Why AI Can’t Tell You What Premium Buyers Want

I’ve been buying from the same coffee business for most of my adult life. Machine, grinder, tamper, cups, and fresh beans every fortnight for years. If you listened to episode 338, the Klaviyo flows episode, you’ve heard that story. What I didn’t tell you is the embarrassing bit. In all that time, I’ve never commented on one of their posts. I’ve never shared them. I’ve never even left them a review.

If that business judged its customers by what it could see on social media, its most loyal customer of several decades wouldn’t exist.

I’m not unusual. I’m normal now. And that’s what this episode is about. The people buying from you have gone quiet. The signals you used to read about them are thinning out and filling up with noise. The shortcut everyone is reaching for to fill the gap is asking AI to figure out who your customer is, then communicate with them on your behalf. That shortcut is built from the same noise.

I use AI every day. I’m not here to tell you to stop. I believe AI is one of the most useful tools a premium brand has ever had for understanding its customers. But I believe something else just as strongly. AI cannot have a human experience. It has never stood in a bathroom at 7am wondering whether a $90 serum is going to do anything. It has never been burnt by a brand that overpromised. It can read about those moments. It can’t have them.

So this episode is about where the line sits. What to hand to AI, what never to hand to it, and how to use it on your customer research without letting it make things up. And before we finish, I’ll give you three quick tests you can run on your own brand this week. Fair warning: at least one of them is likely to be a bit confronting.

Let’s start with what’s happened to social media, because it’s changed more than most of us have noticed.

Last year, during a US court case, Meta presented its own figures on how people use Facebook and Instagram. On Facebook, the share of time people spent viewing content from their actual friends fell from 22 per cent to 17 per cent in two years. On Instagram, it fell from 11 per cent to 7. Seven per cent. On Instagram, more than nine minutes in every ten are spent watching something other than what the people you know have posted.

It’s worth remembering why Meta showed those numbers. It was arguing in court that it isn’t really a social network any more. So Meta had every reason to make that number look small. Even so, it’s a remarkable thing for a company built on friends and family to say about itself.

There’s a name for what’s happening to the rest of us. A writer at The New Yorker called it “posting zero”: ordinary people quietly stopping sharing their lives in public. They still open the apps. They still scroll. They just don’t post, and they don’t comment much either. When they find something they love, they send it to a friend in a private message or a group chat. Mark Zuckerberg said as much in that same court case. People discover things on the platforms, then share them privately.

That private sharing is where a huge amount of word of mouth now lives. Your customer sends your product page to their sister. Their sister buys. Neither of them ever likes, comments or tags. As far as your analytics are concerned, the sister turned up out of nowhere, and the customer who sent her has never engaged with you in their life.

So where does that leave a brand trying to understand its customers? Looking at a comment section that’s getting quieter every year.

Some of you have read that people are leaving social media altogether. Globally, there’s some truth to that. An analysis of 250,000 adults across more than 50 countries, done for the Financial Times, found time on social media peaked in 2022 and had dropped by almost 10 per cent by the end of 2024. The biggest falls were among teenagers and people in their twenties.

But I’m not going to tell you Australians have left, because we haven’t. The latest Digital 2026 report for Australia found we spend an average of 41 hours a week online, more than a full-time job. Nearly half of that is on social media. Your customers are still there. They’ve just stopped talking.

And they’ve stopped believing much of what they see. In the US, Pew Research found that only 37 per cent of adults have at least some trust in the information they get from social media. Among people aged 50 to 64, it’s a third. Over 65, it’s under a quarter. If your customers are women in their forties and fifties with money to spend on something premium, they’re among the most sceptical people in the feed.

I call this trust fatigue, and I think it’s doing more damage to ad results than creative fatigue ever has. Creative fatigue is when people get tired of seeing your ad. Trust fatigue is when people get tired of believing anyone’s ads. You can fix creative fatigue with a new video. Trust fatigue doesn’t care how fresh your video is.

