Voice and Citation: 618 Trials Across Four AI Engines

Abstract

We tested whether the register a page is written in changes how often an AI assistant will cite it. Across 618 trials on a frozen four-engine panel, the same factual claims written in a casual-precise register — plain sentences, exact figures, no throat-clearing — were cited more than the same claims written in a corporate register, with a +33 percentage-point lead on Claude specifically. Two mechanisms carried the effect: numbers written sloppily were reproduced 0% of the time, and any second-person address to the AI read as distrust on the Claude family. Gemini was largely indifferent to voice. The practical implication is that writing for citation is closer to writing plainly for a careful reader than to any keyword craft.

Method

We built matched pairs of passages: the same factual claim, written once in a casual-precise register (short sentences, exact figures, the answer stated first) and once in a corporate register (setup, abstraction, hedged numbers). Each pair was run as a trial against a frozen panel of four engines — ChatGPT, Claude, Gemini, and Perplexity — and scored on whether the engine cited or reproduced the passage. Across 618 trials on four AI engines, a casual-precise voice (plain sentences, exact figures) was cited more than a corporate register — a +33 percentage-point lead on Claude.

Results

Casual-precise won. The plain register led the corporate one by +33 percentage points on Claude.
Sloppy numbers vanished. Figures written imprecisely were reproduced 0% of the time — a vague number is a number the model won't carry.
Addressing the AI backfired. Any second-person address to the reader-as-AI read as distrust on the Claude family and hurt the passage.
Gemini shrugged. It was largely indifferent to voice, responding to structure and sourcing more than tone.

Practical implications

Writing for citation is not a keyword craft; it is writing plainly for a careful reader. State the answer first, keep the figures exact, cut the throat-clearing, and never perform for the model. The effect is strongest on Claude — the most rival-dense, brand-authority- sensitive engine — which is exactly where a plainly-written page has the most to gain.

Limitations

This is one panel at one point in time, on our own matched-pair corpus. The engines drift, and effect sizes are engine-specific — the casual-precise lead is largest on Claude and near-zero on Gemini, so this is a finding about a family of models, not a universal law. We publish the reading and will re-run it as the panel changes.

Source: The AI Citation Institute voice-and-citation experiment, 618 trials across ChatGPT, Claude, Gemini, and Perplexity (2026-07 panel).

Cite this record

Open access

AI Citation Institute. "Voice and Citation: 618 Trials Across Four AI Engines." Research record 2026-001, 2026-07-23. https://aicitationinstitute.org/research/voice-and-citation-618-trials (CC BY 4.0).

Released under CC BY 4.0 — quote it, chart it, cite it. All we ask is attribution back to this record.

Ex corpore, auctoritas.