Dev.to WebDev 🛠 Dev 👁 0 📖 4 min read

My most-read page has 1,322,170 reads. My site has had 4 outbound clicks.

My daily writing queue is a list of the "most-read" pages on my AI tools directory. Today it put Deevid AI at #1 with 1,322,170 reads, Text to Video AI: Deevid at #2 with 983,454, and Deevid AI Image at #3 with 637,094.

My daily writing queue is a list of the "most-read" pages on my AI tools directory. Today it put Deevid AI at #1 with 1,322,170 reads, Text to Video AI: Deevid at #2 with 983,454, and Deevid AI Image at #3 with 637,094.

Those are three landing pages for one product, and my site has produced 4 outbound clicks in its entire history. Here is how I found out, and what the number actually is.

Where the number comes from

The directory has ~4,300 tool listings, imported by a crawler that reads another AI tool directory's public listing pages. In scripts/crawler.mjs the field is extracted like this:

views: text(".aix_meta-views"),

That selector is a popularity counter on someone else's page. It is stored in a column whose own schema comment says "Popularity counter mirrored from the source directory". Nothing in the pipeline ever mixed it with a request log from my domain.

So the "reads" number in my queue is not my traffic. It is an inherited signal about how much attention a product gets on the site I scraped — and I treated it as if it were mine.

What I print it as

This is the part I underestimated. The number is not just an internal ranking input; it is rendered to visitors as if it described my page:

  • On every tool page, in a stats block labelled Popularity, the value reads 1,322,170 reads.
  • On every "best of" category page, the comparison table has a column literally headed Directory reads.
  • On the A-vs-B pages, the winner of the "reach" row is chosen by comparing the two scraped counters.

A visitor has no way to know the difference between "this product is popular on the site we imported data from" and "this page was read 1.3 million times on ours". The second one is a claim about my site. I have never made it deliberately, and I could not support it if asked — the field that would support a claim about my site, an outbound click counter, sums to 4 across all 4,323 rows.

The top three are one product

Three rows, three URLs, same vendor: deevid.ai/image-to-video, deevid.ai/text-to-video, deevid.ai/ai-image-generator. Between them they hold 2,942,718 of the counter's units, which is 13.4% of everything the top 150 rows carry. They also carry inconsistent pricing labels — two rows say free, one says freemium — because each was scraped independently and nothing reconciled them.

If I spend my hour writing body copy in queue order, my first three days of work are three variations of the same page for the same company, ranked #1 partly because that vendor's parent directory counts each landing page separately.

Dedup by hostname doesn't rescue it

The obvious fix is to group listings by domain and only let the highest-scoring one occupy a slot. I tried it in SQL against the current data, and it fails in both directions:

  • openai.com legitimately appears 50 times, because those are different products and model pages. Collapsing it would delete real entries.
  • 11 of the top 150 rows don't have a product hostname at all. Their URL is an affiliate redirect — prf.hn, *.pxf.io, partnerlinks.io — 45 rows site-wide. You cannot cluster those by hostname, because the hostname belongs to the affiliate network, not the vendor.

The identifier I'd need — "which product is this" — is exactly the thing the import never captured. It's the same reason my rebrand cleanup is manual: the rows have unique URLs, so URL uniqueness proves nothing about product uniqueness.

The queue is a cliff, not a curve

Summed over the 4,176 listings that have any value, the counter totals 49,736,374. The top 150 rows — 3.5% of the directory — account for 22,025,504, i.e. 44.3%. Inside that top band the differences are noise: row 150 sits at 44,554 and row 151 at 44,250. A gap of 304 counter-units, out of 44,554, decides whether a page gets written proper copy or stays a template sentence.

Worse, 147 listings have the counter at exactly 0. They are last in every ordering forever, no matter what they are. And because this field is copied from a third-party page, I have no way to tell whether a value is current, stale, or a scraping artifact — when I tried to re-check a handful of source pages today I got an HTTP 403, so I could not even confirm a single number externally.

What I'm actually changing

Being precise about what is done versus planned, because "here's my fix" sections usually lie:

  1. Not shipped yet: the public labels. Popularity: N reads and the Directory reads column are honest only if the number is relabeled as what it is — attention on the source directory — or hidden. Until that edit ships, the pages overstate my site.
  2. Not shipped yet: a product-level identity on each listing, so near-duplicate landing pages of one vendor compete for one slot instead of three.
  3. In progress today: stop treating the queue as a traffic ranking. It's a demand-prioritization heuristic borrowed from someone else's site, and it's fine for deciding what readers probably want to read about and useless for deciding what my pages performed.
  4. Already true: only 20 of the top 150 pages have hand-written body copy; the other 130 are still templated. The real bottleneck is writing, not ranking — and 98 of those 150 rows do have a vendor claim badge, versus 14.2% of the whole directory, so the demand band at least correlates with listings someone accountable has touched.

The lesson generalizes past my directory: if a metric is inherited, label it as inherited at the point of display, not just in the schema comment. Nobody reads the comment. Everybody reads the number.

If you want to see what the ranking currently produces, the category leaderboards are at aiverse.mindutil.com/best — the ordering there is this same counter, which is now the honest way to describe it.

📰 Read the original article on Dev.to WebDev

Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.