That scepticism is earned. Think about what’s filled the space your customers left. Brand content, in enormous volumes. More and more of it made by AI. And bots. In the audits we run on premium brands, we routinely find anywhere from around a fifth to more than half of a brand’s reported website sessions aren’t human. That’s real data founders make real decisions on.

I’ll give you an example of how sneaky this gets. We recently noticed something odd in a client’s email account. Click rates dropped suddenly and sharply. On its own, that looks like a content problem. You’d assume the emails had gone stale, rewrite the subject lines and move on. But there was a second oddity. A bunch of profiles had entered the abandoned checkout flow without having done anything else. They’d never opened an email, never browsed a product, never clicked anything. Real people don’t behave like that. Bots had been getting into the email list, and the only entry point we could find was a subscribe box at checkout that was ticked by default. We caught it early, before it could damage the brand’s sending reputation.

Now imagine we’d handed that data to AI and asked why click rates had fallen. It would have told us, confidently and in beautifully organised dot points, that the emails had become repetitive and the subject lines needed refreshing. It would have been doing its job perfectly, on data that was poisoned. AI analyses whatever you give it. It can’t tell you the data is lying.

That’s the first problem. The second worries me more.

When the customer goes quiet, there’s a very tempting shortcut. Open a chatbot and type, “Describe the ideal customer for a premium natural skincare brand.” Or generate a hundred synthetic customer personas and survey them instead of real people. There are businesses selling exactly this, and it’s quick, cheap and confident.

The research on it is worth knowing. A team of US researchers tested four AI models against 3,200 real people across 16 different identity groups. The AI consistently did two things. It flattened people, collapsing the variety inside a group into one tidy average. And it misportrayed them, describing groups more the way outsiders talk about them than the way they talk about themselves.

Then there’s price, which matters enormously for a premium brand. A UK insight firm, Strat7, compared survey results from synthetic AI respondents against real people this year. When asked what they’d be willing to pay, the AI respondents named prices around 16 per cent higher than the humans did. Ask an AI what your customer will pay and it flatters you. Then you price, forecast and plan your ad spend on a number real people won’t actually pay.

To be fair, it isn’t all bad news for the machines. A study from the London School of Economics and New York University found an AI-generated audience matched human polling results quite closely on broad questions. I believe that. AI is very good at the average. It has read more about the average person than any of us ever will.

But you don’t run an average brand. Premium brands live in the specifics. The particular disappointment that made someone try you. The exact phrase they use to describe what they want. The reason they nearly didn’t buy. Those things aren’t in the average, and they aren’t in a model trained on the whole internet. They’re in your customers’ heads.

There’s a lovely study on this from the University of Maryland and Google DeepMind. They compared tens of thousands of stories written by people with stories written by five different AI models. One of the clearest differences: the humans referred to specific brands, places and texts nearly twice as often as the AI did. People write in specifics because they’ve lived specific lives. AI writes in generalities because it’s averaging across all of them.

If you want to see what happens when AI fills in the gaps on its own, I have a story for you.

Meta would very much like you to use its AI creative enhancements, the settings that take your ad and generate a whole range of variations of it automatically. For one client whose ads we managed, we switched on a couple of them. Not the whole buffet. Two. Meta’s AI rejected the ads. Because we’re an agency, we could ask for a review by an actual person. They do still exist. And Meta told us, more than once, that the problem was the AI-generated variants themselves. They were breaking Meta’s own rules. The machine failed the machine’s test.

Nobody involved in making those variants had ever met the customer. Apparently they hadn’t read the rulebook either.

We’re not the only ones. Agencies here and overseas have been switching these tools off after finding AI-generated variations running that nobody approved. One overseas footwear brand ended up issuing refunds after AI-altered images misrepresented its products.

