Answer-first is 25 words, and three tools had to agree on how to count them
In September 2026 two checkers in this portfolio reported different word counts for the same page. The cause was a function nobody had looked at in a year, and fixing it changed how every number in the family is produced.
Updated
Two visibility checkers in this portfolio were found reporting different word counts for the same page. The cause was str_word_count(), which counts alphabetic runs and therefore silently drops every figure on a page. Fixing it meant agreeing one measurement layer across three tools that otherwise score completely differently on purpose.
This is a small bug with a large lesson attached, which is the only kind worth writing up.
What happened
Two sites in this family publish page checkers. They are meant to disagree: one asks whether an AI system can understand and trust a page, the other asks what should be fixed and whether you can fix it yourself. Different registries, different weights, different wording. Running both and getting two different scores is the product working.
Running both and getting two different word counts is not.
Same page, same afternoon, two numbers. And the moment that happens, every number either tool has ever produced becomes a question.
The cause
One engine used PHP's str_word_count(). It counts runs of alphabetic characters. On a spec table, a price list, a page of measurements, it quietly reports fewer words than the page contains, because it does not consider 7200 or 3.4GHz or 2019 to be words.
The defence at the time was that the alternative behaved badly on non-English pages. That was the wrong reason to leave it, and it took a while to see why: the function drops every figure on a page in any language, so it fails worse at the thing it was supposedly protecting.
There was a second, uglier one underneath it. Both engines used strip_tags() to turn markup into text, which deletes tags outright instead of replacing them with a space. A table row comes out as Served over HTTPS8Structural. That is one word where there were three, and every number inside a table stops being a word at all.
The fix, and the discipline that came with it
Tags become a space. Words are counted by splitting on whitespace. Both engines got the same two functions, character for character.
Then the more important part. Each engine's file now carries a header naming exactly what must never drift between them:
- how markup becomes text
- how a word is counted
- the answer-first window and word bar: 1500 characters, 25 words
- the timeout, redirect limit, byte ceiling and rate limit
- the URL validation and SSRF discipline
Change one, change all of them in the same session. This tool is the third in that family and it was built to that contract from the first line rather than retrofitted into it.
Why 25 words and 1500 characters
Because a bar that cannot be measured is not a bar, it is advice.
"Answer the question early" is true and useless. It cannot fail a page, so it cannot be argued with, so nobody ever fixes anything on the strength of it.
The measured version: within 1500 characters after the closing H1, there is a paragraph of at least 25 words. Now a page passes or fails, the failure comes with evidence attached ("the first paragraph after the H1 carries only 19 words"), and somebody can go and do something.
Both numbers are judgements rather than discoveries, and I would rather say that than dress them up. 1500 characters is roughly the amount of markup a breadcrumb, a byline and a share widget occupy before the writing starts. 25 words is roughly the shortest passage that stands alone once removed from the page around it. Neither is derived from a study. Both are defensible, both are published, and both can be argued with because they are stated precisely enough to argue with.
The general version
Three tools measuring the same page with three different rulers is worse than one tool with a mediocre ruler, and it is worse in a specific way: it destroys the ability to compare anything, including a page with itself over time, which is the single most valuable thing measurement offers.
If you run more than one measuring instrument, the scoring can differ all it likes. The measurement underneath has to be one thing, written down, in a place the next person will read before they change it.
Filed under: answer-first, measurement, word-count