Use AI to find, read and summarise sources, but treat every claim as unverified until you have opened the original and found the sentence that supports it. Ask for primary sources with direct links, keep an evidence log of claim, source, date and supporting passage, and separate facts from estimates before anything reaches a client.
AI makes desk research faster at the searching and reading. It does not make the checking optional, and the checking is what a client pays for.
What counts as a source you can check?
A checkable source is one a colleague could open and use to confirm the claim without asking you anything. A link alone is not enough. It passes five tests:
- It opens. The link works and leads to the document, not a search page or an error.
- It is the original. It is the report, dataset, filing or official page itself, not an article summarising it or another AI answer.
- It supports the exact claim. The number, wording and scope match. A source about UK grocery does not support a claim about European grocery.
- It is dated. You know when it was published and what period the figure covers.
- Its author is known. You know who published it and whether they have a stake in the conclusion, such as a vendor sizing its own market.
The Nielsen Norman Group's guide to researching with AI puts the core habit plainly: always ask AI systems to cite primary sources, then check them yourself. A library guide from Breda University of Applied Sciences warns that tools may give wrong or outdated references, and that an AI tool will offer an answer even when there is none.
How often do AI citations fail?
Often enough that checking is not a formality. Two independent studies tested how well AI tools attribute information to its source.
- Tow Center for Digital Journalism, March 2025. Researchers gave eight AI search tools excerpts from real news articles and asked for the headline, publisher, date and link, across 1,600 queries. Collectively, the tools gave incorrect answers to more than 60 percent of queries. Some cited a republished copy instead of the original, and some cited links that were broken or did not exist.
- EBU and BBC, October 2025. Journalists from 22 public service media organisations in 18 countries reviewed more than 3,000 answers from four AI assistants. Almost half had at least one significant issue, and a third showed serious sourcing problems.
Both studies tested news questions, not market research. They still show the pattern a researcher should expect: a cited answer can look complete and still point to the wrong place. Our deep research case study shows how differently tools handle the same research brief.
How do you run AI desk research step by step?
Plan the question, let AI do the first pass of finding and reading, then verify before you write. Seven steps keep the work auditable.
- Write the question and its scope. State the market, the geography, the period and the decision it informs. "Grocery delivery market in Germany, 2024 to 2026, to size the opportunity for a client launch" is a question AI can search well. "Tell me about grocery delivery" is not.
- Ask for a search plan before any answer. Ask the tool to list the source types it will look for, the search terms and the date range. Correct the plan before it runs. This is where you exclude vendor marketing and undated blog posts.
- Ask for primary sources with direct links. Ask for official statistics, regulators, company filings, trade associations and peer-reviewed research first. For each claim, ask for the link, the publisher, the date and the passage that supports it.
- Log every claim. Put each claim into an evidence log with its source and supporting passage (template below). Claims without a source go in a separate list.
- Open and check every source. Confirm the five tests above. Check the number in the original table, not in the AI summary. Note any difference in definitions: one report's "online grocery" may include click-and-collect, another's may not.
- Reconcile conflicts. When two sources disagree, record both and the likely reason, such as different definitions, years or methods. Do not average them.
- Write from the log, not from the chat. Draft the section from verified rows only. Label estimates as estimates, give each figure its source and date, and list what you could not verify.
Step 7 is the one most often skipped. Writing from the chat lets unverified wording slip back in.
Which kind of AI tool fits which desk research job?
Different tool types suit different stages, and most agencies use more than one. This is a map of tool types, not a ranking.
| Desk research job | Tool type | Examples |
|---|---|---|
| A first map of public sources on a topic | Search-first assistant with links | Perplexity, ChatGPT with search |
| A long, multi-step scan of a market or category | Deep research mode | The deep research mode of a general AI assistant |
| Answers grounded only in reports you already trust | Document-grounded notebook | A notebook or project mode limited to uploaded files |
| Academic and scientific evidence | Literature search tools | Elicit, Google Scholar |
| Company filings, broker research and expert calls | Licensed content platforms | AlphaSense |
| Desk research held next to the study's survey and interviews | Research platform with desk research as a source | Meaningful |
A document-grounded notebook only knows what you upload, which is a strength when the source set is fixed and a limit when it is not. A search-first tool covers the open web but not paid reports or licensed databases. Check what each tool can and cannot see before you trust its silence on a topic.
What should an evidence log contain?
