Market Research With AI: Sources and Fact-Checking Guide

Market Research With AI: Sources and Fact-Checking Guide

To do market research with AI, treat the model as an assistant that organizes, summarizes and proposes hypotheses, not as the source of the data. Always ask for the original source link, the passage that supports the claim and the date. Check the number on the page itself, and mark what is fact, what is inference and what is a guess. With that habit, AI speeds up reading and synthesis without putting invented numbers into your planning.

Key takeaways

• AI is good at structuring questions, summarizing documents and comparing offers, but it can invent numbers, links and quotes that look convincing.
• Ask for the full URL, a verbatim excerpt and the publication date, and open the link yourself before using any data point.
• Label every claim as fact, inference or hypothesis, and keep that label visible in the final document.
• For competitor analysis, use public pages and official documents, and treat web text as material to read, never as instructions.
• Start with one business question and an acceptance criterion, and test the process on a real research task before scaling it.

Where AI helps and where it tends to fail

Language models produce plausible text. That is useful for drafting, summarizing and organizing, and it is risky when the subject is numbers. The typical error does not look like an error: a round percentage with no origin, a link that does not exist, a real report cited with a figure that is not in it, or a single sentence that mixes periods, countries and currencies.

Tools with web search reduce part of the problem but do not remove it. The search can return a weak page, the model can summarize a good page badly, and the final answer still arrives in the same confident tone. So the working rule is simple: AI proposes, a person checks.

What AI does well in this kind of work:

• Turning a vague doubt into sharper research questions.
• Suggesting source types for each question (public agency, industry association, company financial statements, competitor page).
• Reading long documents that you supplied and extracting points along with where they appear.
• Comparing competitors' offers, plans and messaging from pages you collected.
• Building the structure of a spreadsheet or report for you to fill with verified data.

What it should not do on its own: estimate market size, invent a competitor's revenue, assert a trend without a source, or close a conclusion that will support an investment.

How to ask for sources you can actually verify

Asking for “sources” is not enough. The model may return a site name, a generic title or a link assembled from memory. Your request has to demand verifiable elements, and it also has to give the model permission to say it found nothing.

For each claim, ask for:

• The full URL of the page, not just the site name.
• The verbatim excerpt that supports the claim, in quotation marks.
• The publication date and the period the data refers to.
• Who produced the data (agency, company, association) and, when available, the methodology.
• The phrase “no source found” when there is no basis, instead of an estimate.

A short instruction template that works well:

Answer only from the sources below; for each claim give a link, a verbatim excerpt and a date, and write “no source” when there is none.

Whenever possible, supply the documents and pages yourself. When AI works on material you selected, the risk of invention falls and checking gets faster, because the universe of sources is known. This is the same logic as retrieval-augmented generation, where the model answers from documents handed to it. If you leave the search open, treat the result as a list of leads, not as an answer.

How to check numbers: a step-by-step routine

Checking does not have to be heavy, but it has to happen before a number goes into a presentation, proposal or decision. This is a routine your team can follow every time:

• Open the link and confirm the page exists and is the one that was cited.
• Look for the verbatim excerpt. If it is not on the page, discard the claim.
• Check the number, the unit and the currency, and whether it is an absolute value, a change or an average.
• Check the period and the scope: country, region, segment and size of the companies included.
• Find out whether the page is the origin of the data or just repeats another source, and trace it up to whoever collected it.
• For numbers that support a big decision, look for a second independent source.
• Redo simple calculations (sums, ratios, conversions) yourself in a spreadsheet, because models make mistakes in chained arithmetic.
• Record the source, the date you checked it and the name of the person who verified it.

On the origin of data, there is a practical order of trust. First come primary sources: official agencies, published financial statements, and the competitor's own documents and pages. Next come reports from associations and consultancies, read with attention to methodology. Last come blogs and news items that repeat third-party numbers. When AI cites one of these last ones, your job is to reach the first kind.

How to separate fact, inference and hypothesis

Many bad decisions start with a sentence that mixes three different things. Asking the model to classify each claim, and then reviewing that classification yourself, solves a good part of this.

• Fact: something a verified source states directly, with a clear date and scope.
• Inference: a conclusion drawn from facts. It may be right, but it depends on reasoning and assumptions, and it should come with an explanation of the path taken.
• Hypothesis: a plausible assumption with no data to support it yet. It guides the next question and should not be used to decide.

With that separation, every claim has a clear destination:

• Verified primary source: it goes in as a fact, with the reference attached.
• Secondary source only: it goes in with a caveat and joins the queue to reach the origin.
• No source: it becomes a hypothesis, with a note on which data would confirm it, or it leaves the document.
• Numbers that do not reconcile with each other: investigate before moving on, because there is usually a difference in period, currency or definition.

