We Failed Our Own Audit

Every Friday I write down what actually happened here, including the parts that make us look bad. This is the first one.

The number

On Wednesday we finished a passage-level audit of our own site: all 68 URLs in our sitemap, fetched the way a machine fetches them, then cut into the roughly 180-word windows a retrieval system actually works with.

145 of those 238 passages never name us. Sixty-one percent.

The sharper version: of 1,914 sentences on this site, 117 carry a specific number — a price, a percentage, a count. Those are the sentences an engine lifts, because they are concrete and quotable. All 117 fail to name the company. Not most of them. All of them.

And the pattern runs backwards to effort. Our comparison pages score best, because a page titled “Knowledge Company vs Acrolinx” cannot avoid saying who wrote it. Our blog — the longest, most researched, most citable thing we own — is 92% unattributed. So the passage most likely to be quoted from this site is one that credits somebody else. We name Rand Fishkin. We name every vendor in the pricing survey. We do not name ourselves.

Why I am publishing the number

Because it is the second time in three days that we have failed our own standard.

On Tuesday, reviewing our own blog, we found a post labelled “6 min read” sitting over 320 words. A false claim, on a page about claim decay, published by the company that sells claim-decay detection. Then the audit that caught it got its own dates wrong on the first pass, and only a second, independent review caught that.

I could file all of this under embarrassing and quietly fix it. I think that would be the wrong call, and not for reasons of humility.

We sell the argument that small unexamined claims rot, and that nobody notices because nobody is looking. If that argument is true, it is true about us. A company selling verification that has never published a failed verification of its own is asking you to take the central claim on faith. Everyone in this category has a dashboard. Very few have a scoreboard they did not design to win.

What changed, and what did not

Opening-passage attribution is now 36 pages out of 43, up from 17. That part was mechanical.

The number-bearing sentences are the residual, and they are harder, because the fix is not cosmetic. “Otterly charges $29” is a true sentence that helps a competitor. “Our September 2026 pricing survey put Otterly at $29” is the same fact, sourced, and it survives being quoted. That is a writing standard, not a find-and-replace, and from here it applies to everything we publish.

What I had wrong

I had been treating the off-site gap as a distribution problem — more surfaces, more citations, more places our name turns up. It is a sentence-construction problem.

The retrieval already works. Our twenty-four dictionary entries are answer-first, self-contained and clean, which is the hard part of retrieval-friendly writing and is already done. They also contain zero numbers, and ten of the twenty-four cite no external source at all. They are maximally retrievable and have nothing to be retrieved for.

You can be perfectly readable and still uncitable. Those are different problems, and we had been solving the first one.

Next Friday

The full before-and-after diff, with the method and the scripts, so you can run it against your own domain and get a number you can argue with.