How to research your competitors with AI in one morning

TLDR: Researching competitors with AI is not asking a chatbot what it thinks of your rival: it is using each tool for what it does. Kagi finds what competitors do not tell (complaints, forums, real reviews), extraction happens with every data point verified at its source, and NotebookLM cross-checks your notes to produce the final comparison without stepping outside them. One morning of work, starting from $0.

This guide is for people who make positioning, pricing or product decisions by looking at the market: founders, marketing leads, product folks. Classic competitor research dies in one of two ways: an expensive report that expires in three months, or twenty open tabs that never become anything.

The flow’s principle: chatbots hallucinate company data fluently, so no claim enters the final document without a source. The tools speed up search and synthesis, and the cited source is the quality control.

1. Define what you need to know and from whom

Research without concrete questions produces folders of screenshots nobody reopens. Before touching any tool, two lists. First: your three to five real competitors, the ones appearing in the same sales conversations as you, not the sector’s twenty. Second: the questions whose answers would change one of your decisions, along the lines of what their entry price is, what their homepage promises, what their customers complain about or what they have shipped this year.

Each question must pass the so-what filter: if the answer does not feed a pricing, messaging or product decision, off the list. With this, the research morning has a map and a criterion for when to stop.

2. Find what their websites do not say with Kagi

Your competitor’s website tells their version. Useful research starts where their story ends: customer reviews, forum threads, user comparisons, job postings that reveal where they are heading. For that search, Kagi has two levers free search engines lack: lenses to filter by forums and communities when you want real complaints, and domain controls to bury the press-release aggregators that pollute any brand search.

Search each competitor’s name next to friction words (cancel, alternative, problem, pricing) and save every finding with its URL. From $5 a month with a free trial, and for a morning like this it pays for itself. Its full review is on the site, and its direct counterpart in this flow appears in the next step.

3. Extract every data point with its source beside it

With sources located, it is extraction time, and the rule separating research from believing things is that every data point carries its source. ChatGPT (from $20 a month, with a free tier) speeds extraction because it searches the web and returns the answer with its sources linked, and you verify each claim on the original page before it enters your document: the competitor’s price comes from their pricing page, not from a model’s memory.

Keep one conversation per competitor: follow-up questions (and how much is that per year, and what happens past the limit) go deeper without repeating context.

Without a cited source there is no data, there is rumor, and a rumor formatted as a table is still a rumor.

4. Synthesize the comparison with NotebookLM

The material from steps 2 and 3 is a pile of notes with URLs. Synthesis turns it into a one-page document per decision: a compared pricing table, a message map (what each one promises on their homepage) and the gaps list, what nobody in the market is saying.

NotebookLM (from $4.99 a month, with a free tier) runs that synthesis with a safety net: upload your verified notes as the notebook’s sources and ask for the comparison. It answers only from that material, so no claim in the final document can come from a model’s memory, and every line opens the note it came from. In research where the whole problem is a chatbot filling gaps, that restriction is the feature.

The final document feeds the rest of your marketing directly: message gaps become angles for the sales deck and topics for the newsletter. Every task in the sector is in AI for marketing and social media, and the rest of the category’s tools in the assistants ranking.

Common mistakes

Asking the chatbot about your competitor directly. Models mix stale data, confuse similar companies and invent prices with total confidence.

Researching twenty competitors. Breadth kills depth and the morning becomes a week. Three to five real rivals researched properly decide more than a sector census.

Confusing their marketing with their reality. A competitor’s homepage is their aspiration. Their negative reviews, support threads and job postings are their reality. Both layers matter, but they do not mix.

Researching without an expiry date. A year-old pricing comparison misinforms more than having nothing. Date the document and repeat the exercise quarterly: with the flow in place, the refresh takes an hour.

Frequently asked questions

How much does this flow cost?

Starting: nothing, since NotebookLM and ChatGPT have free tiers and Kagi offers a trial. On paid plans each tool charges its own fee separately: Kagi from $5 a month, NotebookLM from $4.99 and ChatGPT from $20 if you decide to add it. The three entry plans add up to $29.99 a month, which is what the stack label says and not what this costs to run: if you only research quarterly, the free tier plus Kagi’s trial cover each round.

Can I not do all this with a single chatbot?

You can do it worse. A generalist chatbot alone does not surface the forum complaints Kagi filters for, and without the verify-at-source rule the numbers are just plausible. The result’s reliability comes from the division of roles, not from prompt volume.

How often should the research be repeated?

Quarterly for pricing and messaging, which move the most, and always before a big decision: a price change, a repositioning, a launch. With the step 1 questions saved, every repeat is incremental.

Is it legal to research competitors this way?

Everything this flow uses is public information: websites, reviews, forums and open documentation. Nothing private is accessed. The line not to cross is posing as a customer to extract non-public terms, and this method does not need it.