Your Customers Already Wrote Your Best Ads: The AI Review-Mining Workflow We Use Before Writing a Word

Let's put a number on what guessing costs you. Say you run dedicated testing - one concept, its own budget, A$100 a day for five days. That's A$500 a concept. Test ten of them and you've spent A$5k in a week. If only one in ten clears your target, that single winner has to pay back the A$4,500 you lost on the other nine before it earns you a cent. That maths doesn't work. You're bleeding from the first test.

Now lift that to three in ten clearing the bar. Same A$5k, but three winners covering the losers and pulling ahead. The exact same testing budget goes from a slow leak to a machine that pays for itself.

The gap between one-in-ten and three-in-ten isn't your creative team getting more talented. It's whether your angles were grounded in what customers actually care about, or whether someone in a meeting decided what they care about. Most brands are guessing. And the wild part is the answer was sitting in their reviews the whole time.

Your customers have already written your best ads. They did it in your reviews, your support inbox, and the comments under your posts. The job isn't to invent angles. It's to go and read what they told you, count it, and hand it back to them in their own words. Here's the exact workflow we run before anyone writes a single hook.

The rule that sits over all of it

One hard line first, because it's the thing that makes this work and the thing most people break: you mine real customer language. You never fabricate.

It's tempting to ask the AI to just invent a glowing review and slap it on a static. Don't. Customers have been lied to with thousands of fake reviews on their feed, and they can smell a made-up one. So can you, honestly, if you read it cold. The most powerful thing in advertising is the truth - the actual words a real person used about a real experience. The AI's job here is to find and sort that truth at speed, never to manufacture it.

Keep that in your head for the whole process. Every output traces back to something a human actually said.

Step 1: Gather the raw material

Pull everything where a customer described your product in their own words. That means:

• Product reviews - all of them, not just the five-stars.
• Support tickets and email replies.
• Comments under your ads and organic posts.
• Post-purchase survey answers.
• Any DMs or chat logs you can export.

Dump it into one document. Don't tidy it. The typos and the run-on sentences are the good stuff - that's how your buyer actually talks, and that's the texture you're trying to capture.

One quick filter that punches above its weight: sort your reviews by length. In a spreadsheet, run a character-count on each one and sort longest to shortest. The longest reviews are where the emotion lives - someone only writes four paragraphs about a knee sleeve because they had a story. Take the longest few hundred and treat them as your richest seam.

Step 2: Tag every piece into an angle type

Now you bring the AI in, and the first job is sorting, not writing.

Feed it the pile and have it tag each review, comment, or ticket as one of a small set of angle types. The set we use:

• Problem-aware - they name the pain they had before buying.
• Solution-aware - they compare against other things they tried.
• Social proof - they talk about trust, reviews, others, "I was sceptical".
• Curiosity - something surprised them, an unexpected use.
• Benefit-driven - they describe the result they got.
• Authority - they reference the science, the ingredients, the credentials.

A prompt as plain as "tag each of the following reviews with one of these six angle types, and pull the exact sentence that earned the tag" gets you most of the way. You're not asking for cleverness. You're asking it to read a thousand reviews in a minute, which is the one thing it's genuinely better at than you are.

The output you want is a table: the customer's words, and the bucket they fall into. Keep the sentence verbatim. You'll need it later.

Step 3: Count the frequency, and let it set your test order

This is the step almost everyone skips, and it's the one that does the work.

Once everything's tagged, count how often each angle type shows up. You're looking for the pattern - which problem comes up again and again, which benefit nearly everyone mentions, which objection keeps surfacing.

That frequency count is your test order. The angle a hundred customers raised gets tested before the one three customers mentioned. You're no longer ranking angles by which one sounds best in a meeting. You're ranking them by how many real buyers already told you it mattered. That single reorder is most of the jump from a one-in-ten hit rate to something you'd actually want to fund.

It also kills a quiet trap: testing five flavours of the same angle. If everyone's running the same generic "before and after" and yours fails, all you've learned is that one angle doesn't work. Five different angles drawn from five different frequency clusters means five different pockets of buyer, and if three land, you've got three customer segments to scale into instead of one dead end.

Step 4: Pull the exact language into hooks

Now, and only now, you write. Except you're barely writing - you're lifting.

Go back to the verbatim sentences under your top-frequency angles and turn them into headlines and hooks. The best ones often need no edit at all. "My secret weapon against the 3pm slump." "I finally found something my fussy toddler will actually eat." Those aren't copywriter lines. They're review quotes that happened to be perfect.

Two specific things to hunt for while you're in there:

The first is the insider phrase - a word or bit of slang only someone with this exact problem would use. When you find one and put it in a hook, it does two jobs at once: it tells your buyer "this is for me" and tells everyone else "this isn't for you". That qualifying-and-disqualifying move is worth more than any clever line you could write, because the right person reads it and thinks these people must get it. They're hard to find. You usually need a big pile of reviews to surface one. When you do, it's gold.

The second is genuinely emotive language - the lines from those long reviews where someone said how the thing made them feel, not just what it did. "For the first time in years I'm not embarrassed to wear shorts." That's the stuff that stops a thumb. The best copy doesn't feel like selling, it feels like being understood, and you can't fake that feeling. You can only find it where a real customer already wrote it.

While you're sorting, get the AI to rank the failed solutions customers mention - the things they tried before you that didn't work. That ranked list is the spine of a comparison angle. You're not guessing what to position against. Your customers ranked it for you.

Step 5: Turn the angle into a brief

A hook on its own isn't a brief. Before this goes to whoever makes the creative, wrap each angle in three things, which is the shape we call APAP - audience, painpoint, angle, product.

• Audience: who specifically is this for? Pull the persona straight from the reviews - the AI can build it from the same pile, so it's grounded in real buyers, not a made-up avatar.
• Painpoint: the exact problem, in their words, from your problem-aware bucket.
• Angle: the specific promise or framing you're leading with.
• Product: what they can actually buy at the other end.

Write a one-line hypothesis too: I believe this angle will land with this person because the reviews show X. If you can't write that sentence honestly, the angle isn't ready and you shouldn't spend money on it. That little discipline stops you funding hunches dressed up as strategy.

Done properly, the brief almost assembles itself: this product, for this person, with this painpoint, led by this angle. The research did the hard part.

Step 6: Match the format to the product

Last thing before it ships, and it's a cheap way to lose hit rate if you skip it: put the angle in the right format for what you're selling.

Products that need explaining - supplements, skincare, anything where the buyer has to understand a mechanism before they'll trust it - generally want video, or a static that only sells the click into an advertorial that does the educating. You can't teach someone why they need a complex thing in one still image.

Products that explain themselves - apparel, jewellery, homewares, the things you understand the second you see them - do fine on static. A great review headline over a clear product shot is often all you need.

Run a supplement on a bare static and you can have the perfect angle and still watch it die, because the format couldn't carry the explaining. Same angle, moved to video, and the number changes. The research can be right and the format can still sink it, so don't treat this as an afterthought.

Where to from here

Notice what we didn't do anywhere in there. We didn't brainstorm. We didn't sit in a room arguing about what the customer wants. We read what they already said, counted it, and handed it back.

If you do nothing else this week, do the first three steps. Pull your reviews into one doc, have the AI tag them into angle types, and count the frequency. You'll likely find one or two angles you've never run that your customers have been telling you about for a year - and probably one you've been spending good money on that almost nobody actually mentions.

Run it, and tell me what surfaced. The angle hiding in your own reviews is usually the one you least expected.