The short version: the best use of AI for a small business isn't generating blog posts or ad copy, it's feeding it your own customer emails, reviews and support chats so it can tell you the exact words your customers already use. Do that before you write a single headline and your conversion rate will move more than any prompt trick will. This is a two-hour job you can do this week with tools you already pay for.
Everyone's using AI backwards
Here's what nearly every small business owner does with AI right now: they open ChatGPT, type "write me a Facebook ad for my accounting firm" and get back something that sounds like every other accounting firm's Facebook ad. Generic, competent, forgettable. Then they wonder why it doesn't convert.
I did this too, for about eight months, back when I was rebuilding my own client base after five years of near-invisibility. I'd ask AI to write the thing, tweak the tone a bit, post it, and get the same flat 1 to 2% engagement I'd get writing it myself. The problem was never the prompt. The problem was that I was asking a machine to guess what my audience wanted instead of showing it what my audience had already told me.
The shift that changed things for me, and for three clients I've worked with since, was using AI the other way round: not to produce words, but to read words I already had and hand me back the patterns.
The client story: a personal training studio and 340 support messages
A client of mine runs a small personal training studio in Kent, two locations, about 400 active members. She'd been writing her own marketing copy for years around the theme of "get fit, feel great, build strength." Standard stuff, and it was underperforming. Her landing page for new-client sign-ups was converting at 2.1%.
We pulled every customer support message, cancellation reason and Facebook comment from the previous 18 months, about 340 pieces of raw text, and dropped them into a document. Then I fed that into Claude with a simple instruction: "Read through these messages and tell me the most common phrases people use to describe why they came, why they stayed, or why they left. Quote them directly, don't paraphrase."
What came back wasn't "get fit." It was "I don't have time" appearing 61 times. It was "I keep starting and stopping." It was "I need someone to notice if I don't show up." Nobody in 18 months of messages had used the word "strength." They used the word "accountability" constantly, and the word "guilt" more than you'd expect from a gym.
We rewrote the landing page headline from "Build Strength That Lasts" to "Someone Will Notice If You Don't Show Up." Ugly, a bit uncomfortable, exactly what her customers say to each other in the group chat. The page went from 2.1% to 3.4% conversion within six weeks, no other changes made to the page. That's not a huge sample size and I won't pretend it's a controlled experiment, but it's the same pattern I've now seen with a bookkeeper, a wedding photographer and a B2B software reseller: the words customers already use outperform the words marketers invent almost every time.
Why this works (and why it's uncomfortable)
Here's the bit that's slightly awkward to admit: most marketing advice, including a fair bit I've given over the years, assumes the business owner's job is to craft the message. It isn't, not really. The message is usually already sitting in your inbox, your reviews and your DMs. Your job is extraction, not invention. That's a less flattering version of "marketing expert" than most of us want to sell, but it's the one that gets results.
It also means the AI tools that generate the flashiest, most polished-sounding copy are often the least useful step in the process. A tool that writes beautifully from a blank prompt is solving the wrong problem. The valuable move is the unglamorous data-in step before you ever ask it to write anything. This is the same instinct behind why a tool like Hotjar built its entire business on watching what real users do rather than what they say they'll do in a survey: real behaviour and real language beat guesswork every time, and it's been true since long before AI existed. It's the same principle Dale Carnegie was teaching in the 1930s, talk in terms of the other person's interests, using their own words, not yours.
How to do this yourself, this week
You don't need a data analyst or an expensive platform. Here's the exact process, and it takes about two hours the first time, less after that.
- Step 1: Pull your raw text. Export the last 6 to 12 months of customer support emails, live chat transcripts, Google or Trustpilot reviews, cancellation survey answers, and Facebook or Instagram comments. Aim for at least 100 to 150 individual pieces of text. Copy them into one plain document. Strip names if you're worried about privacy.
- Step 2: Feed it to the AI in chunks. ChatGPT and Claude both handle large pastes fine, but if you have thousands of messages, break it into batches of 200 to 300 at a time. Don't ask it to write anything yet.
