Free multi-client voice drift check: compare recent pieces with the client's reference

By Jack Stovell · published 2026-09-21 · updated 2026-09-20

Drift is what happens when a client's voice moves away from its own reference over time. It's measurable: a score, plus a named list of the habits that changed (sentence length, contractions, pronoun mix, that sort of thing). No score, no argument. Just numbers.

Paste the reference piece below, then paste up to three recent pieces for the same client. The check runs both through the same counted and judged features and hands back a voice match score, plus the specific habits that moved between them. Nothing is stored.

What drift looks like in numbers

Here's the thing about voice: it's not vibes, it's countable. Every piece gets measured on sentence length and variance, contractions, commas, exclamations, semicolons, brackets, questions, ellipses, pronoun mix, article balance, word length, signature phrases. That's the counted half. The judged half covers formality, humor, expressiveness, confidence, opening style, argument structure, rhythm, complexity, clause ordering, specificity. Weighted, calibrated, squeezed into a number between 0 and 1. Full method's at /reference/voice-measurement-framework, if you want the mechanics.

Against a saved profile, 0.85 to 0.95 counts as on voice. Between 0.5 and 0.8, partial: something's shifted but the bones are still there. Under 0.45, off. The free check works a bit differently, since there's no profile involved, just two pieces compared directly. There, 0.80 and above is on voice, 0.55 to 0.79 is drifting, under 0.55 is off. Each piece needs 120 words minimum to get a reading that means anything.

So what actually moves? Sentence length, usually. Contractions. Formality creeping up or down. Pronoun mix. Signature phrases quietly vanishing. And there's a pattern behind most of it: several writers touching one client, a ghostwriter's own habits leaking through, a new editor with different instincts, or an AI draft flattening the variance and stripping contractions out entirely. That last one is worth sitting with, because it's sneaky. It reads fine sentence by sentence. It just doesn't sound like anyone in particular.

Running the check across a client roster

If you're an agency or a ghostwriter working several accounts, the math gets simple fast. One reference per client. Three recent pieces each. Run the check, watch for anything that lands in the drifting band, because that's your trigger to look closer.

You don't need every client checked every week. You need the ones sitting in that 0.55 to 0.79 range flagged before they slide further. A client sitting at 0.83 one month and 0.79 the next isn't a crisis. A client who was consistently above 0.8 and is now bouncing around 0.6, that's a conversation with whoever's writing for them.

This matters more if you're managing writers across /for-agencies style setups, where the same client might get touched by three different people in a quarter. It matters just as much solo, at /for-ghostwriters scale, where the drift is usually just you, tired, on a Thursday, reaching for shorter sentences because it's faster. Fair enough. Still worth catching.

Fixing drift without a rewrite

Drift doesn't mean starting over. It means editing the named habits back into range and rescoring.

Say the report flags contractions down and formality up. You don't rewrite the piece. You go through and reintroduce contractions where they'd naturally sit, loosen a few sentences that have gone stiff, and check the piece again. That's it. Two passes, maybe three, against the specific attributes the report named rather than a vague sense that "something's off."

This is where having an actual /glossary/voice-drift definition helps, because it stops the conversation from turning into opinion. "This doesn't sound like them" is a feeling. "Contractions dropped 70% and average sentence length rose six words" is a fix list. If you're starting from scratch with a client and want the reference document itself to be sound before you measure anything against it, the /reference/brand-voice-guidelines-template is worth a look. Garbage reference in, garbage score out.

Worth naming directly: this isn't about hunting down AI-written sentences and deleting them on principle. Some AI drafts land fine. The published research backs that up. In one study (SGR-002, 20 September 2026), a measured-voice generation approach scored a mean 0.83 against a study profile built from 2,447 words across three posts, against 0.74 for bare GPT-5 and 0.62 for GPT-5 working from a one-page measured prompt. Against the held-out post itself, those numbers were 0.91, 0.82, and 0.80. So it's not "AI bad." It's "measure it and see."

That same study found the prompt approach asked for roughly 14 contractions per 1,000 words and a formality of 3.7, and GPT-5 delivered 0.5 contractions per 1,000 words with a judged formality of 7.2, even while it obeyed the easy countable rules like no em dashes and more short sentences. Bare GPT-5, meanwhile, wrote 2,309 words against a 1,393 word target with 17.5 em dashes per draft. The measured generation ran 1,302 words with none. And in an earlier study (SGR-001, 28 July 2026), running a banking brief through a profile dropped average sentence length from about 17 words to about 11, doubled the share of sentences under eight words from 24% to 48%, lifted contractions by roughly 70%, and turned ten em dashes into zero.

Watching it monthly

The free check is a snapshot. If you want the drift caught before it shows up in a client call, that's what the monthly monitor is for.

Paid plans watch up to 5 pages a month per profile on Writer and Operator, 15 on Studio, 50 on Agency, and email you when something moves. Studio, at $134, holds 10 client profiles across 3 seats. Agency, at $405, goes unlimited on profiles with 10 seats. If you're already running the roster check by hand every few weeks, this is the same logic on autopilot: same reference, same drifting band, just scheduled instead of manual. You can point it at the /tools/brand-voice-analyzer for the deeper per-piece breakdown when a monthly email flags something worth a proper look.

Questions

What counts as a reference piece?

Any piece that genuinely represents the client's voice at its best: a well-received blog post, a signature email, a page they've approved without heavy edits. It needs at least 120 words for the free check to produce a reliable score. Pick something typical, not their most unusual or most heavily edited piece.

How many recent pieces should I check at once?

Three per client is the practical number for the roster check: enough to catch a pattern rather than a one-off bad day, not so many that the review becomes a chore. If all three land in the drifting band together, that's a stronger signal than one alone.

Does drift always mean something's gone wrong?

Not necessarily. A single piece scoring a bit lower could just be an unusual topic or format. Consistent drift across multiple pieces, especially a downward trend over months, is the pattern worth acting on. One low score is noise. A trend is data.

Can I check pieces without a saved profile?

Yes, that's exactly what the free check does: it compares two pieces directly, reference against recent, with no profile stored anywhere. The scoring bands shift slightly because there's no calibrated profile involved (0.80 and above on voice, 0.55 to 0.79 drifting, under 0.55 off).

What's the difference between the free check and the monitor?

The free check is manual and one-off: paste, compare, done, nothing saved. The monitor runs automatically against a saved profile, checking a set number of pages monthly depending on plan, and emails you when the score moves. One's a spot check. The other's ongoing.