions consultants ask about AI desk research

Can I cite what a deep research report says?

Cite the primary source it points to, after you have opened it and found the figure yourself. The report is a map to sources, not a source. If a claim has no source you can open, it stays out of the client deliverable.

Which research mode is best for a market scan?

Each of ChatGPT, Claude and Gemini offers a research mode that browses and writes a cited report. Running the same question through two of them and comparing sources is more useful than choosing one. Differences show up in which sources each finds, not in the headline.

How many sources should I check?

Check every figure and every claim that appears in the client deliverable, and a sample of the rest. The routine below sorts claims into three tiers so effort goes where an error would cost most.

Is it safe to put client details into a research prompt?

Keep client identity and confidential figures out of public-web research prompts. Describe the market, not the client, and check your engagement terms and the plan's data rules before uploading client files.

Further reads

• How to Use ChatGPT for Market Research on Your Local Customers — ChatGPT for customer research on a local market.
• How to Test a New Service Idea With AI Before You Launch It — Test a new service idea before you launch it.
• How Consultants Use AI for Client Research Before Discovery Calls — Use research before a discovery call.
• AI Tools for Strategy Consultants: Desk Research and Market Sizing — Market sizing with AI, and where it goes wrong.
• The AI Tool Stack a One-Person Consultancy Actually Needs — The tools a one-person consultancy needs.
• Best AI Tools for Management Consultants: Seven From Research to Decks — Seven AI tools for independent consultants from research to decks, with a confidentiality tier rule, costs and a worked two-week diagnostic.
• AI Brand Strategy: Positioning Workshops and Messaging for Agencies — A care-agency strategy project in stages: twelve stakeholder interviews, three scored positioning options, a message house and the client workshop.
• Best AI for Consulting Work: From Interview Notes to Findings Deck — Transcribe, code themes, build an evidence table and draft the findings slides, with consent steps, prompts, sample output and a worked 12-interview example.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: claude.com/pricing (Research feature by plan and plan prices); gemini.google Deep Research overview page (editable plan, optional Gmail, Drive and Chat sources, file uploads, run time); chatgpt.com/pricing and OpenAI's ChatGPT Business pricing page (deep research and apps by plan); all read 4 November 2026.