Competitive research is a folder of their pages, not a vibe
Using AI for competitive research and market analysis means gathering what a rival, a substitute, or a market actually published, then using a model to sort your notes. It does not mean asking "what is the competitive landscape" and pasting the paragraph into a deck. That paragraph is a genre. It sounds like analysis and cites nothing you can open.
This workflow keeps you on pages and filings you can file. It uses the Perplexity practice to find those pages and a writing model only after the notes exist. The tool directory orients you to products. None of them have your competitors' private numbers, and you should be suspicious of any tool that implies it does.
Define the rival as a choice a customer has
A competitor is not only a company in your category slide. It is the other way a customer gets the job done, including a spreadsheet, a freelancer, or doing nothing. Write three of those down before you open a tool. If you ask a model for your competitors first, you will get the list from its training, which is a year out of date and confident.
For each name, write the question you need answered in public sources: pricing page, docs, job posts, a filing, a changelog, a review page you will not treat as a survey. Job posts are particularly honest. A company hiring five people for a feature is a stronger signal than a slogan.
Collect pages, then notes
Search for the primary page. Save the URL and the date. Do not save the model's retelling as the source. If the pricing page is a login wall, write "not public" and move on. Do not guess the price. A guessed price in a deck becomes a number someone repeats to a customer.
Use a model on the file you downloaded: summarize the changelog into the features that appeared this quarter, list the claims on the homepage, extract the integration names from the docs. The summarisation guide is the same discipline on PDFs. Check two bullets against the page before you trust the rest of the list.
When two sources disagree, keep the disagreement. Markets are not required to be consistent. A model that reconciles them into one story has done damage.
A small comparative table you fill
| Question | Your note | Rival A, with URL | Rival B, with URL |
|---|---|---|---|
| What job does the homepage promise? | |||
| What does the public price page actually list? | |||
| What did the changelog add lately? | |||
| What role are they hiring? | |||
| What do they refuse to support, in the docs? |
Empty cells are allowed. An empty cell is more professional than a invented one. If a colleague pressures you to fill it, the answer is the URL you do not have.
Analysis is the sentence after the table
Only now ask what it means for a decision you are actually making: a feature, a price, a page you will publish, a segment you will ignore. The model can propose interpretations. You pick one and attach it to a row in the table. An interpretation that does not point at a row is a mood.
"They are winning" is not analysis. "Their docs describe a export we do not have, and two job posts mention that export" is analysis. You can decide whether the export matters to your customers. The model cannot.
Avoid fake precision. Market size paragraphs with three significant figures and no dataset are a known failure mode. If you need a size, cite the dataset's own page or do not state a size. The SEO writing guide has the same ban on decorative statistics, because a blog and a deck fail the same way.
Ethics and access
Use public pages and documents you are allowed to hold. Do not ask a agent to log into a rival's customer portal, scrape behind a login, or evade a technical control. That is not competitive research. The agents guide draws the same line for browser tools. A test account of your own product is fine. A test account of theirs, created to see private pricing, is not a clever growth hack.
Do not feed a customer's confidential roadmap into a public chatbot in order to "compare." Use notes with the names removed, on a tool your company allows.
A monthly cadence
Once a month, refresh the table for the three rivals that matter, not for a list of twenty. Update the date. Archive the old note so you can see what changed. A agent that rebuilds a forty-page landscape every Monday will bury the change in a new preface.
Share the table, not the chat. The chat is how a stray fake price escapes. The table has URLs.
Frequently asked questions
Q: How do I use AI for competitor analysis?
A: Name the alternatives a customer has, collect their public pages, extract notes with a model, and check two bullets per page. Write the decision in a sentence that points at a row. Do not ask for a landscape and paste it.
Q: Can AI tell me the size of my market?
A: It can point you at a dataset if one exists, and it can invent a number if you insist. Use the dataset or leave the number out. Invented precision is worse than a gap.
Q: What sources are fair game?
A: Public sites, docs, filings, job posts, and reviews you label as anecdotes. Not private portals, not data you do not have rights to, not a customer's secrets.
Q: How often should I refresh?
A: Monthly for a short table, or when you are about to make a pricing or feature decision. Continuous monitoring is a feed you will ignore. A dated table is a tool.
Q: Are review sites useful?
A: As anecdotes about language customers use, yes. As a score, no. A star average is not a market study. Quote a specific complaint and look for it in your own support mail before you treat it as a trend.
Q: Can I automate this with an agent?
A: You can automate collection of public URLs into a folder. You should not automate the conclusion, and you should not automate logins. Read the trace.
Q: What if the model names a competitor I had not considered?
A: Treat it as a hypothesis. Open their site. If the job they do is a real alternative, add a column. If the page is dead or unrelated, drop the name. Do not add it because the list looked short.
Q: How do I present this to a founder who wants certainty?
A: Show the table and the empty cells. Certainty that is not on a page is a story. You can still recommend a decision. Label it as a decision, not as a fact the model found.
Bring the URL
Competitive research with AI is a faster way to fill a table you are willing to check, and a faster way to embarrass yourself if you skip the check. Build the three-column table this month. Leave blanks. The blog covers the research and writing habits around it. The glossary will translate a vendor's "market intelligence" claim into ordinary words: it searches the web, and you still have to read.
