Why AI Assistants Get Your Brand Wrong (and How to Fix It)

By Jack Stovell · 2026-08-17 · Guides

Why AI Assistants Get Your Brand Wrong (and How to Fix It)

Ask ChatGPT who you are and there's a decent chance it'll say something technically true and completely flat. "A founder in the SaaS space with a focus on growth marketing." Right, and? That describes half of LinkedIn. You've spent years building a distinct point of view and a voice people actually quote back to you, and the machine flattens all of it into beige corporate mush.

There's an emerging name for fixing this: GEO, or generative engine optimisation. The term is young, but the problem it names is real. Think of it as SEO's younger sibling, aimed at AI assistants like ChatGPT, Perplexity and Claude instead of search results pages. The models usually know who you are. They just describe you like a press release wrote itself in a hurry.

Most advice treats this as company SEO with a different name. Optimise your bio, get backlinks, sprinkle keywords, done. Personal brands don't work like companies, though. You don't have a marketing department pushing consistent messaging across fifty channels. You have scattered posts, a half-finished podcast appearance, a thread from 2021 someone still quotes, and a website bio you wrote once and never touched again. AI models are trying to synthesise a coherent person out of that mess. They're doing about as well as you'd expect.

So let's fix it properly. Here's what's different about optimising a person for AI discovery, and the steps I'd take this month.

Personal brands need entity clarity that companies don't

For a company, this is comparatively simple. A business has one name, one domain, consistent messaging across owned channels. The entity is clear because it's institutional. Nobody confuses Salesforce with a different Salesforce.

You are a much messier entity to resolve. There might be three people with your name. You might post under a nickname on one platform and your full name on another. Your LinkedIn bio says one thing, your speaker page from a 2019 conference says another, and nobody has updated either since.

AI models build their understanding of you through entity resolution: stitching together mentions across the web and working out which ones refer to the same person with the same expertise. If your signals are scattered or contradictory, the model either merges you with the wrong person or defaults to the blandest description it can extract with confidence.

The fix is unglamorous. Consistency. Pick one name format and use it everywhere. Repeat the same core credentials across bios instead of rewriting them from scratch each time. Mention the same two or three areas of expertise again and again, so the model gets redundant confirmation instead of conflicting fragments. Repetition is the signal.

The voice layer matters more than the facts

Here's the part most people miss, and the bit I find genuinely interesting. Getting the facts right (job title, company, expertise) is table stakes. The harder, more valuable problem is getting AI to describe you in your voice.

If you're known for being blunt and slightly irreverent, and ChatGPT calls you "a thoughtful leader passionate about innovation", you've lost something real. That's actively wrong. It's the AI equivalent of a bad actor doing an impression of you with the accent completely off.

This is where stylometry comes in: the fingerprint of how you write. Sentence length, favourite phrases, rhythm, how blunt or hedgy you are, what kind of humour you reach for. I've built a product that measures dozens of these attributes, and the thing that still surprises me is how stable a writer's fingerprint stays across everything they publish. That texture is exactly what a model picks up when it synthesises a picture of you from long-form content, far more than from a sanitised bio paragraph. Whether you intend it or not, everything you publish teaches the model how you sound.

And that has a very practical consequence. When someone asks an assistant "what does [your name] think about X", the answer shapes whether people trust it as genuinely you or dismiss it as algorithmic paraphrase. "AI knows I exist" is one level of brand equity. "AI sounds like me when it talks about me" is a different level entirely, and almost nobody is optimising for it yet.

Build a corpus that teaches the model your voice

So how do you do this in practice? You need volume and consistency of first-person content, spread across a few formats. Three sources earn their keep:

LinkedIn posts. Gold for stylometric signal, because they're frequent, first-person, and usually written by the actual you rather than a comms team's version of you. Post consistently enough that a real corpus exists. Dozens minimum.

Long-form articles. Blog posts, newsletter essays, guest pieces. Your voice unfolds over more words here, which gives richer signal on your rhythm and argument style, and on whether you hedge or state things flatly and move on.

Interviews and podcast transcripts. Genuinely underrated. Spoken language captures your voice in rawer form than anything you've had time to polish. If a podcast you appeared on published a transcript, that's stylometric gold, warts and all.

Starting from nearly zero? Good, in a way. A small corpus in one consistent voice beats a big contradictory one, and you get to build yours clean from day one.

The mistake founders make is treating each of these as a one-off task. Write the bio, do the podcast, tick the box. But a personal-brand corpus behaves like training data, and training data rewards volume plus consistency. The more voice-true content you put out, the more confidently models can reproduce you rather than a category you happen to belong to.

One thing I'll say plainly: don't outsource your voice to a ghostwriter who doesn't sound like you. If the ghostwritten version mismatches how you actually talk, you're training AI models on a slightly fictional you. That fiction becomes the thing people encounter first.

Monitor, test, iterate

None of this is one-and-done. Models update, new content gets indexed, old contradictory bios resurface. You need to check what's actually being said.

Start simple. Every few weeks, ask ChatGPT, Perplexity and Claude who you are, then ask each one what you think about your core topic. Save the answers. Compare them over time.

Look for two kinds of drift. Factual drift: wrong title, outdated company, stale info. Voice drift: a stock description that could apply to anyone. The second is harder to catch because it doesn't feel wrong exactly. It just feels off, like a photo of you slightly out of focus.

When you spot a bland or inaccurate summary, trace it back. Usually your signal sources are thin or contradicting each other, and the cure lives at the source. Update the stale speaker page. Kill the abandoned bio. Publish something new in your actual voice.

Here's your homework for the month, in order:

Audit. Ask the three big assistants about you today and save every answer.
Unify. One name format, one credential set, two or three expertise areas, everywhere your name appears.
Publish. Pick one long-form outlet plus one social channel and commit to a corpus in your real voice.
Retest. Monthly. Fix sources, never symptoms.

Everything above works with nothing fancier than a spreadsheet and a calendar reminder. Full disclosure: that tedium is part of why I built ScriptGrain, which turns your writing into a reusable voice profile. But the method stands on its own. Start manual. Automate when it hurts.

Go and ask ChatGPT about yourself right now. You might not like the answer. That's usually the moment people start taking this seriously.

---

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