Skim three corporate blogs from this week and you will notice it: the texts could swap places without anyone spotting the difference. Same openings, same transitions, same polite blandness. That is not a coincidence, it is statistics - roughly half of all newly published articles on the web now come out of language models, and without steering they all write in the same register. Which is precisely why brand voice has moved from branding footnote to a hard competitive question.
The good news: the uniformity is an opportunity for everyone who does not sound like it. The data on this is remarkably consistent - across rankings, citations and trust. Let us walk through it.
What is brand voice - and what is it not?
Brand voice is the constant character a brand speaks with: stance, word choice, rhythm, humour - and just as importantly, the things it would never say. It is not the same as tone of voice, which adapts to channel and situation: the same brand sounds different in a LinkedIn post than in a support email, but the character behind it stays recognisable. Voice is the personality, tone is the mood of the day.
And it is not an appendix to the logo. A brand voice shows up in decisions every text makes anew: formal or informal address, active or passive, concrete or abstract, a number with a source or a claim without one. You can recognise brands like Mailchimp or Oatly after two sentences with the sender stripped away - that is the benchmark, not the colour scheme.
How much AI content is actually out there?
The most robust figure comes from Graphite: in an analysis of 55,400 randomly sampled English-language articles from Common Crawl, the share of primarily AI-generated articles stood at around 50 per cent in early 2026 - machines and humans now publish in equal measure. The second finding is the remarkable one: since early 2025 that share has plateaued across five consecutive quarters. The great flood has levelled off.
Why the plateau? Because mass-generated text simply does not perform. Graphite's companion study on search shows that 86 per cent of articles ranking in Google Search are human-written - and so are 82 per cent of the articles that ChatGPT and Perplexity cite as sources. In fairness: AI detectors carry error rates, and individual classifications remain uncertain. But the direction holds across studies and quarters - generic AI text gets published, yet hardly gets found.
What does visible AI content do to trust?
Here too we now have large-scale numbers instead of gut feeling. Klaviyo and Datalily surveyed around 8,000 consumers across eight markets in December 2025: only 7 per cent say visibly AI-generated marketing increases their trust in a brand - 31 per cent say it decreases it. And 91 per cent expect brands to disclose where AI is involved in their marketing.
The real punchline of the study is the contradiction behind it: 77 per cent of the senior marketing decision-makers surveyed plan to shift budget towards generative AI content at the same time. The market is producing more of exactly what consumers demonstrably trust less. For you, that does not mean "no AI" - it means the gap between interchangeable output and a recognisable voice is widening, and that gap is your margin of attention.
How do you define a brand voice that machines understand too?
The classic brand style guide - a 40-page PDF on values and logo clearances - does not help here, because neither your writers in their day-to-day nor a language model in a prompt can work with it. What works is a style guide built as a working document: three to five voice attributes, each with a write-this/not-that example pair, plus a banned-words list ("revolutionary", "game-changer", superlative filler) and firm rules for numbers, sources and address. Short enough to fit into a system prompt - precise enough that two different authors hit the same sound with it.
That turns voice from a feeling into a specification: any text can be checked against the banned list and the example pairs, whether a human or a model wrote it. And the content itself needs what generic AI text cannot supply - first-hand experience, proprietary data, real examples. It is the same reason Google leans on E-E-A-T when judging content: lived experience cannot be generated.
Three levers to start with
Write your voice down on one page. Three attributes, one write-this/not-that pair each, a banned-words list, rules for address and numbers. One page that gets used beats forty pages sitting on a shared drive - and from day one it doubles as your AI prompt.
Run the blind test. Take three recent texts from your brand, strip logo and sender, and show them to someone who knows you. If the brand is not recognised, you do not have a voice, you have content - sharpen it before you publish anything else.
Build the voice into your AI workflow, not against it. Style guide in the prompt, banned list as a pre-publish check, a human as the final judge of stance and facts. That way you get the machine's speed without buying its uniform tone.
In 2026, brand voice is not a garnish for agency slides - it is a visibility and trust lever with data behind it. If you want to know what your brand sounds like, and whether anyone can tell, drop us a line. 🎙️