Every figure inside an AI generated content brief is a target, not a finding

A figure inside an AI generated content brief is a target the model made up, and it carries no information about the world. A count taken across many briefs is a different thing entirely and is a real measurement, because the denominator is the number of conversations and you ran them yourself. Printed, they are indistinguishable. Only the first can be used. Our own records hold both, which is the cleanest way to show the difference.

The practical rule is short. Strip every number out of a brief, keep the list of fields it asked for, then fill each field from a reading you took. We sell both halves of that, the taking of the reading and the writing that follows it, so the standard applies to our own pages before we quote it at anybody.

The measurement: what 102 conversations asked for

There is a tally behind that specification and we produced it ourselves. The 102 conversations were held with ChatGPT and with Gemini on 15 October 2026, spread across three markets, and each one began by putting an answer the engine had just written back in front of it. The question that followed was what a page would need before it earned the citation instead. Ten requirements recurred often enough to count, and 102 is the denominator on every row.

Four of the ten are enough to show what a tally looks like. Every one of the 102 wanted a date, a line saying when the facts were last checked and a statement of how often they get re-checked. 99 wanted a method a reader could take apart, handed over with the figure. 80 wanted the sample size. 68 wanted the results that had gone against the author printed too. The remaining six, with their counts, are set out in what the two engines said they need before citing a vendor page.

Each of those figures is a count of conversations we ran and saved. None of them describes how an engine works inside, and none of them is a ranking factor. They are a specification for a page, tallied, with a date on the tally.

The invention: what one brief asked for

The same programme holds 423 page briefs produced by asking engines what to publish. They are not evidence, and the clearest example is a single row. One brief asked for a page carrying 8,728 buyer questions.

We have never run 8,728 questions. Our two citation runs were 102 questions per market and then 210 per market, which is 612 readings attempted and then 1,260. A writer who took that brief at face value would have published a figure roughly seven times larger than anything we have measured, in a sentence that reads exactly like the counted rows above it.

The pattern repeats outside arithmetic. Briefs in the same set reached for authority they had not checked: one of them rested a page on a monetary threshold in a statutory provision that contains no figure, no table and no threshold of any kind. The brief was confident in exactly the register it uses for the parts that are correct, which is what makes the whole class of output dangerous rather than merely unreliable.

Why a brief gets this wrong by design

A brief is produced by asking a model what a citable page looks like. The model answers with the shape of such a page, and a shape includes slots where numbers go. It fills those slots with plausible values because a brief with empty slots reads as unfinished. Nothing in the request asked it to check anything, and nothing in the output marks the difference between a field name and a value.

So the output is genuinely useful at the level it was asked about. It tells you that a page needs a sample size, a window, a named engine list and a disclosure. It tells you nothing about what your sample size is, and the slot where your sample size belongs arrives pre filled.

Four steps that turn a brief into something you can fill

• Delete every number. Counts, percentages, prices, dates and sample sizes. Keep the field names and the order they appeared in.
• Mark each field with where its value will come from. Your own reading, an export, a document you can quote, or nowhere. The fields marked nowhere are the page you cannot write yet.
• Fill the fields from your own records, unrounded. A figure that sits where the brief put a target has to be traceable to a file you can open.
• Keep the fields you could not fill, in print. A page that names what it has not measured is doing the thing 68 of 102 conversations asked for, and it also stops the gap being quietly filled next time.

Judging generated output before you act on it is a general problem rather than one about briefs, and the version of it that applies to buying written work is set out in the five kinds of content supplier and how to judge output before you sign.

The two kinds of figure, side by side

Property	A count across briefs	A figure inside a brief
Where it came from	Conversations you ran and saved	A slot the model filled
Does it have a denominator	Yes, and you chose it	No
Can a stranger reproduce it	Yes, by running the same conversations	No, because there was nothing to reproduce
What it is about	A request two products made on a stated day	Nothing at all
Safe to print	With its date and its denominator	Never as a fact

What this page is not claiming

It is not claiming that briefs are worthless. Ours produced the ten counted requirements above, and those requirements are why our own pages carry a date, a denominator and a disclosure. The counts came from treating the briefs as a population to be tallied rather than as statements to be believed.

The tally also belongs to one date and to two products. The conversations reached ChatGPT and Gemini only, and the other three surfaces our product reads for a customer were never asked. We hold no measurement of how often a generated brief contains an invented figure, which is why this page gives named examples instead of a rate.

Questions about numbers inside an AI generated content brief

How do I spot an invented figure in a brief I have been handed?

Ask what file it would come out of. A real figure has a source you can name in one sentence: an export, a reading log, a document. A target has a round confidence to it and no file behind it, and the test takes about ten seconds per number.

Are the ten requirements above also something the engine invented?

They are a tally of what 102 separate conversations asked for, which is a count we produced from our own saved records rather than a claim any engine made about itself. The thing being measured is the request, and the request is the data.

What should I do with a brief that asks for a measurement I have not taken?

Take the measurement at the smallest honest size, or write the page without that claim and say which question it leaves open. A small number printed with its denominator is a different act from a large number printed without one, and 80 of 102 conversations asked for the denominator specifically.

Can I quote a brief as a source if I attribute it?

Quote it as what it is, which is one engine's description of what would earn a citation on a stated date. That is a legitimate and interesting thing to publish. Attributing an invented figure to the engine that invented it still puts the figure on your page, where a reader will treat it as yours.

Why publish the example at all, given it makes the briefs look bad?

Because the figure is memorable and the lesson transfers. A request for 8,728 buyer questions in a programme whose largest run was 1,260 readings is the clearest version of this failure we have, and 68 of 102 conversations asked to see exactly this kind of result alongside the useful ones.

The brief tells you the shape of the page and nothing else

Read a generated brief the way you would read a blank form: the field names are the work product, and the sample values are placeholders somebody forgot to clear before handing it over. Filling those fields from readings that actually happened, on a stated schedule, is what customers pay us for. The files our own values are computed out of, rather than typed from, are at aiknowsus.com/proof with the window on each one.