A physician client of mine sent me an email a few weeks ago that started with this line: “Ok, at the risk of you rolling your eyes like I do when my patients come in with their ChatGPT version of medical advice…”
She’d asked ChatGPT to “find hidden problems” with her website. It gave her a clean, confident, seven-point audit. Headings, prioritized findings, the whole package. It looked like the kind of thing you’d pay a consultant a few thousand dollars for. And to her credit, instead of acting on it, she forwarded it to me and asked whether any of it was legit.
Here’s what I want you to notice about that sentence she opened with. She already knew. Not about the audit specifically, but about the pattern. She watches patients walk into her practice every week with printouts and screenshots and confident theories about their own bodies, and she knows that the confidence and the correctness are not the same thing. She knows it because she’s the expert in the room, and she can see the gap between what the model said and what’s actually true. What she couldn’t do, sitting at her desk between patients, was apply that same instinct to her own website. Different domain. No yardstick. So she did the smart thing and asked someone who had one.
What the AI Got Right and Where It Quietly Failed
I went through the audit point by point. Some of it was real. Form error text was leaking into her search snippets, which was a legitimate find and worth fixing. Template placeholders reading “No items found” were showing up on service pages, which made the site feel unfinished. There were small copy errors in prominent places. Those were good catches, and I told her so.
But some of it was wrong in the quiet way that matters. The very first item in the audit flagged her homepage title for containing “PPC” and assumed it was an accidental campaign tag left in by mistake. That’s a reasonable guess if you’re a language model that has seen the acronym PPC used to mean “pay per click” ten million times in marketing contexts. It’s the wrong guess if you know her practice, because in her case PPC stands for Physicians Primary Care, which is a relevant search term for her market and is in the title on purpose. The model couldn’t know that. It pattern-matched on the acronym, decided the most statistically likely explanation, and presented its guess in the same authoritative tone as the things it had right. If she’d acted on it, she would have removed a keyword that was actively helping her.
That’s the whole problem in one bullet point. The audit wasn’t useless. It also wasn’t trustworthy on its own. It was a mix of real findings, reasonable misreads, and confident wrong answers, and there was no marker on any of them telling her which was which. The marker is the part you can’t generate. The marker is the expertise.
The “Anyone Can Do Marketing” Problem Isn’t New > But AI Made It Harder to Spot
I’ve been hearing some version of “anyone can do marketing” for over ten years now. It used to come from clients who’d watched a YouTube video about Facebook ads. Then it came from clients who’d hired a cousin’s kid who was good with computers. Now it comes from clients who’ve spent an afternoon with ChatGPT and produced something that looks, on its surface, like the kind of work they used to pay an agency for. And I want to be clear about something: in all three eras, they were partly right. You can do marketing yourself. People do it all the time. The question was never whether you can. The question is whether what you produced actually works, and whether you’d know if it didn’t.
That second part is the one that’s getting harder, not easier, in the AI era. And it’s the part nobody is talking about honestly.
The Danger Isn’t Bad Advice > It’s Advice That’s Hard to Evaluate
The danger with AI output isn’t that the model gives you bad advice. The danger is that it gives you confident advice, in clean prose, with reasonable-sounding justifications, and you have no internal yardstick to measure it against. If a model tells a doctor how to perform a surgery, the doctor knows which parts to trust and which to interrogate, because she’s done the surgery before. If a model tells a patient how to perform the same surgery, the patient reads the same words and has no idea where the trapdoors are. The output is identical. The reader is not. Most marketing clients are patients, not doctors. They’ve just been handed a tool that talks like an expert, and they have no way to audit the expert.
I had another client run our latest scope of work through what I’m pretty sure was ChatGPT to stress-test it before signing. I have no problem with that. I stress-tested my response with Claude. That’s the world we’re in now, and pretending otherwise is silly. But notice what was actually happening in that exchange: two humans, each using a model as a second set of eyes, each bringing their own expertise to what the model got right and what it got wrong. The models weren’t making the decision. We were. The models were just faster at surfacing things to argue about. That’s the right way to use these tools. It’s also exponentially harder to do well than people think, because it requires you to already know enough to disagree with the model when it’s wrong. And the model is wrong more often than its tone suggests.
Plausibility Is Easy to Create, Expertise Still Matters
So here’s where I’ve landed.
Plausibility used to be expensive. Producing something that looked professional, sounded authoritative, and read like it came from someone who knew what they were doing used to require, you know, someone who knew what they were doing. That’s no longer true. Plausibility is now free, or close to it. Anyone with twenty dollars a month and a lunch break can produce a marketing audit, a content calendar, a brand voice guide, a competitor analysis, a launch plan, all of it, and all of it will look like the real thing.
What hasn’t gotten cheaper is expertise. Expertise is the thing that tells you the audit got the prioritization backwards. Expertise is the thing that knows PPC means Physicians Primary Care in this client’s title and Pay Per Click in someone else’s. Expertise is the thing that reads a proposal and knows which line item is the one that’s actually going to break in month three. Expertise comes from having done the work, watched it succeed, watched it fail, and built up an internal library of pattern matches that no model has access to because the model has never had to live with the consequences of its own advice.
Which Marketing Agencies Will Survive the AI Era
A few things follow from that.
The agencies that survive the next few years are not going to be the ones that pretend AI doesn’t exist, and they’re not going to be the ones that loudly announce they “use AI” as if that were a differentiator. Everyone uses AI now. Saying you use AI in 2026 is like saying you use email. The agencies that survive are going to be the ones that can answer a simple question: what do you bring that the model can’t? And the honest answer for most of us, if we’re being real, has to be some version of “we know what the model got wrong, because we’ve been doing this long enough to tell.”
The clients who should be paying for that expertise are the ones with the most to lose from acting on confident-sounding bad advice. Regulated industries. High-ticket B2B with long buying cycles. Anyone whose mistake-recovery cost is high enough that “let’s just A/B test it” isn’t a real option. Those are the clients for whom an agency relationship is not a luxury but a form of insurance, and those are the clients for whom “I asked ChatGPT” is the most dangerous thing they could possibly say out loud.
Your AI Fluency Has to Be Higher Than Your Clients’
And your own AI fluency, if you’re on the agency side of this, has to be higher than your clients’, not lower. The instinct to hide how much you use these tools is exactly backwards. You should be using them more aggressively than anyone you’re selling to, because that’s the only way to know where they break. I run audits through models. I run my proposals through models. I run my own thinking through models. Not because the model is doing the work, but because I need to know what the model would tell my client if my client asked it the same question, so that I can be ready to explain why the answer is incomplete, or wrong, or right but for the wrong reason.
That’s the job now. Not producing marketing. Producing the insight about marketing that the tools can’t produce on their own.
The Right Way to Use AI for Your Marketing
I want to come back to the doctor, because the way that exchange ended is the reason I’m writing this at all. After I went through her audit and flagged what was real and what wasn’t, I told her something I genuinely meant: this is exactly the right way to use ChatGPT. You asked a good question, you got an answer that looked authoritative, and then you looped in someone with the context to tell you which parts to trust. That last step is the whole game. The people who skip it are the ones who are going to spend the next few years acting on confident wrong answers and wondering why the needle isn’t moving.
If you’ve run an AI audit on your own marketing and you’re not sure whether to trust it, that’s not a weird thing to feel. Send it to someone who’s been doing this longer than the model has existed. Get a second read from a human who has watched these recommendations play out in the wild. That’s not a sales pitch. It’s just the only sane thing to do with a tool this new and this confident.
