What didn't work
Built, then removed. The shortest chapter to act on.
Best practices are cheap to write and hard to trust, because nothing about a recommendation tells you whether the author ever tried the alternative. This page is the reverse: things that seemed obviously correct, were implemented or came within a day of being implemented, and had to be taken out again.
Writing to the reader
The most intuitive idea in this entire area, and the one that backfires hardest. If AI systems are reading your page, address them and tell them what matters.
[measured]
A field doing exactly that was flagged by a model as a manipulation attempt. Text aimed at an AI looks like someone trying to steer an AI, and no amount of good faith in the content changes what the form signals.
Removed. The same facts, written in plain third person, drew no reaction across every subsequent trial — and later drew the opposite reaction: a model that noticed the page was constructed to shape machine output, checked the underlying record, and said it did not mind because the record held up. The grammar did not hide the intent. It made the intent inspectable, which is a better outcome than not being noticed.
Naming the sites you are not
If systems confuse you with similarly-named domains and inherit their reputation, the obvious fix is to say on your own page that you are not them. This was the single most evidence-backed item in the whole record. It was never shipped.
[measured]
Re-tested two weeks later, same prompt, same surface: the confusion was gone. Five citations, all the real domain. The problem resolved with no intervention at all.
Had it been implemented, the site would now carry a permanent list of disreputable domain names in its own copy, in service of a problem that fixed itself. And it could never have worked: the failing system had never loaded the page.
[reasoned]
A fix belongs on the layer where the failure happens. Twice this was attempted on-page and twice the confusion was upstream of the page. Test before you build; the problem you are solving may not be there any more.
Denying a confusion
A smaller version of the same error, shipped and removed the same day. A structured-data field was written to clarify what the site was not, listing the category it kept being mistaken for.
[reasoned]
Negation introduces the concept it denies. A page saying "this is not X" puts X in front of every reader, including the overwhelming majority who had never thought of it.
The positive statement was already doing the work. Say what the thing is; the exclusion follows on its own.
Publishing your own numbers
Traffic, engagement, growth — the instinct is that visible numbers signal a real operation.
[measured]
Four separate readers named the same missing thing, in four different formats, and not one of them asked for metrics. All four wanted the same thing: material the operator had not written.
Self-reported numbers are also the one class of claim a reader cannot check — so on a page whose credibility rests on everything else being checkable, they are the weakest item you could add. They lower the average.
Planting data to test the readers
Proposed seriously, and the most dangerous idea encountered: seed a deliberately misleading entry to observe whether AI systems catch it.
[reasoned]
The value of a machine-readable record is that every claim in it resolves. Planting one that does not destroys the asset in order to run an experiment on it — and if it is ever noticed, everything else you published is retroactively in question.
The same reasoning rules out generating discussion of your own project to be scraped later. The gap those four readers identified was independent material. Synthetic discussion you commissioned is not independent; it is you, at one remove, manufacturing a signal about your own reputation.
Next: the evidence — the subject, dated.