An engine's mistakes repeat, so review the rule, not the page

When a human writer gets something wrong, the mistake lives on one page. When an engine gets something wrong, it lives on every page the same rule produced, the ones already published and the ones scheduled for next week. So reviewing an automated blog has to start with a different question. Not "is this page right?" but "where else did this rule run?" A fix that touches only today's post is not a fix.

A writer makes slips. An engine makes policies.

Human errors are mostly idiosyncratic. A misremembered figure, a broken link, a sentence that went wrong at five in the afternoon. They scatter across an archive more or less at random, and an editor reading page by page catches them at roughly the rate they occur.

An engine does not have bad afternoons. It has rules, and it applies them with perfect discipline on every page it writes. That is the whole reason to use one: the voice holds, the structure holds, the cadence holds. The same property that makes its good habits compound makes its bad ones compound. A wrong rule is not a slip. It is a policy, enforced everywhere.

This is a place where automated content is genuinely worse than a freelancer, and we would rather say so than have a client find out. A freelancer's bad assumption gets noticed by an editor around the second article. An engine's runs silently until someone happens to listen. And scale is exactly what Google's guidance is concerned with: as we argued in Google doesn't penalize AI content, the problem it targets is unhelpful content produced at scale. An engine's mistakes arrive at scale by construction.

Our own record, read as classes

On 6 October, while writing about alt text for decorative images, we checked what our own hero images told a screen reader. The illustration step had been copying the image prompt into the alt text, cut at 120 characters mid-word, in English even on Finnish pages. It was not one bad description. It was every hero the blog had ever published.

Before 4 October, the Finnish version of every post linked to other posts at their English addresses, because the writing rule said to link to the related post and never said which language. Each instance looked like a small slip. Together they told every Finnish reader who followed a link that the language toggle was decorative.

We have written before about the storage failure that left most of our heroes broken for five days. What is worth repeating here is how it was caught: not by looking harder at the newest image, which was fine, but by a check that walks the entire archive. Every one of these defects was found on one page and existed on dozens.

Page review and class review ask different questions

Page review asks whether this post is accurate, argued and on brief. It is still necessary, and it is most of what a ten-minute approval should be. But it cannot see a class, because a class is invisible from inside any single member of it.

Class review starts when a defect is found, and it asks three things before anyone edits a word. Which rule produced this? Which pages did that rule touch? Is it still running? Engineering teams settled this long ago. Google's own SRE book defines a postmortem as including the follow-up actions that prevent an incident from recurring, and warns that without a formal process for learning from incidents they may recur indefinitely. Fixing the symptom is the start of the work, not the end of it.

The two repairs are separate, and they fail separately. Fixing the rule stops new instances. Fixing the instances repairs the archive. Do only the first and the old pages keep the defect forever. Do only the second and the engine quietly reintroduces it tomorrow morning.

We fixed the rule and left the instances

Here is the part that costs us something. On 6 October we fixed the illustration step, and every hero since has carried an empty alt, as a decorative image should. We checked this morning before writing: the 45 older heroes still carry the old truncated descriptions. By the argument of this post, that is half a fix, and it is the half a screen-reader user still meets on 45 pages.

The Finnish links are a different case, and a fair one. We decided to correct them as each post is next updated rather than in a single sweep, because the cost to a reader is one extra click on a page that offers a way back. A deferred class fix is a legitimate choice when the harm is small. What separates it from a forgotten one is that the class has been named and counted.

Ask "where else" before "how bad"

If you run an automated blog, or are buying one, treat every defect report as the start of a query rather than the end of a ticket. When you find a mistake, ask how many pages share it before you ask how bad it is on this one. And ask your vendor the same about theirs: when you last found a mistake, how many pages did it touch, and how do you know? An answer with a number means someone reviews by class. An answer with an apology means someone reviews by page, and the rest of the class is still out there.