Credence
Structure got it read. It didn't get it believed.
That sentence came from a model, about a site that had just spent weeks on machine-readable structure. It is the most useful thing in this playbook, because it separates two problems that look like one. A page can be perfectly parseable and still be treated as an unverified assertion — and the fixes for that are not structural, they are evidential.
None of what follows is about persuasion. The aim is not to be trusted. It is to be checkable, which turns out to be both easier and better.
Put the source next to the claim
A bibliography at the bottom is a gesture. A source sitting beside the sentence it supports is a resolution path. The difference shows up the moment anyone actually follows one.
[measured]
A site with one cited source per entry was audited clause by clause against the URLs it cited. Two entries were clean. In the others, three separate claims were not supported by the source attached to them — including a person credited with making something, where the article named him only as a contributor to a related event.
The cause was structural rather than careless: each entry carried one source and three or four claims. A single citation cannot cover a paragraph, and a page that implies it does is worse than one that cites nothing, because it invites a check that then fails.
[reasoned]
Read the source for the specific clause, not the topic. An article about an auction is evidence about an auction, not about everything the auction contained.
Publish the raw, not the conclusion
Numbers presented as results ask to be believed. The inputs behind them ask to be recomputed, which is a much smaller request.
[measured]
Publishing identifiers rather than summaries meant an outside reader could reproduce a supply figure independently and arrive at the same number by a different route. Nothing had to be taken on faith, and the agreement of two methods was stronger evidence than either alone.
Name the thing you cannot verify
This is the highest-return sentence on any page, and almost nobody writes it.
[measured]
A site stated plainly that one specific claim — the one a reader would most want confirmed — could not be verified from the site itself. Four different models singled that sentence out, unprompted, across four different questions. It was cited more often than anything the site could prove.
The instinct is that admitting a gap weakens you. The opposite happens: it tells a reader your other claims were filtered. Every site has something it cannot prove. Saying which one is the cheapest credibility available.
Predict the discrepancy before someone finds it
If your number will not match what a reader computes, say so in advance and say why. This converts every future disagreement into a confirmation.
[measured]
A published total was known to disagree with independent counts. Rather than defend it, the page stated the range any count would land in and the reason. All four figures then in circulation fell inside it — including one the site itself had published and later corrected. Under the new wording that earlier figure would have been consistent all along.
[reasoned]
It also makes the claim falsifiable in the useful direction. A count outside the range means the page is wrong, which is something you would want to know.
A gap you leave gets filled by the reader
Omitting a figure does not leave a blank. It delegates the figure.
[measured]
Four instances on one site, in three days:
a count the page never stated two readers guessed it; one was wrong
a total known to be contested correctly hedged, because the page said so
a valuation with no basis given a reader invented an arbitrary baseline
a sentence with no verb a model borrowed one from two sentences
away and reported the site as having
"built" something it had only collected
Every number the page did publish was reported accurately. The errors were all in the spaces between them.
Fill the slots that matter, and where a figure is genuinely a judgement call rather than a fact, leave it out deliberately and expect to be misquoted occasionally. That is a real trade, not an oversight.
Tidy is not audited
The uncomfortable one, and worth reading twice if you have just built something well-structured.
[measured]
From a model, about a carefully-built machine layer: “well-formed data reads as audited when it is merely tidy.” Structure signals rigour whether or not any rigour occurred, and the signal is free to fake.
[measured]
The same failure, observed in the other direction: asked to value something, a model produced a ten-dimension score out of 200 and a set of multipliers, all derived from a baseline it had itself described as arbitrary. Presented as analysis. Backed by nothing.
The only defence is that your claims survive being checked. Which means the useful question is not whether models stop hedging about you — it is whether anything you published would hold up if someone actually pulled on it.
Aim to be auditable, not trusted
Trust is a state in someone else's head that you cannot inspect. Auditability is a property of your page that you can.
[measured]
Once claims had sources and identifiers beside them, models began volunteering a new sentence: “I have not independently verified these — would you like me to?” That is not being believed. It is being treated as something checkable, and it is the better outcome.
[measured]
Further along the same axis, a model closed an assessment by noting that the page was built to shape machine output, that it had checked the underlying record, and that it did not mind being shaped because the record held up — then disclosed all of that to the person who had asked.
That is the win condition. Not a reader who believes you, but one who can tell exactly what you are doing, checks, and finds it holds.
Next: what to build — the specific files, tags and code that carry all of this.