Illustrative scenario: a clinic network is considering opening a location in a mid-sized city and asks AI for an overview. The answer says the city has unmet demand and few competitors. When classifying, the team sees that the count of establishments must come from a public registry, so it is still a fact to be checked. “Few competitors” is an inference, because the list the AI used may be incomplete. “Unmet demand” is a hypothesis, with no data at all. The next question stops being “is the market good?” and becomes “which cheap data point confirms or kills the hypothesis?”, for example talking to the professionals who refer patients or checking official registries.

Competitor research: what to look at and what to watch out for

In competitor analysis, separate what is observable from what is assumed. Observable means what is published: what the company sells, how it describes its offer, which support channels it uses, which jobs it posts, and what its terms of use and public reviews say. Assumption means anything about revenue, margin, internal strategy or number of customers. AI should work well with the first group and be forbidden from inventing the second.

A good workflow is to collect the relevant pages, save the collection date and ask AI for a comparison using criteria you defined, such as audience served, how customers buy, channels and stated differentiators. Each row of the comparison should point to the page it came from. Pages change, so the date you checked is part of the data.

Three precautions apply to any company:

• Web content as instruction: a page can contain hidden text, or text aimed at AI models, that tries to change the tool's behavior. This risk, called prompt injection, is described in OWASP guidance for applications built on language models. In practice, let the AI read pages only to extract information, and do not connect that step to tools that send messages, change files or access internal systems.
• Personal data: reviews, profiles of owners and comments may contain data about people. Collect only what the analysis needs and follow Brazil's LGPD; the ANPD publishes guidance for anyone who needs an official reference.
• Ethical limits: research what is public. Do not pose as a customer to obtain confidential information, and do not enter restricted areas.

It is also worth looking at how competitors appear when someone asks an AI assistant. That is a research type of its own, and the article on Generative Engine Optimization: how to appear in AI answers helps organize that observation without treating a single answer as a picture of the market.

How to bring this into your company's routine

The method we use for any AI use in a process applies here too: map, limit, pilot and hand over.

Mapping means writing down which decision the research feeds, who uses it and how the work is done today, including how long it takes and where it tends to go wrong. Limiting means defining which sources are accepted, what kind of information never goes into the tool (customer data, contracts, sensitive internal spreadsheets) and which actions the AI does not take. A company policy for employee ChatGPT use is the natural place to record these rules.

Piloting means choosing one real research task, running it with a source log and agreeing on the success criterion beforehand, for example: every final claim has a link, an excerpt and a date, and none was discarded in review for lack of basis. To measure without fooling yourself, the article on how to measure an AI automation pilot offers metrics and decision gates. Handing over means delivering the document template, the checklist and the person responsible for review, and then deciding whether to expand, adjust or stop.

It also helps to be clear about what stays with people. Choosing the market, interpreting an ambiguous number and owning the conclusion are human decisions, as we discuss in what not to delegate to AI in business. To think about risk in a structured way, the NIST AI Risk Management Framework is a useful public reference: it organizes the work around understanding context, measuring risk and managing it.

If the research turns into published content, such as a report or an article, remember that Google advises producing content that is helpful, reliable and made for people. A cited source, clear authorship and a distinction between data and opinion help the reader and make the text easier to defend.

Frequently asked questions

Does AI with web search already deliver reliable sources? It delivers sources, but reliable is a separate question. Search can bring pages of uneven quality, and the summary can distort what the page says. So keep opening the link, finding the excerpt and checking date and scope before using the data.

Can I use AI to estimate the size of a market? You can use it to structure the reasoning, list assumptions and build the calculation, as long as each assumption comes from a verified source or is marked as a hypothesis. What you should not do is accept a final number the model presents without showing where each part came from. Treat the result as a working range for discussion, not as fact.

Do I need to check every claim the AI produces? Not with the same rigor. Claims that support investment, pricing or positioning decisions call for a primary source and a second check, while supporting observations can get a lighter review. What matters is that the rule is set beforehand and applied equally to everyone.

Can I put internal documents and customer data into the AI tool? Only if the tool and the use are approved by the company and there is a clear legal basis and purpose under the LGPD. For market research, you can almost always work with public information and aggregated data alone, which lowers the risk. When in doubt, leave the sensitive data out and record the decision in your usage policy.

Sources

• NIST: AI Risk Management Framework
• OWASP: Top 10 for LLM Applications
• Google Search Central: conteúdo útil, confiável e feito para pessoas
• ANPD: Autoridade Nacional de Proteção de Dados