This is a repeatable pattern, not a one-off prompt. It works for any competitor that has a public website, a public changelog or docs, some press coverage, and — ideally — a pricing page. It does not work well when you need to analyse something private (internal strategy decks, unreleased roadmaps).
When to use it
- You need to brief a founder or exec on a specific competitor this week.
- You're scoping a new feature and want to see how three adjacent tools handle the same problem.
- You're entering a category and want a lay of the land before you commit to positioning.
Skip it when you already have good primary research — customer interviews with people who have evaluated the competitor will always beat this.
Inputs you gather (30 minutes)
- The competitor's homepage, pricing page, product or feature pages, and any public changelog.
- One recent review source — G2, Capterra, Trustpilot, or a Reddit thread. You want unfiltered language.
- One recent press or investor source — a funding announcement, Crunchbase row, a YC launch post.
- Two or three screenshots if the product has a public-facing UI worth reading.
Save them into a single local folder. Paste the text of each source into a working doc with clear
delimiters (--- SOURCE: homepage ---). The LLM performs better when sources are labeled than
when they are bulk-dumped.
Step one — the structured read (20 minutes)
Ask the model for a structured read against a fixed schema. The schema is the whole trick; without it, you get generic marketing summary. Use something like:
Read the following sources. Produce a teardown with these exact sections:
1. What they say they do (one sentence, their words).
2. What they actually do (one sentence, your words — be blunt if the claim doesn't match the product).
3. Who the buyer is. Role, seniority, what they are accountable for.
4. Who the buyer is NOT. Segments they tried to serve and quietly dropped.
5. Pricing shape. Per-seat, per-usage, tiered, enterprise-only. Call out the friction points.
6. Distribution. How do they actually reach new buyers? PLG, sales-led, community, SEO, paid?
7. Three strongest specifics in the product (feature, not benefit — name the thing).
8. Three weakest specifics. Things that feel thin, missing, or bolted on.
9. The one thing they would panic about if a well-funded competitor copied it tomorrow.
10. The one thing they'd be happy to lose.
Cite the source for every claim with the delimiter label. Do not generalise.
If the model refuses to be blunt, push back once: "You are being too diplomatic. Name the specific thing, not the category." Don't wrestle with it more than once — move on.
Step two — the adversarial pass (15 minutes)
Start a fresh chat. Paste the structured read from step one as input. Ask:
The following is a competitive teardown. Play the role of the competitor's Head of Product reading
this in private. For each section, note:
- Which specific claim is weakest or most overstated.
- Which claim you'd privately agree with but publicly push back on.
- One thing the teardown missed entirely.
This step catches ~30% of the errors the first pass makes with high confidence. If the model's adversarial response is vague, give it a concrete persona: "You are an early employee who left the company in 2024. Write like you're DMing a friend."
Step three — the human checkpoint (30 minutes)
Now you read. Not the sources — the two outputs. You are looking for:
- Claims with no citation. Delete or verify. No negotiation on this.
- Adjective stacks. "Innovative, modern, AI-first" is a smell. Rewrite as a specific behaviour.
- Suspiciously round numbers. Check every statistic. LLMs hallucinate them confidently.
- Things you already knew to be wrong. The model is confidently wrong about 10–15% of any given domain. Your job is to catch those.
Mark each claim as verified, plausible, or unchecked. Do not ship unchecked claims.
Step four — the one-page brief (20 minutes)
Convert what survived into a one-page brief with a fixed structure:
- One-sentence positioning (theirs, then yours-about-theirs).
- Three specifics that make them dangerous.
- Three specifics that make them beatable.
- The wedge you'd use if you were competing against them next quarter.
- Two things you'd want to validate with a real customer before acting on any of this.
Save the structured read and adversarial pass as appendices — not in the brief itself. The brief goes to decision-makers; the long versions stay available when someone asks "how did you conclude that?"
What this is not
It is not a substitute for customer research. It does not tell you why people actually buy. It tells you what's visible from the outside, structured enough to reason about. That's useful for a first pass and for staying oriented; it's not sufficient for a pricing change or a positioning pivot.
Variants worth trying
- The internal tool teardown. Same structure, applied to something your company already ships. Forces you to treat your own product as an outsider would.
- The adjacent-category teardown. Apply this to a competitor from a neighbouring category to see what conventions transfer. Especially useful before a category expansion.
- The quarterly re-run. Schedule it. Most competitive intelligence decays within a quarter because changelogs and pricing move. A re-run is cheap if you have the input folder preserved.
If you build your own variant, write it down. The teardown is only as good as the schema — and the schema gets better every time someone refines it against a specific mistake.