Every agency that scales AI blog writing for clients hits the same wall around client number six. The posts are grammatically fine, the keyword is in the H1, the word count is right — and a plumbing company in Tulsa, a B2B SaaS, and a boutique law firm all sound like the same mildly enthusiastic consultant. One client notices. Then another. Then you're rewriting from scratch and your margin is gone.
The fix isn't a better model or a cleverer prompt. It's treating voice as a documented asset per client, built once and reused, the same way you'd treat a keyword map or a link prospect list.
Why AI drafts converge on the same voice
Models are trained to produce the statistically likely next sentence. Absent strong constraints, "likely" means the median internet blog post: 18-20 word sentences, three-item lists, transitional phrases like in today's fast-paced world, and a closing paragraph that restates the intro.
When you prompt with "write in a friendly, professional tone," you've given almost no constraint. Friendly and professional describes 80% of the training data. You've asked the model to be average, and it obliges.
The agencies that get this right do two things differently. They measure voice instead of describing it, and they feed examples instead of adjectives.
What "voice" actually breaks down into
Voice feels unquantifiable until you decompose it. Four components cover most of what a reader perceives:
Lexicon
The words this client uses and refuses. A commercial HVAC company says "equipment," never "units." A fintech says "customers," never "users." A trades client says "we'll come out and take a look," never "we'll conduct an assessment." Every client has 15-30 of these preferences, and getting them wrong is the single fastest way to make a post feel foreign.
Rhythm
Average sentence length and its variance. One of our reference clients averages 12.4 words per sentence with heavy fragment use. Another averages 24.1 with multi-clause sentences and semicolons. Those two read nothing alike even on identical topics. Run their existing posts through a readability tool and write the numbers down — it takes four minutes and does more for voice matching than a paragraph of adjectives.
Stance
Who the author is relative to the reader. Peer ("we've all botched a migration"), authority ("here's what the data shows"), or servant ("we'll handle this for you"). Also: does the client make claims or hedge? A law firm hedges on everything. A contractor says "this will fail within two years" without flinching.
Proof furniture
What the client reaches for when making a point. Case studies? Internal stats? Analogies from a specific industry? Customer quotes? A post that argues with a different kind of evidence than the client normally uses reads as inauthentic even when the words are right.
Build a voice profile in about 90 minutes
Do this once per client, at onboarding, and bill it as part of setup.
1. Pull 8-12 source pieces
Not their whole blog. Ask the client: "Which five pieces of writing best represent how you want to sound?" Then add three to five more from their site — ideally pages their sales team sends to prospects, plus any founder LinkedIn posts, which are usually the least sanitized version of their voice. If the client has a brand guide, read it, but weight it below the actual writing. Brand guides describe aspiration; the writing describes reality.
2. Measure before you describe
Record, for the sample set: average sentence length, Flesch reading ease, paragraph length in sentences, first-person plural vs. second-person ratio, contraction usage, and whether they use em dashes, semicolons, or neither. Fifteen minutes of counting gives you a spec no amount of vibes-based prompting will match.
3. Write the do/don't list
Twenty lines maximum, split in two columns. Example from a real roofing client profile:
- Do: Name specific materials and brands. Give price ranges with "depending on." Open with the homeowner's problem, not the company.
- Don't: Use "solutions," "seamless," "in today's market." Never promise timelines. Never use exclamation points.
4. Build a sample bank
Extract three paragraphs from the client's own work: one opening, one explanatory middle section, one closing or CTA. These go into every prompt verbatim as few-shot examples. In testing across our own client set, adding real sample paragraphs reduced first-draft editing time by roughly 40% compared to a prompt with voice described in prose only. The model imitates structure it can see far better than structure it's told about.
The prompt structure that holds up
Order matters. A reliable structure for AI blog writing for clients, in sequence:
- Role and audience — who's writing, who's reading, what the reader knows already.
- The measured spec — sentence length target, reading level, paragraph length, person/POV.
- Do/don't list — the banned-words list is the highest-leverage item here.
- Sample paragraphs — labeled as style reference, not content to reuse.
- The brief — primary keyword, search intent, required sections, internal links, factual constraints (service areas, licensing, guarantees, pricing you're allowed to state).
- Explicit outputs — H2/H3 outline, word range, meta title and description.
Topic selection lives upstream of all this. The brief is only as good as the opportunity behind it, which is why we pull most client topics from query data rather than brainstorming — Search Console is the cheapest topic research tool you already own. Storing the voice profile against the client record rather than in a doc somewhere is the difference between this surviving staff turnover and not; it's one of the reasons we built voice profiles directly into PeakKR's AI blog writer for agencies instead of leaving them in prompt files.
QA: catch drift before the client does
Voice erodes quietly. A junior editor "improves" a fragment into a full sentence, someone swaps in a stock transition, and eleven posts later the client emails: "these don't sound like us anymore."
Run a five-point check on every draft, 90 seconds:
- Ctrl+F the banned words list. Zero hits required.
- Spot-check three sentences against the length target. If the spec says 13 and you're seeing 22, reject and regenerate rather than editing down.
- Read the opening paragraph aloud. If it could open a post for any client in your book, rewrite it.
- Verify every factual claim about the client's business — service area, warranty terms, pricing, credentials. This is where AI drafts genuinely hurt you.
- Check that the proof matches the client's habit. If they always use customer stories and the draft uses a generic industry statistic, swap it.
Quarterly, re-sample: pull three recent published pos

Nick Quirk

