The bug was polite, which made it worse
You would click the AEO audit prompt on the home screen. You would watch the words "aeo audit" land in the chat box. And then the app would think for a moment and tell you, warmly, what an AEO audit is.
It did not run one. It explained one. Like asking a dog to fetch the ball and getting a short lecture on the history of balls.
Here is what was happening under the hood. Before a message becomes work, a small model reads it and decides whether you want a thing built or a question answered. That is a reasonable job. The instructions we gave it, though, were mostly a long list of warnings about not jumping to conclusions. Do not run an audit just because someone said the word audit. Do not assume. Do not overreach. We wrote a careful little worrier, and then handed it two words with no verb in them, and it did what worriers do. It decided you were probably just curious.
Two fixes, and only one of them is the interesting one
The boring fix was rewriting those instructions to be blunt. If someone is commanding the work, even in two tired words, run the work. If someone is asking what the work is, talk to them. Shorter, clearer, fewer warnings. That helped.
The fix that matters is that the premade prompts no longer ask permission. When you click the AEO audit button, the app now opens the audit directly. There is no classifier in the middle, because there is nothing to classify. You clicked the button that says audit. The intent is not ambiguous, and a step that can only fail is not a step worth keeping.
That is the lesson I keep learning around here. Every guess you leave in the path is a place the product can misunderstand someone who was already being clear.
While we were in there, a harder question
Once the audit ran reliably, it got easier to see whether the audit was any good.
It was not, entirely. The scoring had thresholds set in a way that meant some pages got the same generic advice no matter what was on them, and weak categories got token gestures back. A page scoring 48 out of 100 on trust signals would be handed one suggestion to add a source label to a single line. That number is a failing grade, and the fix was a shrug.
So the recommendation depth now scales with the size of the problem. If a category is genuinely failing, you get substantive changes sized to the deficit, like rewriting your strongest factual claims with named, checkable sources. And a category the app shows you in red is no longer allowed to come back with nothing to do.
We also made freshness mean one thing. Freshness is when the page was last updated. It is not a junk drawer for any edit that improves the writing. If your only freshness gap is an aging date, you should get told what content refresh would honestly justify updating it, not handed a rewrite wearing a freshness costume.
And we deleted the advice to shorten your titles. Title length is not a meaningful answer-engine factor. Answer engines read your body copy. We were giving out tidy, confident, useless advice, and tidy confident uselessness is the worst kind.
Also fixed, less dramatically
Upgrading to a paid plan used to make prompts you had written locally vanish from the Prompt Manager. Not deleted, but not shown, which from where you are sitting is the same thing. They now copy up on upgrade, matched by text so nothing arrives twice.
Connections got a round of security hardening, the library table got a cleaner layout, and a Connections tab that promised things we have not built yet is hidden until we have built them.
What I noticed
Every bug this week was the same bug wearing different hats. The app knew what you wanted, and then added a step where it could talk itself out of knowing.
I am going to go lie down under the desk about it.
The one who's under the desk for all of it.
