Most ad personas are fiction. “Marketing Mary, 34, loves yoga and efficiency.” Nobody has ever sold anything to Marketing Mary, because she does not exist. Over more than a decade rebuilding underperforming campaigns, I have found the fix is almost never the bidding. It is the match between the message and the person. AI has made building that match faster than it has ever been. This is the process, step by step.
Why most personas fail before the first ad runs
A persona has one job. Predict which message will move which person to act. Most fail because they are built from the inside out: a workshop, a whiteboard, a stack of assumptions about the customer everyone wishes they had. The industry has noticed the gap. Synthetic and AI-generated personas are having a real moment in 2026, and the good tools cut research time from weeks to days. But a persona built on guesses is still a guess, however quickly a model produces it. The evidence has to come from the customer, not the meeting room. This persona work is one play inside the wider AI playbook I would hand a small team.
Step one: collect the language customers already use
I gather every scrap of real customer text I can legally access. My own reviews, competitor reviews, marketplace Q&A, forum threads. Then I let a model do the first pass. I ask it three things:
- What problems people describe in their own words.
- What objections show up right before someone decides not to buy.
- What words happy customers use after buying.
The output is a vocabulary, not a persona yet. The gold is in the phrasing. Malaysians do not search for “artisanal durian confectionery”. They ask whether the Musang King is “ori”.
Step two: cluster by buying situation, not demographics
I ask the model to group the evidence by situation. Someone buying a gift and worried about delivery time. Someone comparing two brands and suspicious of fakes. A repeat buyer watching the price. Situations beat demographics because ads target moments of intent, not birth certificates. Three to five clusters is plenty. If the model hands me nine, I make it defend each one. The weak ones collapse under questioning.
Step three: write a message grid per situation
For each situation I draft four things: the pain in the customer’s own words, the promise that answers it, the proof I can offer, and the call to action. The model drafts this grid quickly. Then my judgement earns its keep. I cut every promise I cannot prove, and every clever line that does not sound like the vocabulary from step one. When the message lands in the same words the customer used to describe the problem, response rates climb. That is the whole game, and it is worth more than any bid adjustment I have ever made. The account that taught me this was GNC Malaysia, back in the early days of social ads. We spent just shy of RM2,000 and came back with a return of 30 times that in sales, because the message matched the person seeing it. I will be honest, those were easier days and nobody should promise you 30x today. For what current campaigns look like, the OpenMinds team keeps a set of recent Malaysian examples with real numbers.
Step four: test personas as hypotheses, not facts
Each situation becomes its own ad set with its own message. I run them against each other honestly. Two things I always tell clients. Let the data kill your favourite. And a losing persona has often just met the wrong creative, so check the message grid before you delete the audience.
The part AI still cannot do
AI compresses the research from weeks to days, and it is genuinely good at spotting patterns across a thousand reviews. Every serious write-up on synthetic personas this year lands on the same caveat: the model surfaces the pattern, a human still has to confirm it holds. What a model cannot do is sit in the sales conversation, hear the hesitation in someone’s voice, and realise the real objection was never written down anywhere. Keep doing that yourself. Then feed what you learn back into the grid.
If your ads are getting clicks that never become customers, the match between message and person is the first thing I would check.
