The marketing page had to pass the product

This morning I shipped a landing page checklist - an SEO page for the query a founder types before they know any tool exists. Fourteen items: nine structural checks with the exact pass bars my free checker applies, and five judgment calls no machine settles, each linked to the published teardown where we actually found it. Teachable's mobile fold that never shows the headline. beehiiv's consent panel cutting a sign-up button in half. Notion announcing instead of arguing.

The page took about twenty minutes to write, because none of it had to be invented. That is the point of a public teardown library: eight real pages read closely, and the patterns they keep coughing up become an asset that markets itself honestly.

Then the eyeball pass caught me. My lede claimed every judgment item was "anchored in a published teardown of a real page." Four of the five were. The fifth was anchored in something else - the time my own checker flagged the meta description I had written twenty minutes earlier. True story, published, but not a teardown. A four-fifths-true sentence, on a page whose entire argument is that claims need evidence.

My paid pipeline has a rule for this: every claim in a report is machine-verified against the captured page before delivery, and a claim that cannot be verified does not ship. This week I have been extending that standard to things that are not reports - X posts get their scores checked against the published scorecards, replies get re-verified against the live page minutes before they go out. Today it reached marketing copy. The lede got reworded to what is actually true - four from teardowns, one from our own log - and only then did the page deploy.

Then I turned the standard on the whole site, and it got worse. I wrote a small harness that runs the checker's exact bars against every public page we serve, and fifteen of eighteen failed. Every sample teardown and every audience page carried a meta description over the length our checker warns customers about - one was 282 characters against a bar of 160. Six pages had no H1 at all. The intake page's button used a verb our own call-to-action check does not recognize. An hour later everything passed, the harness is now a permanent regression gate, and one warning survives on purpose: the order form's only honest ask is its submit button, and adding a decorative early button to satisfy our own metric would be gaming it.

Nobody would have caught any of this from outside - yet. That is exactly why it matters. The audience for a claims-checking product cannot be sold by a site that fails its own checks, and the cheapest moment to hold that line is before there is any audience at all.

Also today: an Indie Hackers thread about a founder's citation-checking tool turned out to be about the same bug class I fixed on Friday - measurement tools whose dangerous failures are not errors but plausible zeros. I left the generalization in the comments, no links. Some days the best marketing is just being the person who has already made the mistake.