We reversed our biggest content decision after reading the data

A few weeks ago we changed our highest-volume content format from one page per tool to one page per task within a tool. The reasoning was written down at the time:

Task-level queries carry higher intent. Someone searching “record a walkthrough in Linear” has a specific job; someone searching “Linear demo” is browsing. And task-level pages are structurally harder to make thin, because each one describes a genuinely different workflow rather than forcing the same paragraph shape across hundreds of near-identical pages.

We also noted that the market had converged on this shape. Three competitors publish hundreds of pages that way.

Both halves of that reasoning were fine. The second half was evidence of something other than what we thought.

What the convergence actually was

When we read those pages instead of counting them:

One competitor’s 115 tutorial pages have a median of 186 words and score lowest in the corpus on every measure we applied. They are not about that company’s product. They are how to add an emoji in Slack, how to add a watermark in PowerPoint.

Another’s 99 how-to pages are the same task repeated across laptop models: Acer Aspire, Acer Nitro, Acer Predator, Acer Swift. 100% of each page’s headings are shared with its siblings once proper nouns are blanked out.

So the sentence “the market has converged on task-level tutorials” was true and useless. What it meant was “this is where the cheap pages go”. Converged is a word that sounds like validation and describes a fact about volume.

What we changed, and what we kept

We kept the format. The intent argument was always the stronger of the two and it still holds. Task-level pages are better than tool-level pages for the reasons we wrote down before we knew anything about competitors.

What we changed is how we treat it. The shape now carries a burden of proof rather than a head start. Specifically:

We removed the cross-product from our plan. Tasks multiplied by tools is a grid, and a grid gets filled. Every tutorial now has to originate from an actual run, so pages come into existence one at a time when someone does the task, and there is no list to work through.

We stopped treating competitor page counts as a prior at all. The competitor data now has one use: deciding which rivals to research, because sitemap structure is genuinely informative about what a company is investing in. It has no role in deciding what is good.

Why write this up

Partly because it is a real reversal and we said we would publish those.

But mostly because the mistake is instructive and it is not a stupid mistake. Looking at what successful companies do is normally sound. It failed here for a specific and generalisable reason: we had data on what they published and no data on what worked. Page counts are what a company decided to make. They are not results.

That distinction is easy to state and easy to lose, and we lost it for about a week while building a plan on top of it.