Content that sounds like you

If you publish AI-written content without a voice system, you are publishing roughly the same article as everyone else in your category. The model trained on the whole internet and optimised for the middle, so every team that uses it off the shelf draws from the same centre.

That is not a hunch. When researchers had people write with a feedback-tuned language model, the assistance measurably reduced the diversity of what they produced, and the homogenising text came from the model, not the writers. A separate study found that giving writers AI story ideas lifted individual creativity but made their stories about 11% more similar to each other.

Individually better, collectively the same.

That sameness is the slop. Generic to the point of invisibility.

And the web is filling up with it. By early 2026, AI-generated articles had drawn level with human-written ones, at roughly half of everything newly published online. A separate analysis of the live web by researchers at Stanford and the Internet Archive put AI-generated or AI-assisted pages at about a third of new sites by mid-2025, up from near zero before ChatGPT. Ahrefs, scanning 900,000 new pages in 2025, found around 74% contained some AI-written text.

Most brands using AI right now are not making a bad product. They are making an average one, which in content marketing is often worse.

Why AI content sounds the same

Most teams start by hunting for AI tells. Strip the em dashes. Kill the hollow adjectives. Delete the constructions that inflate a simple point into two clauses. That work is real. It is also the prequel.

Content that does not read like AI still needs to read like something. If the output is clean but could have come from any brand in your sector, nothing material has changed. You have removed the model's fingerprints and replaced them with nothing.

The useful test is sharper: would a reader know this came from your brand if the logo disappeared?

What is an AI content system? An AI content system is the set of files, prompts, proof points and review gates that tell a model how your brand thinks before it writes. It usually includes a brand voice pack, audience notes, structural patterns, sentence-level preferences, a claims register, source material, banned phrases, examples of good and bad writing, and a QA scorecard. The point is not to automate judgement away. The point is to make judgement repeatable. A model without this context defaults to average language from its training data. A model with this context has something specific to write from, and a reviewer has a clear standard to judge the draft against.

Build the voice before the prompt

A clever prompt can get you a better draft. It cannot give you a voice. For that you need two things working together.

Start with a brand voice pack. A prompt tells the model what to do this time. A voice pack tells it how your brand makes decisions every time. That distinction matters because AI does not struggle with consistency. It is too consistent already. The work is to add the useful irregularities: the sentence shapes, examples, refusals and judgement calls that make the writing belong to a particular brand.

A useful voice pack should include:

• Voice boundaries: the tensions that define the tone. Plain but not flat. Expert but not academic. Direct but not abrasive.
• Structural patterns: how a piece usually opens, how quickly it reaches the point, how it moves from example to argument, and what a good ending does.
• Sentence preferences: the punctuation, rhythm, concrete nouns, verbs and paragraph shapes that feel right for the brand.
• Signature moves: the things the brand does especially well, such as starting with a practical problem, naming the hidden system, then giving the reader a usable framework.
• Anti-patterns: the tells to cut on sight, including parallel headings, fake profundity, generic transitions, tidy summary endings, and "not X, but Y" scaffolding.
• Positive and negative examples: one paragraph that sounds right, one paragraph that misses, and a note explaining why.
• A claims register: the things the brand can say, cannot say, and needs evidence for.

Add a scorecard. This grades each draft against Google's people-first, helpful-content guidance and names the AI tells if they appear. It turns "does this feel right?" into something specific enough to act on and consistent enough to run across a team.

The voice pack gives the AI something to write from. The scorecard tells you whether it worked.

The fastest way to build the first version is not to ask the team to describe the brand from memory. Most people are better editors than self-analysts. Give the model three to five examples of writing that feel close, then make it ask questions one at a time. Which opening sounds more like us? Which paragraph is too polished? Which line would we never publish? Which example has the right amount of friction?

Those reactions are the raw material. Turn them into rules only after they have been tested against real drafts.

Make the workflow repeatable

Both tools only hold up inside a repeatable process. The system has seven layers. Brand DNA, examples, research, draft, stop-slop pass, humaniser pass and QA gate. In that order, every time.

The research layer is the one most teams skip, and it is the one that earns the right to make a claim. Most AI content fails because it draws only on what the model already knows. The same sources, the same conclusions.

Feeding in primary documents, your own data and real interviews is what gives a draft something to say that other content cannot replicate. Every statistic in this article came from that layer, not from the model's memory.

After publishing, edits and performance data feed back into the voice pack and the research. Every repeated correction becomes a new guardrail. If the model keeps writing symmetrical headings, add that to the anti-pattern list. If endings keep collapsing into summary, add a rule that endings must extend the idea or give the reader a next decision. If a technical section keeps getting too abstract, add a positive example with concrete nouns, named tools and real constraints.

Without that loop the voice pack goes stale and the output drifts back to the average. The model-level failure mode already has a name. Train AI on enough AI output and the distribution collapses toward a narrow mean. The same gravity pulls on your content if nothing original feeds it.

How do you stop AI content sounding generic? Stop treating the prompt as the system. Start by giving the model real inputs: your voice rules, source documents, proof points, audience context, structural preferences and examples of copy you would actually publish. Then run the draft through a stop-slop pass, a humaniser pass and a QA gate before it reaches the CMS. The draft should be checked for unsupported claims, generic language, borrowed tone, missing proof, weak usefulness and repeated AI patterns. After publishing, feed edits and performance data back into the voice pack. That loop is what keeps the output from drifting back to the middle.

A before-and-after example

Take a Jersey tourism page, the kind that reads fine and says almost nothing. The intro runs something like this:

Our island may be just nine by five miles, but it is bursting with adventure at every turn. Whether you are uncovering our rich history, exploring breathtaking landscapes or treating yourself to a little retail therapy, there is something for every kind of escape. Your dream holiday is just a doorstep away.

Smooth, and completely interchangeable. It promises everything to everyone and leaves a visitor with not one fact to plan around, so it could be any island in any year.

Now the same intro, written with a brand voice pack, a visitor persona and a claims register feeding it:

Jersey is nine miles by five. In one day you can tour Mont Orgueil Castle, follow the coast toward La Corbiere, eat in St Helier, or spend the afternoon at St Ouen's Bay. Decide the kind of day you want first, then use the guide below to plan the rest.

Same length. The second names real places and gives the reader an actual decision to make, where the first gave only a mood. Neither is factually wrong. Only one of them helps you plan a trip.

Where SEO fits

Google could not be clearer that generic, scaled output is a liability. Its people-first guidance asks for content "created primarily for people, and not to manipulate search engine rankings", and for first-hand expertise, the kind "that comes from having actually used a product or service, or visiting a place". In March 2024 it made "scaled content abuse", generating many pages "without adding value", a spam policy, naming generative AI tools by name.

The flip side is what teams get wrong about AI search. There is no special lever for it. Google's own guidance on AI features states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that you "don't need to create new machine readable files, AI text files, or markup".

The way to be cited by AI is the same as the way to rank. Say something a person with real experience would say, that no one else can.

Sounding like you and ranking for the right queries are not separate jobs. The brand work carries the search work.

Get the toolkit

At Ikigai, we built a free version of the system that runs in ChatGPT, Claude or Copilot. It includes a brand voice pack template, a spot-the-slop cheat sheet, copy-paste prompts and the skills behind each layer.

If you are spending time cleaning up AI content that still reads like everyone else's, the brand voice pack is where to start. Everything else builds from it.

Download the free AI Content System toolkit