Most competitive teardowns die in a Slack thread. Someone signs up for a rival product, screenshots the onboarding, pastes a few observations, and the artifact evaporates by the next planning cycle. Three months later a new PM asks "what does Competitor X actually do for activation?" and the work starts from zero.
The failure isn't effort — it's that teardowns are treated as one-off reactions instead of a repeatable process with a durable output. An LLM doesn't fix that on its own. Used carelessly, it makes it worse: a model will happily describe a pricing page it never saw and invent a feature that sounds plausible. Used as a structured research assistant on inputs you supply, it compresses the boring parts of a teardown so you spend your time on judgment instead of transcription.
We ran this workflow across several real teardowns to find where AI helps and where it quietly fabricates. Here's the version that survived.
The four-pass workflow
Treat a teardown as four passes, each with a defined input and output. The point of separating them is that the AI's job changes at each stage — and so does how much you can trust it.
Pass 1 — Capture (you, not the model). Sign up for the competitor. Walk the real flows: signup, first-run, the core job-to-be-done, upgrade prompts, cancellation. Save raw evidence as you go — screenshots, the actual pricing table copied as text, the onboarding email sequence, support-doc URLs. This is the one pass you cannot delegate, because it's the only source of ground truth. Everything downstream is only as honest as what you captured here.
Pass 2 — Structure. Hand the model your raw notes and screenshots and ask it to organize them into a fixed schema: positioning, target segment, pricing tiers, activation path, key features, gaps. The constraint that matters: tell it to fill a field with UNKNOWN — not observed whenever your evidence doesn't cover it, and never to infer. This single instruction is the difference between a useful summary and a confident fiction.
Pass 3 — Analyze. Now ask sharper questions against the structured data. Where does their activation path assume context a new user won't have? Which pricing tier is doing the conversion work? What's the implied roadmap given what they shipped last? The model is good at surfacing patterns across your notes; it is not good at knowing which pattern matters to your product. You own that call.
Pass 4 — Decide. Convert observations into claims your team can act on: what to copy, what to deliberately not do, what to watch. A teardown with no decision attached is just a book report.
The model will fabricate specifics most readily in Pass 2 and Pass 3 — exact prices, seat limits, launch dates, percentages. Anything numeric or dated that you didn't personally capture should be treated as unverified until you check the source. Ask the model to tag every factual claim with the evidence it came from; claims it can't attribute are claims to delete.
Why a fixed schema beats a clever prompt
The instinct is to write one elaborate prompt that produces a polished teardown in a single shot. It looks impressive and it's a trap. A monolithic prompt blends evidence you captured with inferences the model generated, and you can no longer tell them apart in the output. It also produces a different shape every time, so two teardowns can't be compared side by side.
A fixed schema solves both. When every teardown fills the same fields — positioning, segment, pricing, activation, features, gaps, your decision — three things happen:
- Comparison becomes mechanical. You can stack five competitors in one view and read down a column instead of re-reading five prose documents.
-
Gaps become visible. An empty
activationfield is a research to-do, not a silent omission you discover during a roadmap debate. -
The model has less room to wander. A schema with explicit fields and an
UNKNOWNescape hatch gives the AI a smaller surface to hallucinate into than an open-ended "write me a teardown."
The schema is also what makes the workflow repeatable by someone who isn't you. A new PM can run the same four passes against the same fields and produce a comparable artifact without inheriting your private context.
Where the artifacts live
A teardown is only an asset if the next person can find it. Scattered docs and Slack threads fail this test, which is why most competitive intelligence has a half-life of weeks. You want one searchable repository where every teardown shares a schema, links to its raw evidence, and carries a date so readers know how stale it is.
A structured workspace handles this well: a database where each row is a competitor, each column is a schema field, and the page body holds the captured screenshots and source links. Filtered views let you read all teardowns in a segment at once, and the date field makes freshness obvious at a glance.
Whatever tool you pick, the requirement is the same: shared schema, linked evidence, visible dates. The tool is replaceable; the discipline isn't.
Version your teardowns instead of overwriting them. When a competitor ships a redesign, add a dated entry rather than editing the old one. The diff between last quarter's teardown and this one is often more useful than either snapshot alone — it shows you the direction they're moving, not just where they are.
What the AI does not replace
It's worth being precise about the division of labor, because the workflow breaks when it blurs. The model is a fast, tireless research assistant: it structures, summarizes, finds patterns, and drafts. It does not have an opinion about your strategy, it cannot see anything you didn't show it, and it will assert false specifics with the same confidence as true ones.
So the human keeps three jobs that don't delegate: capturing real evidence in Pass 1, deciding which observations matter in Pass 3, and committing to a position in Pass 4. The AI makes those three jobs faster by clearing the transcription and organization out of your way. That's the whole bargain — and it's a good one, as long as you don't mistake the assistant for the analyst.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)