Competitor Analysis With AI: Sources, Signals and Limits

Competitor Analysis With AI: Sources, Signals and Limits

To run competitor analysis with AI, first gather the public material (website, service pages, ads, reviews, job postings, social profiles), hand it to the AI with a closed question, and require every claim to come with a quote and a link to its source. Then separate what is verifiable fact from what is interpretation, and only then decide. AI speeds up reading and comparison, but it does not know what a competitor does beyond what the competitor publishes.

Key takeaways

• AI is good at reading, organizing and comparing a lot of public material. It is weak at guessing another company's strategy, margins or internal results.
• You choose the sources. If the input is weak or outdated, the comparison will be weak, even when the text sounds convincing.
• Every claim needs a cited source and a label: fact, inference or hypothesis. Without that, the report becomes well-written opinion.
• Web pages are untrusted input. Instructions hidden in a site can influence the model, so the AI reads but does not take actions.
• Start from a concrete decision (repositioning a service, revising an offer, responding to a move) and not from a generic report about the market.

What kind of question can AI answer well about competitors?

The quality of the analysis depends on the question. Open questions such as "what is my competitor doing?" produce vague summaries. Closed questions tied to a decision produce usable material.

Good questions for the AI, based on material you supplied:

• Which services does each competitor list on its site, and how does it describe each one?
• What promises and conditions appear on the sales pages (lead time, stated guarantee, contact channel)?
• Which complaints and compliments repeat in public reviews, and on what topic?
• What kinds of roles is the company hiring for, and what does that suggest about where it is investing?
• Where does our offer say something theirs does not, and vice versa?

Questions the AI cannot answer safely: how much the competitor earns, what their margin is, how many customers they have, whether their operation runs well internally. Any answer to these will be a guess. If a number or conclusion like that shows up without a source, treat it as an error.

Which public sources are worth gathering?

The rule is simple: only include what you can open, read and point to. Each source reveals a different kind of signal.

• Company website and service pages: how the company positions itself, who it speaks to and what it promises.
• Pricing or plan pages, when public: how the offer is structured, what is charged separately, which conditions appear. Record the date you checked.
• Ads and the ad libraries of the platforms: which messages and offers the company is testing.
• Reviews on Google, Reclame Aqui (Brazil's consumer complaint site) and marketplaces: recurring customer pain points and praise. To see how to handle this kind of text, read the article on handling consumer complaint sites with AI.
• Open job postings: hiring signals, tools mentioned, areas that are expanding.
• Published content (blog, videos, newsletters, posts): topics the company wants to own.
• Official documents and public records, when relevant to your industry.

One important detail: save the content at the time you consult it (copied text, screenshot or saved file, with date and address). Pages change, and without that record you cannot show what the AI read when someone questions a conclusion.

How do you ask AI for a comparison with cited sources?

What works is to supply the material and ask the model to work only on it. This is the same principle as retrieval-augmented generation (RAG): the answer rests on documents you provide, not only on the model's memory.

A practical step-by-step:

• Define the decision. Example: deciding whether to expand a service that two competitors already offer.
• List the competitors and the comparison dimensions (offer, audience, contact channel, social proof, stated conditions).
• Collect each competitor's public material and label every file with a name, address and date consulted.
• Give the material to the AI one competitor at a time, asking for a structured profile with the same dimensions.
• Ask for the comparison across profiles only afterward, and require each line to point to its source passage.
• Check a sample of the citations against the original pages before trusting the rest.
• Record what is fact, what is inference and what is hypothesis, and carry only what passed the check into the decision.

Instructions that usually make a difference in the request:

• Use only the documents provided. If the information is not in them, answer "not in the material".
• For each claim, cite the document name and the exact passage that supports it.
• Do not fill gaps with assumptions. If you must assume, mark it as a hypothesis.
• Split the answer into three blocks: facts, inferences and open questions.
• If two documents contradict each other, show both passages instead of choosing one.

Because the volume of text can be large, it helps to understand the limits of what a model can consider at once. The article on what a context window is in AI helps you decide when to split the material into parts.

How do you separate fact, inference and hypothesis?

This is the step that prevents the most decision errors. AI writes everything in the same confident tone, so the label has to be demanded explicitly.

• Fact: it is written in a source you opened and can cite. Example: the services page lists a certain kind of support. The test is simple: someone can go to the link and see the same thing.
• Inference: a reasonable conclusion drawn from facts that may still be wrong. Example: several support job postings suggest the company is expanding its support team. Record which facts back the inference.
• Hypothesis: a plausible explanation without direct proof. Example: the competitor probably prioritizes larger clients. It is useful for guiding a test or a question, never for closing a decision.