I’m not anti-Meta. We recommend Meta ads to most of our clients, and we run them for our own brand. But attention on Meta costs more every year. In its most recent results, Meta reported that its average price per ad was up 12 per cent on the year before. And on today’s Meta, your creative does most of the targeting. The words and images in your ad are how the algorithm decides who should see it. So if your hooks come from an AI’s guess about your customer, you get generic ads, served into a feed already overflowing with generic content, to people suffering trust fatigue, at a higher price every year. What earns attention in that feed is specificity. And specificity comes from real people.

So let me show you what real people sound like.

A couple of years ago, we used a PR platform to help get exposure for Indagare, my natural skincare brand. Their team kept encouraging us to use big-promise headlines. Things like “fewer wrinkles in seven days”. That’s the kind of promise we couldn’t stand behind anyway, and it’s the kind of claim that gets skincare brands into trouble with regulators. But that wasn’t the most interesting part.

When we ran our last customer survey, we asked a simple question: “What’s the single most important thing a skincare brand could do to earn your long-term loyalty?” Here’s one answer, word for word.

“Real results, natural ingredients, no hype about doing this and this. If I’m going to spend money I want the right products with good results. Sick of promises that don’t deliver eg dark circles under the eyes and skin discolouration.”

No AI wrote that sentence. Look at what’s in it. “No hype about doing this and this” is exactly how a real person talks when they’re tired of being sold to. The specific disappointments. The link between spending money and wanting something that works. Our own customers told us, in their own words, that the headline we were being pushed towards was the thing most likely to lose their trust.

You won’t find that by reading comments, because that customer probably never comments. You won’t find it by asking a chatbot, because a chatbot would have suggested “fewer wrinkles in seven days”. You find it by asking real people the right question and listening to the answer.

This is where AI becomes extraordinary.

One survey answer is a story. Four hundred survey answers is a research project. Add your reviews, your customer service emails, the search terms bringing people to your website, and the performance data from your email and ads, and it becomes a research project most founders never get to, because it takes weeks.

For Indagare, I used AI to combine all of it. Our survey responses, our customer reviews, the organic search queries from Google Search Console, our Klaviyo data, and the performance data from our organic and paid channels: which hooks stopped the scroll, which content got clicked, which pages held people’s attention longest, and which of those paths brought in our best customers. And then to find the specific patterns and common themes leading to the best and worst results. Doing it alone, I’d have been at it for weeks. With AI, it took a fraction of the time.

What came out was a concrete picture of our best-performing path from first glance to loyal customer. The hooks that earn the best attention. An ad structure that looks very little like the standard eCommerce setup. The landing pages people actually read. The email opt-in offer that brings in the customers with the highest lifetime value, which, for the record, isn’t the one most brands use. And the subject lines and content that get our flows opened and converting.

We’re still implementing it. That’s true of every brand, and we’re scaling at the same time. But in the period I compared, Indagare’s MER, our total revenue divided by total marketing spend, more than tripled on the same time the year before. MER is a blended number, so it isn’t flattered by any change in how Meta attributes sales. It’s the money in against the money out.

AI didn’t have a single insight about our customers. Our customers had the insights. AI found them, sorted them and showed me where they agreed with each other. That’s the job I want AI doing.

So how do you get the extraordinary version and not the one that makes things up? These are the guardrails I use.

Start with real people, every time. Synthetic respondents can help you test whether a survey question makes sense before you send it. They can’t be the answer. If the voices going into the analysis aren’t real customers, nothing that comes out is real either.

Clean the data before AI sees it. Strip out bots, test orders, staff purchases and anything that doesn’t behave like a human. The email story is why. AI will happily build a confident conclusion on rubbish.

Make it quote. Every theme the AI gives you should come with the customers’ actual words beside it, and some way to trace each quote back to the person who said it. If it can’t show you the quote, the theme doesn’t exist yet.

Make it count. “Many customers said” isn’t a finding. Three out of 400 is a very different story from 150 out of 400, and AI loves turning three people into “many”.

Ask it what doesn’t fit. Get it to show you the responses that contradict the main themes, the outliers and the minority views. The most valuable insight in a survey is often the one only a handful of people mentioned, and an AI looking for the average will smooth it away.

