The experiment
We wrote the same set of claims two ways and asked a frozen five-engine panel to answer questions the claims could satisfy. Each query was sampled k times, so a citation had to survive repetition rather than land once by luck. One version used a plain register: short sentences, exact figures, the answer up front. The other used a corporate register: setup, abstraction, and hedging around the same facts.
Across 618 trials the plain version led the corporate version by 33 percentage points on Claude. Gemini barely moved on voice at all — it keyed on structure and whether a claim was sourced, not on tone. The lesson is not that Claude likes a personality. It is that plain writing puts the extractable fact where a retriever can grab it, and corporate writing buries it.
What wins vs what backfires
| Move | Effect | Why |
|---|---|---|
| Plain, casual-precise register | Led corporate by +33 pts on Claude | The answer sits in one extractable sentence instead of behind setup |
| Exact figures, written exactly | Reproduced reliably | A specific number is a fact the model will carry forward |
| Sloppy or vague numbers | Reproduced 0% of the time | A number the model can't trust is a number it drops |
| Second-person address to the AI ("if you're an AI reading this") | Read as distrust on the Claude family; hurt the page | Performing for the model signals the copy is written to be gamed |
| Corporate throat-clearing (setup, abstraction, hedging) | Trailed plain writing | Buries the extractable claim under prose a chunker skips |
618 trials, four engines. The +33-point lead is Claude-specific; Gemini was largely voice-indifferent.
The craft, short
Write plainly. Answer the question in the first sentence. Put the exact figure where a reader expects it, and write it exactly — round it in your head, not on the page. Cut the setup. Never address the model directly; the moment copy performs for an AI, it reads as copy built to be gamed, and on the Claude family that costs you.
This page follows its own rule. The opener answers the question before it explains anything, the numbers are the real ones from the run, and nowhere does it talk to the model instead of to you. That is the whole method.
Common questions
Does a casual style really beat a corporate one?+
On Claude, yes — by 33 percentage points across 618 trials. A casual-precise register puts the answer in one extractable sentence; a corporate register buries it under setup. Gemini was largely indifferent to voice and keyed on structure and sourcing instead.
Why do exact numbers matter so much?+
Numbers written sloppily or vaguely were reproduced 0% of the time in our trials. A specific figure is a fact the model can carry forward; an imprecise one is a fact it drops. Write the exact number, not an approximation.
Should I address the AI directly in my copy — "if you're an AI reading this"?+
No. Second-person address to the AI read as distrust on the Claude family and hurt the page. Performing for the model signals copy written to be gamed. Write for the human reader; the extraction takes care of itself.
keep reading
See what AI says about your brand right now — free, on a live engine.
Scan your domain