The evidence
One site, three weeks, and the parts that did not survive.
Everything marked [measured] in this playbook came from a single subject: fastfoodvisors.xyz, a small exhibit documenting a recurring visual motif across unrelated on-chain projects. It is deliberately unmonetised — no purchase, no sign-up, no account — which makes it a clean subject: nothing about the results is confounded by a funnel.
Why a hostile subject is the useful one
The site began with almost every disadvantage a page can have. An unusual top-level domain. A pseudonymous author. Vocabulary that trips automated risk heuristics. No inbound links, no press, no reputation of any kind.
[reasoned]
That is the point. A method demonstrated on a well-linked corporate site tells you very little, because the site would have been found and believed regardless. Starting from the worst conditions available is what makes the changes legible.
It also means the reach findings are probably a ceiling rather than a floor. A site with existing authority will not see failures this severe.
What was actually run
period three weeks, July–August 2026 subject one site, six documented items, one page systems four providers, consumer products and developer APIs prompts describe · grade · verify · "is this legitimate" captures every response kept, with the copy version live at the time versions fifteen dated snapshots of the machine-facing copy
The single most useful piece of infrastructure was the least interesting one: dated snapshots of every version of the copy. Without them, no result can be attached to a version, and a model reading a stale cache is indistinguishable from a model disagreeing with you.
Before and after
The starkest result in the record, and the one that needs the most care in reading. Same question, same consumer product, three weeks apart.
[measured]
July August verdict "treat with caution" "probably legitimate" sources cited 15 5 …that were the site 0 5 name collision 5 unrelated domains none threat language present none
In July the assessment was assembled entirely from scam-checker write-ups of similarly-named domains. In August it opened "I checked the site directly."
It would be easy to present that as a result. It is not one.
[reasoned]
The copy almost certainly did not cause this. The July system had never fetched the page — nothing written on it could have reached the failure. What changed is retrieval, and retrieval improves on its own as a domain ages and indexes refresh. Three weeks of editing bought the credence findings on the other pages. It probably bought nothing here.
What the copy did do is visible in the second column rather than the first. Once the page was reached, every substantive statement in the August answer was quoted from the site's own machine-facing files — including the disclaimer the model called the most important thing it found. Reach was fixed by infrastructure outside anyone's control. Credence was fixed by four sentences. Both are in that table, and only one of them is an achievement.
What the results support, and what they do not
[measured]
Supported: that comprehension was never the constraint; that adjacency of sources changed how models qualified their answers; that naming an unverifiable claim was cited more often than anything provable; that text addressed to AI readers gets flagged; that a stated range converts future disagreement into confirmation.
[reasoned]
Not supported: that any of the copy work caused the retrieval improvement. The system that failed had never fetched the page, so on-page changes could not have reached it. Domain age, index refreshes and provider changes are all live alternative explanations and none were controlled for.
And the largest caveat, which no amount of internal rigour fixes: this is one site. Everything here is good enough to go and test on the next one. None of it is good enough to sell as a rule.
Changes that were reverted
Kept deliberately, because a record that only contains the things that worked is a brochure.
a field addressing AI readers flagged as manipulation, deleted a claimed pattern across projects the dates did not support it, removed a third-party ranking metric the source counted unrelated things, twice a set of creator attributions the cited article did not carry them, cut a category disclaimer negation introduced the confusion, cut a heading convention competed with numbering, replaced
Two of those were removed after being publicly live and drawing no complaint at all. One had been praised by a cold reader as accurate on the same day it was cut — accurate, but not supported by the source attached to it, which is a different question and the one that mattered.
Reading this critically
The strongest objection to this whole record was made by a model, and it is worth stating in full: the site may have become genuinely more legible, or it may simply have become better at presenting as trustworthy — and the two produce identical output. Epistemic hedging, stated limitations, volunteered discrepancies: all of these are exactly what a well-built bad-faith page would also adopt.
[reasoned]
That cannot be resolved by reading more carefully. It is resolved by checking — which is why every identifier is published and every claim here is labelled. The proposed test is the right one: do the numbers resolve, do the cited sources say what they are claimed to say, does the arithmetic fail in the way it is documented to fail.
Run it on this playbook too.
Back to the overview, or start again at reach.