- Step 3: Ask for direct quotes, not summaries. The prompt that matters most is something like: "Read this text and pull out the most frequently repeated phrases customers use to describe their problem, their hesitation, or their reason for buying. Quote them word for word. Do not paraphrase or make them sound more polished."
- Step 4: Look for the phrase you'd never have written yourself. You're hunting for the language that feels a bit rough, a bit too honest, the kind of thing you'd never put on a brochure. That's usually the winning line.
- Step 5: Test it in one place first. A headline, a subject line, one ad. Not your whole website. Give it two to three weeks and compare against what you had before.
If you run a service business with recurring customers, do this quarterly. Language shifts as your customer base changes, and a phrase that worked in January can go stale by autumn.
Where this connects to bigger AI decisions
This one exercise is a small, doable thing you can run in an afternoon. But it points at a bigger truth about how small businesses should be using AI generally: the value isn't in the flashy generation, it's in the unsexy analysis of information you already own. Brands that get this right build entire strategies around it. Duolingo's marketing team is famously obsessive about the exact language and behaviour of their users before a single social post goes out, and it shows in how specific and odd their content feels compared to competitors who just guess. Revolut built its early growth on solving the specific, plainly stated frustrations customers had with old banking apps, not on inventing a new emotional pitch from scratch.
If you're already collecting customer language this way, the natural next step is turning it into something interactive that gets visitors giving you even more of it, which is exactly the thinking behind using interactive calculators and quizzes on your site: every answer someone gives you is more raw material for this same process.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
None of this requires expensive software. A free ChatGPT account or the free tier of Claude is enough to run the whole exercise. Where I do see business owners get stuck isn't the tool, it's the time and objectivity to read their own inbox without flinching, and to trust an ugly phrase over a polished one. If that's you, that's usually the point where bringing in outside help pays for itself, whether that's a few hours with an AI consultant who can run this analysis across your whole customer base and build it into your ongoing marketing process, rather than a one-off exercise you do once and forget.
What this doesn't fix
I want to be honest about the limits here, because I've seen business owners treat this as a silver bullet. It won't fix a product people don't want. It won't fix pricing that's wrong for your market. And it works better for businesses with some volume of existing customer contact, meaning if you've only got 12 customers and three reviews, you won't have enough raw text to spot real patterns yet. In that case, the move is to start collecting it deliberately, ask every customer one open-ended question after purchase, "what almost stopped you from buying?", and build your data set from there before you run the analysis.
It's also worth saying plainly: the AI isn't finding anything magical. It's doing pattern-matching across text that a patient human could do by hand given enough time. The value it adds is speed, not insight you couldn't get yourself. I've read all 340 of that Kent client's messages myself in the past, over the course of a long, tedious Sunday. The AI did the same job in about four minutes. That's the entire pitch: not a new capability, just hours given back to you.
Frequently asked questions
What's the difference between this and just asking AI to write better marketing copy?
Asking AI to write from scratch means it's guessing at your audience using general internet patterns. Feeding it your actual customer messages first means it's reporting back real, specific language your customers already use, which nearly always outperforms invented copy because it sounds like something a real person would say, not a marketer.
How much customer text do I need before this is worth doing?
Aim for at least 100 pieces of text, reviews, support messages, comments, cancellation reasons, combined. Below that the patterns are too thin to trust. If you're under that number, start asking one open-ended question to every new customer for the next two months and build your data set before running the analysis.
Is it safe to paste customer messages into ChatGPT or Claude?
Strip out names, emails, phone numbers and any payment details first, and stick to the free or standard consumer tiers only for non-sensitive text. For anything involving health, financial or otherwise sensitive customer detail, use a business or enterprise account with data controls, or don't paste it at all and work with an anonymised summary instead.
Can this work for a business-to-business company, not just consumer brands?
Yes, and it often works even better because B2B sales calls and support tickets tend to contain very specific, quotable objections, "we tried a similar tool and it broke integration with our CRM" is exactly the kind of line that makes a landing page far more convincing than generic B2B copy about efficiency and growth.
Related reading: How to Get More Views on Instagram Reels (What Moved the Needle for Me) and How to Find Out Who Viewed Your Instagram Profile (What Happens When You Try).
More on this here: write for us about fitness.
