Everyone's LLM Sounds the Same, and Your Buyers Have Noticed

By Jack Stovell · 2026-08-03 · Opinion

Everyone's LLM Sounds the Same, and Your Buyers Have Noticed

Read three SaaS blog posts back to back. Go on, actually do it. I'll wait.

Notice anything? The same tidy three-part structure. The same "let's dive in" energy, the same oddly upbeat ending about unlocking potential. You could swap the logos at the top and nobody would spot the difference for a solid paragraph or two.

That's convergence. It happens to most companies that plug a generic prompt into an LLM and call it a content strategy, and I say that as someone who stares at AI-generated drafts for a living. I build ScriptGrain, a tool that measures how writing actually sounds, so I've read more machine-written marketing copy than any human reasonably should. The beige middle of AI output is very beige indeed.

The convergence problem

Here's the uncomfortable bit: the content isn't bad. It's usually well structured, grammatically clean, technically "on brand" if your brand guidelines only ever specified a colour palette and a logo lockup. The problem is that it's fine in exactly the same way everyone else's content is fine.

Ask an LLM to write "in a confident, friendly, professional tone for a B2B SaaS audience" and you'll get roughly the same output whether you sell payroll software, project management, or niche compliance tooling. Sentence lengths flatten out. Transitions become interchangeable: "let's break this down", "here's the thing", "at the end of the day". Whatever personality your brand once had gets smoothed into professional beige.

I can't prove every buyer consciously clocks this. What I can tell you is that when I read three converged blogs in a row, the sameness is unmissable, and your buyers read far more of your category's content than I do. They skim your blog, then a competitor's, then another, and something in the back of their brain quietly files all three under "same company, probably". Trust takes a knock. Memorability takes a bigger one.

There's also a second audience now: the AI assistants themselves. When an assistant summarises your category for a buyer doing research, it's drawing on a sea of content that increasingly reads like one author wrote all of it. If your voice is statistically identical to your competitors', you're easy to blur into a bland list item.

What "voice-trained" actually means

So the fix is obvious, right? Just tell the model to sound like you. Add a paragraph to the prompt: witty, irreverent, short punchy sentences, no corporate jargon. Job done.

Except it isn't. That's prompt engineering, and prompt engineering is instructional: you're describing your voice to a model the way you'd brief a freelancer who's never read a word you've written. The model complies with the description, and descriptions are lossy. "Witty" means something different on every run. "Short punchy sentences" drifts the moment you tweak the prompt for a different content type.

Voice training works from the other direction. It's stylometric, meaning it's calibrated against measurable patterns in your real authored writing: sentence length variance, punctuation habits, how often you use contractions, whether you open paragraphs with "I" or "The" or "So", how dense your commas run, whether you reach for metaphor or avoid it. Dozens of small, mostly invisible fingerprints that make writing feel like a specific person produced it.

None of that fits in a prompt. You can't tell a model "use slightly more parentheticals than average" and expect consistency across fifty pieces. You have to measure a body of real writing, work out what's genuinely distinctive about it, and build those measurements into the generation process itself. It's the difference between describing your handwriting down the phone and training a system on actual samples of it.

Why voice is turning into a moat

This is where it stops being an aesthetic nitpick.

Voice used to be a slide in the brand deck. Something marketing discussed at kickoff and quietly abandoned once deadlines hit. Then the same tools commodified everything around it. Think about what's actually scarce in content now. Information isn't; assistants summarise it for free. Speed isn't either, because everyone can produce volume. What's scarce is distinctiveness: the quality that makes a reader go "oh, this is that company" before they've seen a logo.

That scarcity compounds because of where discovery increasingly happens. Buyers ask an assistant to compare tools in your category, and what comes back is shaped by how consistently and distinctively you show up across the web. Generic content is easy to compress into a safe, forgettable list item. A genuinely distinct voice gives the model something quotable, and gives the human on the other end something to remember.

Most teams thinking about AI-driven search focus on structured data and citations. Fair enough. That's the plumbing. Voice is what makes you memorable inside the answer itself, and when an AI summary is a buyer's first touchpoint with your category, memorable is most of the fight.

What the fix looks like in practice

None of this means binning your AI tools and hiring five copywriters. Here's the operational version, in four steps.

Build a measured voice profile. Take your best authored content, the pieces that genuinely sound like you, and extract a stylometric baseline from them. Measured attributes, with numbers attached, rather than a vibes-based style guide someone wrote in an afternoon.

Generate against the profile. The profile becomes the calibration target every new piece is checked against, so "sounds like us" stops being a matter of opinion in a review thread.

Check drift before publishing. Is the sentence length variance still yours? Are contractions appearing at the rate you actually use them? Catch the beige creep per piece, before it compounds across fifty posts.

Re-measure on a schedule. Voices evolve and prompts get tweaked. A quarterly re-baseline keeps the target honest.

You can run a rough manual version of this today. Pick five pieces you're proud of, note your average sentence length and its spread, list your habitual openers and transitions, then hold every AI draft up against that list before it ships. It's tedious. It works.

If you'd rather have it measured properly, that's the exact job ScriptGrain was built for: it extracts a detailed stylometric profile from your writing samples and generates against it, with the drift checks built in. Obviously I'm biased, I built the thing. So don't take my word for it: paste a writing sample into the voice preview and see what your voice actually measures as. It takes about a minute, and it's a strange little mirror the first time you look into it.

Because the alternative, the thing most teams are doing right now without realising it, is spending real budget on AI content that makes them sound exactly like everyone they're trying to beat.

---

By Jack Stovell, founder of ScriptGrain. I build the product this blog writes about; everything here comes from shipping it. See how ScriptGrain works.