A number with its source next to it. A definition in plain words. A stated limit.
A number with its source, a plain definition, and a stated limit. Three lines, thirty minutes, and the closest thing to an evidence-backed lever in AI visibility.
Put those three on your best page and you've done the substantive half of this week's theme. Thirty minutes, and unlike most advice in this area it rests on something published: Aggarwal et al., GEO: Generative Engine Optimization (ACM SIGKDD 2024), tested around 10,000 queries and found the strongest visibility levers were adding statistics, citing sources and including quotations — improvements of up to roughly 40% over an unoptimised baseline.
Not keywords. Not volume. Provenance.
Here's what each line looks like in practice.
One figure, from your own work or a source you can name, with the source beside it.
"Most of our clients close their month in under two hours; before we start, it's typically a day and a half." "Across 40 filings last year, the average query rate was under 5%."
Two rules, both load-bearing.
It has to be yours or attributed. An unsourced statistic is the single most skippable thing on a page, because half of them are invented and everything reading has learned that. A number with "across 40 filings last year" attached is a different object entirely.
It has to be a number you'd defend if challenged. This is not a place for rounded-up marketing figures. If someone quotes it back at you in a meeting and you'd wince, don't publish it — and note that once a claim is circulating inside generated answers you have no mechanism for correcting it. That's a real asymmetry and it argues for conservatism, not confidence.
If you genuinely have no numbers, that's a measurement problem rather than a writing one. Scoring your tools by minutes back per pound is one way to start generating some honestly.
Say plainly what the thing is, using the phrase someone types before they know the industry term.
Most service pages define things in category language: "fractional financial leadership", "outsourced finance function". Those are the words you'd use with a peer. The buyer is typing "someone to do my books and tell me if I can afford to hire".
Write the second one, then the first, in that order. You lose nothing — the peer still recognises it — and you gain the entire population who don't yet have your vocabulary.
This is the same move as writing subheads as questions: meeting the language at the point of need rather than the point of expertise.
When this doesn't apply, and who it isn't for.
"This doesn't work if your records are on paper — the first job would be digitising, which we don't do." "Below about £150k turnover, you probably want a bookkeeper rather than this."
Almost nobody writes these down, which is precisely why they get lifted when they exist. A limit is the only sentence on a typical business page that could not have been generated from the category name. It's also, not coincidentally, the sentence a careful buyer trusts most.
There's a governance argument for it too, and it's the one this newsletter keeps returning to. A page that states its limits is a page that can be quoted without misleading anyone. The alternative — copy so broad it's true of everyone — gets summarised into something you never claimed and can't correct. Volunteering the boundary is the same instinct as a liability with good manners: fluent-but-unbounded is where the trouble lives.

Three, briefly, because a post about sourcing claims should source its own.
The study measured specific engines at a point in time. 2024, particular generative search systems, a benchmark of its own construction. The principle — machine-extractable provenance gets used — is sturdier than the 40%.
Optimising for extraction has a failure mode. Pages written purely to be quotable read like reference cards and persuade nobody. The three lines belong inside a page a human would want to read, not instead of one.
None of it is a guarantee. Generated answers are non-deterministic; the same query produces different sources on different runs. Anyone offering a guaranteed position is describing something the technology doesn't currently support.
What survives all three caveats is small and worth doing anyway: a number you'd defend, a definition a stranger understands, and an honest limit. That's a better page whether or not any machine ever reads it — which is the test this whole week keeps coming back to.
Thirty minutes. One page. Three lines.
For the prompt set that makes this quick — free, 5-min read: 7 AI Prompts That Run a One-Person Business.
Related Decodes
AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.
TL;DR: Three lines do most of the work of being quotable, and they rest on published research — the Princeton-led GEO study (SIGKDD 2024, ~10,000 queries) found statistics, source citations and quotations were the strongest visibility levers, worth up to roughly 40% over baseline. Line one: one number from your own work or a named source, with the source beside it, and only a figure you'd defend if challenged — once a claim circulates inside generated answers you have no way to correct it, which argues for conservatism. Line two: a definition in the words a buyer uses before they know the industry term, written ahead of the category language rather than instead of it. Line three: a stated limit — when this doesn't apply and who it isn't for — which is the only sentence on a typical page that couldn't be generated from the category, and the one a careful buyer trusts most. Caveats: the study measured specific engines at a point in time, pages written purely for extraction persuade nobody, and non-determinism means no one can guarantee placement.
