Brand Voice for AI Content: How to Give a Tool Your Site's Tone (With Examples)
"Write it in our brand voice" is the instruction most people give an AI tool and the one it can do least with, because "our brand voice" isn't an instruction. The model needs to know what it would do differently.
The good news is that voice, for a buyer guide, comes down to five measurable things, and a brief that sets all five takes ten minutes to write and changes every page the tool produces from then on.
This post covers what a voice setting can change and what it can't, the five dimensions that account for most of what readers hear as "tone", how to write the brief, the same product section written in three voices so the difference is visible, and the mistakes that make a voice brief useless.
It's the setup half of the editing post: the voice gets the first draft close, the edit adds what no voice can.
The short answer
A voice setting controls the measurable surface of writing: sentence length, person, verdict bluntness, hedging and humour. Set each with one concrete sentence, in one paragraph, once per site. It can't supply opinions or experience; those are the editing line. Give each site its own brief, flipping at least one dimension between sites, so they don't read as the same tool.
What a voice setting can and can't change
| Can change | Can't change |
|---|---|
| Sentence length and rhythm | Your opinion of the products |
| Whether the reader is "you" | Your experience with them |
| How bluntly verdicts are stated | Facts not in the data |
| How much hedging ("might", "could") survives | Humour that's actually funny |
| Vocabulary limits and banned words | The structure of a buyer guide (and shouldn't) |
| Formality: contractions, register | What makes your site different from the next one |
The five dimensions
1. Sentence length
The single biggest lever. "Sentences under 20 words, one idea each" produces a clipped, confident page; "mix short sentences with longer ones that explain" produces a more conversational one. Say which, with a number.
2. Person
Does the guide address the reader as "you"? Does the writer appear as "I" or "we", or not at all? "Second person throughout, no first person" and "we, as a team that tested" are different sites. For a site working from listing data, no first person is safer; it avoids implying hands-on experience the site doesn't have.
3. Verdict bluntness
"Say which product to buy and the one reason, in a single sentence, before anything else" versus "present the trade-offs and let the reader decide". The first converts better and reads as more confident; the second suits a site whose readers are experts. Most affiliate sites want the first and get the second by default.
4. Hedging
Models hedge: "might", "could potentially", "may be a good option for some users". A brief that bans those words, and says "state facts as facts and opinions as opinions, without softening either", removes most of what makes AI text sound like AI text.
5. Humour and warmth
The riskiest dimension, because model humour is weak. The safe settings are "none; dry and factual" or "one light aside per section at most, never in the verdict". "Be witty" produces puns about product names. If your site is funny, the humour is yours to add in the edit.
Writing the brief
One paragraph, one sentence per dimension, each with something the model can act on. A working example:
"Sentences under 20 words, one idea each. Address the reader as 'you'; no 'I' or 'we'. Name the product to buy and the one reason in the first sentence of the verdict and of each section.
No hedging words: no 'might', 'could', 'may', 'potentially', 'some users'. Dry and factual; no jokes. Contractions are fine. Specs over adjectives: never 'durable' or 'versatile', always the number or the fact."
That's 70 words and it changes everything the tool writes for the site. Compare the version people usually write, "friendly, authoritative, engaging and professional", which sets nothing.
One section, three voices
The same product, the same facts, three briefs.
Understated and factual (short sentences, no "you", blunt verdict, no hedging): "The X is the pick for unlit roads. 1,200 lumens, six hours on high, USB-C. It weighs 180 g, 40 g more than the Y. The beam is the reason to accept the weight."
Direct second person (short sentences, "you", blunt verdict, no hedging): "If you ride unlit roads, buy the X. You get 1,200 lumens and six hours on high, which covers a week of commutes on one USB-C charge. It's 40 g heavier than the Y; you'll notice that on the shelf and not on the bike."
Scenario-led (mixed sentence length, "you", trade-offs first, no hedging): "On a country lane at 11 pm, the difference between the X and the Y is whether you see the pothole at 30 metres or at 15. The X's 1,200 lumens and six-hour runtime are what put it first here; the 40 g it adds over the Y is the price of that beam."
Same facts, same verdict, three sites. None of them is better; the point is that a reader of each site hears one voice across every guide on it.
Mistakes that make a brief useless
- Adjectives instead of instructions. "Authoritative" tells the model nothing; "state the verdict in one sentence, first" does.
- Contradictions. "Warm and friendly" plus "dry and factual". Pick one.
- Asking for a famous writer's style. The model produces a parody, and it's not your voice anyway.
- Setting the voice per guide. It's per site, or the site has fifty voices.
- Expecting the brief to add opinions. It can't. That's the experience only you supply, in the edit.
Doing it with AffSERP
AffSERP's X Best Review tool takes a brand voice brief per site and applies it to every guide written for that site, alongside a house voice that's assigned per site so that two sites without a custom brief still read differently.
The banned-word list and the specs-over-adjectives rule are built in whatever the brief says, so the brief only has to set the five dimensions above. The scale post covers how that voice combines with the rotations that keep fifty guides on one site from repeating each other.
FAQ
What is a brand voice for AI content?
A short set of instructions that fixes how the model writes for your site: sentence length, whether it addresses the reader as 'you', how bluntly it gives verdicts, how much it hedges, and whether it's allowed any humour. Set once per site, it makes fifty guides read as one site's work rather than fifty prompts' output.
Can AI really match my writing style?
It can match the measurable parts: rhythm, person, directness, vocabulary limits. It can't supply your opinions, your experience or your jokes, because it doesn't have them. A voice setting gets the text 80% of the way; the editing pass that adds judgement does the rest.
How long should a brand voice brief be?
One paragraph, five to eight sentences. Each sentence sets one dimension with a concrete instruction ('sentences under 20 words', 'say which one to buy and why in one line'). Longer briefs full of adjectives ('friendly, authoritative, engaging') give the model nothing to act on.
Should every site I run have a different voice?
Yes, if they're on different topics or for different readers, and especially if a reader might visit both. Two sites with the same voice and the same tool read as the same site. Pick one of the five dimensions to flip between them and the difference is audible.
Keep reading
