How to Do a Competitive Analysis with AI (Without Publishing Wrong Facts)

A competitive analysis is mostly two activities: gathering facts about competitors, and making sense of them. Language models are close to opposite in skill across those two.

Good: structure, synthesis, first drafts

Give a model 200 reviews and it will find the themes. Give it two pricing pages and it will build a clean comparison. Give it your notes and it will draft a battlecard outline in seconds. This work is reliable because the source material is in front of it.

Bad: recalling facts about companies

Ask what a competitor charges and you will often get a confident, well-formatted, wrong answer. Training data ages, pricing pages change weekly, and models fill gaps rather than admit them. Even guides that recommend this approach tend to add that the output can be misleading, which is a strange thing to build a competitive analysis on.

Everything below follows from that split: use the model for thinking, not for remembering.

This is the sequence we would run for any serious competitive analysis. It takes roughly a morning for three competitors, versus a week done manually.

1. 1Define scope and pick the competitor set. Name the decision the analysis serves and choose three to five competitors you actually lose deals to.
2. 2Gather sources yourself. Save pricing pages, homepages, docs, changelogs, review exports, and call transcripts. This is the step people skip, and it is the one that determines accuracy.
3. 3Feed the model your sources, not its memory. Upload or paste the material and instruct it to work only from what you provided.
4. 4Prompt for structure, not conclusions. Ask it to extract, organise, and compare. Ask for the reasoning, and ask it to flag anything it could not find.
5. 5Verify every factual claim against a primary source before it goes anywhere near a customer.
6. 6Turn it into a decision, then put it on a refresh cycle. An analysis nobody acts on and nobody updates is a document, not intelligence.

Any capable model handles the reasoning. What actually differs is how easily it takes in your sources, whether it cites them, and where your data ends up. Capabilities and plan names change frequently, so confirm current details with each provider before committing.

The practical answer: use whichever your company already pays for and trusts with data, and put the effort into sourcing and verification instead. Those two steps affect quality far more than the choice of model.

The workflow above is most of the value; the prompts are the part people come for. Two are printed here in full because they show the discipline the rest inherit, and the whole library is maintained separately so each one can carry the thing a prompt on its own cannot: what has to be in the context window before you send it.

Positioning teardown

Note the last sentence. Without it, a model asked for a competitor’s wording will produce plausible marketing copy the company never published, inside quotation marks, where it reads as evidence rather than as a claim.

Battlecard first draft

The unsourced marker is what turns checking from a re-read into a pass down a list, which is the difference between a card that gets verified and one that gets shipped unverified.

The rest of the library

Each of these is maintained as its own page with the sources it requires, where the output should land, and in several cases what came back when we ran it against a live page.

Several of these have a structured destination rather than a chat window to land in: the feature comparison matrix, the pricing matrix, the SWOT builder, the sales battlecard template, the win/loss report template and the executive brief template. Pasting a model’s output into a structure your team already reads is what stops a good analysis from being ignored because it arrived in an unfamiliar shape.

This gets asked constantly and answered rarely. A competitive analysis involves your most sensitive material, and a consumer chat account is not the place for it.

• Unreleased roadmap and pricing plans. Your future moves are the single most valuable thing a competitor could learn.
• Named customer data. Deal details, contract values, and anything a customer told you in confidence. This is often a contractual obligation, not a preference.
• Raw call recordings with personal data. Transcripts carry names, phone numbers, and opinions people did not expect to be processed elsewhere. Anonymise first.
• Anything under NDA. Including competitor material a prospect forwarded you.

Two practical guardrails. Use a business or enterprise plan with contractual data protections and training turned off. And anonymise before uploading: replace customer names with identifiers, and strip anything you would not want read aloud in a meeting.

Run this before any AI-assisted analysis reaches a rep, a customer, or a board deck. It is the difference between a useful shortcut and a credibility problem.

1. 1Every price, plan name, and billing unit checked against the competitor's live pricing page today.
2. 2Every feature claim traced to a page, doc, or changelog entry you can link.
3. 3Every quote confirmed verbatim in the source, not paraphrased into something the reviewer never said.
4. 4Statistics traced to the original study rather than a blog citing a blog.
5. 5Anything the model could not source removed, not softened.
6. 6Competitor claims stated factually and fairly. Never disparage: it is a legal risk and it makes your team look unserious.
7. 7A date on the document, because a competitive analysis without one will be quoted long after it stopped being true.

If that verification burden feels heavy, it is the honest cost of the approach. It is also the argument for tooling that maintains sourced competitor data continuously rather than rebuilding it by hand each quarter, which is the case for competitive intelligence software rather than a quarterly rebuild. Where AI genuinely helps and where it adds noise is a different question, and competitive intelligence with AI is the one that answers it.

It is worth being precise about what an assistant actually changed here. Drafting collapsed from days to minutes. Verification did not move at all. The cost of a competitive analysis has therefore not fallen so much as relocated: nearly all of it now sits in sourcing and checking, which is why the checklist above is long and why it stays long. A team that speeds up drafting without speeding up sourcing has built a faster way to produce claims it cannot stand behind, and the failure shows up in the worst possible place, which is a rep repeating one of them to a buyer.

That inverts what is scarce. The bottleneck is no longer writing capacity, it is having a trustworthy, dated, sourced picture of each competitor to draft against. Flares maintains that picture across corporate structure, positioning, feature parity, pricing model, customer voice and distribution, and refreshes it as things change, which shortens the checklist rather than deleting it. It never reaches zero. Anything going to a rep, a customer or a board still needs a person to open the source and read it, and if you would rather assemble the picture yourself, the public sources are all named.