Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 4 min read

I rewrote my 20 highest-traffic pages. The other 130 still show three sentences I never wrote.

My directory lists 4,323 AI tools. Every day I get one hour to make the highest-traffic pages better, and I have been spending that hour wrong. Here is the number that finally showed me. What "written" means on

My directory lists 4,323 AI tools. Every day I get one hour to make the highest-traffic pages better, and I have been spending that hour wrong. Here is the number that finally showed me.

What "written" means on my site

Each listing has a description body. Most of them were filled in by a crawler, and the crawler copy is a template with the product name slotted in. My enrichment queue (scripts/seo-queue.mjs) strips four known boilerplate sentences, a templated opening line, affiliate-tracking URLs, and the tagline; whatever word count is left is the only part I'm willing to call "written." Target is 80+ unique words.

I ranked all 4,323 rows by the mirrored read count, took the top 150, and ran that filter. The result:

  • 20 pages pass. 130 don't.
  • Those 130 templated pages are 86.7% of the top 150 by count.
  • They are 64.0% of the top 150 by traffic β€” 14,103,474 of 22,025,504 mirrored reads.
  • My 20 hand-written pages, the ones I've been grinding on, hold 36.0%.

So the effort curve and the traffic curve point in opposite directions. I write a page, feel good, and it's one of the lower-read ones in the set. The bulk of the people landing on my "best of" rows are reading text a script wrote.

What a reader actually sees on an unwritten page

The single most-read page I haven't rewritten is "Muse Wonder" β€” 211,026 mirrored reads. The raw description in my database runs longer than the reader sees, because my renderer (the same filter) drops the boilerplate sentences at render time. After filtering, the entire "About" paragraph a visitor reads is three sentences:

Muse Wonder is a freemium AI tool in the LLM models category. It currently ranks #4 in the LLM models category. Common tags for this tool include Latest AI.

I pulled that off the live page with curl, no cookies, to make sure I wasn't describing the database instead of the site. Those three sentences contain: one price label, one category name repeated twice, one rank, and one tag. There is no fact in there about what the product does that the vendor would recognize as theirs. And the next most-read unwritten pages β€” "Gemini 4 Argon" at 210,792 reads, "GPT-6 Sol and Luna" at 207,352 β€” render the identical shape with a different proper noun.

The part that actually stings

81 of those 130 templated pages carry a "claimed" badge β€” the vendor came to my site, verified they own the listing, and said "yes, this is ours." Some of those show "Verified Gold."

Think about the sequence of events that implies. A vendor was enough of a human to find their listing and claim it. Then they left. Nobody β€” not them, not me β€” put a single real sentence about their product on the page the claimer's own customers would land on. The claim badge and the empty body coexist on 62% of my unwritten top pages. The overall claim rate across the whole directory is only 14.2% (614 of 4,323), so this isn't random noise: high-traffic rows are far more likely to be claimed and far more likely to still be a template.

It isn't a free-vs-paid story

I expected the gap to sit in one pricing bucket, which would at least make it a clean bug. It doesn't. The 130 templated pages split free 67 / freemium 44 / paid 19. The free bucket is biggest simply because free is the biggest bucket overall β€” about half my directory is labeled free. So the /free-ai-tools list people filter on is, for its highest-traffic rows, a list of pages whose only body is those three template sentences. The filter is honest about the price label (that comes from the claim, where it can be); it says nothing about whether the description is worth reading, because for 67 of them there isn't one.

What I'm changing

The queue already sorts by mirrored reads, so it has been handing me the right order β€” I was just treating it as a "nice to have" backlog instead of the top of a funnel. Two concrete changes:

  1. Stop counting "written" per page and count it per read. My daily hour should be judged by how much of the 14.1M-read templated block I convert, not by how many pages leave the todo list. Converting the top 10 by reads is worth more than converting 40 near the tail.
  2. Prioritize the 81 claimed-but-empty pages specifically. That's a list I can email, because the claim already gave me a contact and a demonstrated willingness to engage.

None of the fix is clever. The number that was missing is the traffic-weight one. A page with zero unique words and 211,026 reads is not a minor content task β€” it's the most-read thing on my site that I refuse to describe.

I'm keeping the whole thing reproducible rather than asking you to trust my arithmetic: the same filter lives in one file (scripts/seo-queue.mjs), and every figure above is findMany over one SQLite table plus a cookie-less curl against a live URL. The full directory, including the free filter that 67 of these pages feed, is at aiverse.mindutil.com/free-ai-tools if you want to go find a three-sentence page yourself. The experiment I'd rather run is the boring one: rewrite the top ten, then check whether the reads move.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.