Check its work. Read a sample of the raw responses yourself before you read the AI’s summary, then spot-check its quotes against the original file. When it’s wrong, it’s usually wrong confidently, which is exactly why you check.

And keep your customers’ words. Don’t let AI polish them. “No hype about doing this and this” is clumsy. It’s also perfect, because it’s how your next customer talks too. The moment AI tidies it into “consumers value authenticity and transparency”, you’ve lost the thing that made it useful.

Now for those three tests. None of them takes more than an afternoon, and none of them needs a big survey.

The first is the quiet customer test. Export your top 50 customers by total spend from Shopify. Then go looking for them on your social channels. How many follow you? How many have ever liked a post, left a comment or sent you a message? Most founders who try this are surprised by how few of their most valuable customers they can find. If the people who’ve spent the most with you are invisible on social, then the people you can see, and the metrics you’ve been reading, aren’t telling you much about your best customers at all.

The second is to ask the question you can’t see the answer to. Add one question to your post-purchase survey or your order confirmation page: “How did you first hear about us?” Make it a free-text box, not a drop-down list. A drop-down only lets people choose the options you thought of. A free-text box lets them tell you “my sister sent me the link” or “a friend wouldn’t stop talking about it”. Give it a month and count how many answers point to a friend, a group chat or a recommendation. That’s your private word of mouth, made visible. It’s often a much bigger slice of your business than your attribution reports suggest, and once you can see it, you can start working out how to earn more of it.

The third is the chatbot test, and it’s the confronting one. Open your AI tool of choice. Ask it to describe the ideal customer for a brand like yours: who they are, why they buy, what they worry about, what they’d happily pay. Then ask it to write three ad hooks for that customer. 

Now put that next to twenty real reviews or survey answers from your actual customers, and go through both with a highlighter. Highlight every phrase that only a real person could have written. The specific disappointment. The odd turn of phrase. The thing nobody else in your category says. Then look at the AI’s version and see how much of it would work just as well for your three nearest competitors. Check the price it guessed against what your customers actually pay, too.

If you want to go one step further, give AI those twenty real answers without any of the guardrails I just described, and ask it for the main themes. Then check every quote it gives you against the originals. Count how many it got exactly right, how many it tidied up, and whether it invented any. That tells you very quickly how much to trust it when you’re not checking.

Whatever you find, the next question is what to do with it, because insights sitting in a spreadsheet don’t sell anything.

Your customers’ words belong in your ad hooks and in the first line of every email. The phrases real people use are the ones that stop the next person like them mid-scroll.

Their hesitations belong on your product pages and in your welcome emails. Every “I wasn’t sure it would work for me” is an objection you can answer before the next customer has to ask.

The questions they had before they bought belong in your content. That’s what people type into Google and ask ChatGPT, so answering them is how new customers find you when they’re searching for exactly what you sell.

What earns their loyalty belongs in your offer. If your customers say they’re sick of hype, your welcome offer and your launch campaigns shouldn’t sound like hype.

And whenever you do use AI to write, including Meta’s own tools, give it your customers’ real words as the brief. AI writes far better when it’s working from what real people said than when it’s left to guess.

What all of that looks like for your brand depends entirely on your brand. Which questions to ask, which sources to combine, what your customers’ answers mean for your ads, your pages and your emails, and what to fix first. That’s what a Brand Growth Strategy Session is for. We’ll look at what you already know about your customers, where the gaps are, and what your real customers might be trying to tell you that you haven’t heard yet. If you run any of those three tests and the results make you wince, bring them with you. You can book a session at the link in the show notes.

AI reads, sorts, counts and connects at a scale no founder could manage alone. People supply the experience. When you let AI do the people’s part, you get an average customer who doesn’t exist, writing headlines your real customers don’t trust.

And if there’s a brand you’ve loved quietly for years and never told, maybe go and leave them a review. I’m going to go and write mine.