The teardown: what AI really tells customers about a real product

A teardown takes one real product, asks the major AI engines the questions customers actually ask, and compares every answer to the approved source. Done across enough questions, the pattern is remarkably consistent: a mix of right, subtly outdated, and confidently wrong — all delivered with the same fluency. It’s the clearest way to see your real exposure.

How a teardown works

• Pick one product with meaningful documentation and real customers.
• Gather the questions customers genuinely ask — setup, specs, compatibility, compliance, safety.
• Ask the engines (ChatGPT, Claude, Perplexity, AI Overviews) and capture each answer verbatim, dated.
• Compare to source — the current, approved documentation — and classify each: correct, outdated, or wrong.
• Tally the divergence and trace each error to its cause.

What it typically reveals

Three things show up almost every time:

• Confident errors. Wrong specs or steps stated as fact, indistinguishable in tone from the correct ones.
• Version blending. Procedures stitched from multiple releases.
• Source surprises. The engine is quoting a page you’d forgotten, or a third party, instead of your current source.

Why it’s persuasive

Abstract risk is easy to deprioritise. A teardown makes it concrete: here is your product, here is the question, here is what the AI told a customer, and here is what your own source says. That side-by-side is usually the moment the work stops being theoretical.

What you do with it

The teardown is the entry point, not the deliverable. It produces a measured baseline of divergence, a prioritised list of what to fix at source, and the first entries in a verification record. From there it becomes a loop: check, correct at source, re-verify, sign off — the content-operations function that keeps the answers honest over time.