A useful habit is to ask the AI itself to attack the report. In a second pass, ask it to point out which conclusions depend on a single source, which use guessing language, and what would change if the conclusion were wrong. The article on using AI as devil's advocate before deciding covers this technique in detail. And before circulating the report, run the text through a review checklist for AI-written content.

What does it look like in practice? An illustrative scenario

Illustrative scenario: a regional equipment maintenance company wants to decide whether to create a periodic-visit plan, because it suspects competitors already offer something similar.

The team defines the decision (create the plan or not) and picks three local competitors. For each one, it saves the service pages, the most recent public reviews and the open job postings, noting address and date. It then asks the AI for a profile per competitor: what is described about recurring maintenance, what conditions appear, which complaints repeat about delays or follow-up.

In the comparison, the AI reports that two competitors describe periodic visits and one does not mention the topic. That is fact, with a quoted passage. It also reports that the second competitor's reviews mention slow follow-up. That is fact with a caveat, because it depends on a sample of reviews. The sentence "the market is moving toward recurring plans" appears in the summary, but with no source. The team marks it as a hypothesis and does not use it.

The resulting decision is small and testable: offer the plan to a limited group of current clients and track interest, instead of reshaping the company based on an AI report. The analysis helped design the test, not replace it.

What are the limits and risks?

First, coverage. The AI only sees what was supplied. What a competitor agrees by phone, negotiates on WhatsApp or delivers in the field appears on no public page. Draw conclusions about what is published, not about what the company is.

Second, outdated material and source bias. Reviews tend to concentrate extreme cases, and a services page describes commercial intent, not delivery quality. Always state in the analysis where each signal came from.

Third, web content as untrusted input. A site can contain text written to manipulate an AI model, and the risk grows when the tool browses on its own and can take actions. OWASP guidance on prompt injection prevention addresses this. In practice: let the AI read and summarize, with no permission to send messages, fill in forms or access internal systems during the research.

Fourth, personal data and ethics. Public information about a company is different from information about people. Profiles of owners, employees and customers who appear in reviews can be personal data, and LGPD (Brazil's data protection law) still applies even when the data is accessible. Keep the focus on the company and its offer, avoid building dossiers on individuals, and do not use fake accounts, scraping that violates platform terms of use, or any way of obtaining material that is not public. If your own company's material (internal spreadsheets, proposals) gets mixed into the analysis, define beforehand what may go into the tool. A company policy for employee ChatGPT use is a good starting point, and if the service processes data outside Brazil, it is worth knowing the rules on international data transfer under LGPD.

Fifth, risk management. The NIST AI Risk Management Framework reminds us that AI use should have a defined context, an owner and a review step. For competitor analysis, that means one person who owns the report and has the authority to discard conclusions without a source.

How to apply it with a method: map, limit, pilot, transfer

The same method we use for any AI process applies here, without turning the analysis into a large project.

• Map: describe how the company does competitor research today, who does it, how often and which decision it feeds. Record current time and quality as a baseline.
• Limit: define the allowed sources, the data that must not enter the tool and the actions the AI may not perform (sending, publishing, accessing accounts).
• Pilot: choose one real decision, run the analysis while recording sources and checks, and agree beforehand on a success criterion, for example whether the analysis reduced rework or improved the quality of the decision.
• Transfer: document the script and the profile template, define who maintains them, then decide whether to expand, adjust or stop using it.

If the team plans to repeat this process often, the script can become a standardized workflow, with human review kept in every round. For deciding what should stay with people, see what not to delegate to AI in business.

Frequently asked questions

Can AI research competitors on the internet by itself?

It can, in tools with browsing, but the result is harder to control: you do not know well what was read or what was left out. For important decisions, it is safer to choose the sources, give them to the model and require a citation for every claim.

Is it legal to analyze competitors with AI?

Analyzing information that a competing company published itself is a common market practice. The precautions are to avoid violating terms of use and restricted access, to avoid processing personal data unnecessarily and to respect LGPD. If you are unsure about a specific use, consult a lawyer.

How do I know if the AI made up information?

Ask for the literal passage and open the source to check a sample. If the passage does not exist or says something else, discard the claim and increase checking on the rest of the report. Claims without a source should be classified as hypotheses.

How often should the analysis be repeated?

It depends on how fast your industry changes and on the decision the analysis feeds. A practical rule is to repeat it when there is a new decision to make or a visible move by a competitor, and to keep earlier versions with dates so you can compare what changed.

Sources

• AWS: What is retrieval-augmented generation
• OWASP: prompt injection prevention
• NIST AI Risk Management Framework