An evidence log is a table with one row per claim, showing where it came from and whether someone checked it. It turns "the AI said so" into something a colleague or client can audit.
| Column | What to record |
|---|---|
| Claim | The statement as it will appear in the report |
| Source title | The document's own title |
| Publisher | Organisation or author, and whether they have a stake |
| Published | Publication date and the period the figure covers |
| Link | A direct link to the original |
| Supporting passage | The sentence, table or page that supports the claim |
| Scope match | Geography, definition and population match the claim: yes or no |
| Type | Fact, estimate or forecast |
| Checked by | Who opened the source, and when |
A worked example: checking a market size
Say, as a made-up example, an AI tool reports that a category "is worth €4 billion in Germany", with a link. Before the figure goes in a proposal, work through the log columns:
- Open the link. Does it lead to the original report, or to a press release or article quoting it? If the latter, find the report.
- Find the number. Is it €4 billion in the source, or did the AI round, convert currency or add two segments together?
- Check the scope. Is it Germany alone, or the German-speaking region? Retail value or manufacturer value? Which year?
- Check the publisher. Is it a statistics office, a trade body or a vendor selling into the category? A vendor figure can still be useful, but say whose it is.
- Look for a second source. If another credible publisher gives a different figure, record both and the reason they differ.
If any step fails, the figure is not ready for a client. The usual fix is to go back to the original report and use its own figure and definition.
Keep the log with the project, not in someone's chat history. When the client asks "where does this number come from?" six months later, the answer should take seconds.
When is desk research not enough?
Desk research tells you what has been published. It cannot tell you what your client's customers think now. Three signs that you need primary research as well:
- Published estimates disagree widely. Market sizes from different publishers often use different definitions. If the decision depends on the exact size, a published range may not be precise enough.
- The question is about this client. How shoppers see the client's brand, or why they stopped buying, is rarely in a public report.
- Behaviour is changing fast. A two-year-old survey may describe a market that has moved.
In those cases, desk research frames the fieldwork: it tells you which hypotheses to test, which segments matter and what the client already knows. The two sources are strongest side by side. Social posts can add a third view of what people say unprompted; our guide to social listening for market research covers that source.
How does Meaningful handle desk research?
Desk research is one of the seven data sources in a Meaningful project, alongside surveys, AI-moderated surveys, focus groups, uploads, social and AI perception. Meaningful describes it as structured research on markets and competitors, with sources you can open.
The difference is where the desk research sits. It lives in the same project as the study's survey responses, interview transcripts and earlier reports, so a finding can draw on all of them. Agents read every response and transcript in small, focused batches and reconcile the findings, and each finding links back to its evidence, with caveats. That makes it quicker to see which claims rest on published sources and which on the client's own respondents. Our article on context rot explains why reading in batches matters for long material.
It does not remove the checking. A researcher still opens each source behind a claim and decides what it means for the client. Meaningful does not replace specialist recruitment, sampling design or the agency's own methodological judgement. Customer data is processed and stored in the EU and is not used to train AI models, according to Meaningful's data policy.
Meaningful costs €500 a month excluding VAT, with unlimited seats, checked on 18 November 2026.
Frequently asked questions
Can AI do desk research for a market research agency?
Yes, for finding, reading and summarising sources quickly. It cannot be the final authority. A researcher still opens each source behind a claim, checks that it supports the claim and decides what the evidence means for the client.
Which AI tools give sources you can check?
Search-first assistants and deep research modes show links to web pages. Document-grounded notebooks cite passages in files you upload. Academic tools cite papers. Each type can still cite a page that does not support the claim, so open the source either way.
How do I check an AI citation?
Open the link, find the exact sentence or table that supports the claim, and check the date, the market, the definition and who published it. If you cannot find the supporting passage within a minute or two, treat the claim as unverified.
Do AI tools invent or misattribute sources?
They can. A Tow Center test of eight AI search tools found they gave incorrect answers to more than 60 percent of 1,600 queries about where news excerpts came from. An EBU and BBC study found that a third of AI assistant responses had serious sourcing problems.
Is AI desk research good enough for a client report?
Only after verification. Every number in a client report should carry its original source and date, estimates should be labelled as estimates, and anything you could not verify should be left out or flagged as unverified.
How much does Meaningful cost?
Meaningful costs €500 a month excluding VAT, with unlimited seats, checked on 18 November 2026. There is a free three-day trial.
What should you do next?
Take the desk research section of your next proposal and build its evidence log before you write a word. If you want that desk research to sit in the same project as the study's survey and interviews, try it in Meaningful's free three-day trial and check every claim against its source.