Buyer personas in the age of AI, from real data

Anna, 34: buyer persona, workshop version DOESN'T EXIST
Occupation	marketing manager
Family	husband, two children
Location	Budapest, District XI
Hobbies	yoga, travel, food
Goal	staying in shape
Source	the team's assumptions

We tested what happens when public data replaces the assumptions. We picked a single category, yoga, and collected 2,757 Hungarian search terms for it, plus 351 written Google reviews of twelve yoga studios in Budapest. AI helped with the processing. By the end, six and a half Annas had taken shape, each with a channel, a message and a format.

Where did Anna come from, and why is she still here?

The persona started out as a software design tool. Alan Cooper introduced it in 1999, in his book The Inmates Are Running the Asylum. His aim was to get developers to design for one concrete person rather than a faceless user. Marketing borrowed the idea and added a photo, an age and an income. That is how Anna was born. A face is easier to remember than a spreadsheet.

In most companies the buyer persona is put together in a workshop, where the team says who it thinks the buyer is. The loudest person's opinion ends up on the slide. A buyer is rarely in the room. Nobody checks the finished document against reality, and the file stays unchanged for years. Anna is still 34, even though she was drawn up five years ago.

The second problem is that age is not behaviour. Two 34-year-old women can go to the same yoga class for completely different reasons: one wants to protect her lower back, the other wants to become a teacher. Ad platforms have no button called Anna either. You can target searches, interests and customer lists, in other words what people do. Researchers have known this for a long time: personas have been built from behavioural data since the mid-2000s, and Bernard Jansen and Joni Salminen wrote a book on the method. Until recently, it took a team of data scientists. Language models have made it simple.

Let's see it in practice

We worked from two public sources. Searches show what someone types before they know us. Reviews show what they say after they have bought. Between the two lies the whole buyer journey. We chose yoga because it almost always appears among the invented Anna's hobbies. If the classic buyer persona is going to hold up anywhere, it should be here.

On the search side, we pulled every search term in Hungary that contains the Hungarian word for yoga, using monthly averages from Google Ads data for Hungary. That came to 2,757 terms, together 90,120 searches a month, excluding the bare word "yoga" itself. The reviews come from the Google profiles of twelve Budapest studios: 351 written reviews, 309 of them in Hungarian and 39 in English. The sample is skewed. 91% of the reviews are five-star, and a third of the texts belong to a single studio, so satisfied regulars sound louder than they really are.

AI labelled every search term and every review against a codebook fixed in advance. We describe the steps below so anyone can repeat them. The analysis took one afternoon. A company working with its own data can go further: customer service emails, sales call transcripts and purchase data can be processed the same way, once names are redacted.

What do searches reveal about the buyer?

The biggest group is branded search. 29% of monthly searches contain the name of a studio or a teacher, and that is a lower estimate, because smaller names are hiding among the unclassified terms. These people had already decided by the time they typed into the search box. Someone recommended the place to them. For them, most of the buyer journey happens in conversations that no measurement tool can see.

Local searches make up 13.4%, yoga styles 11.9%, equipment and clothing 8.2%. The surprise is at the bottom of the list. Searches about weight loss and body shaping sit at around 0.5%. Chair yoga, senior yoga and exercises for the over-fifties, on the other hand, add up to 1,900 searches a month, almost four times as many. In January this jumps to 5,360, and by April it falls back to 930. A studio that advertises to Anna knows nothing about this audience.

880 people a month search for yoga teacher training, exactly as many as search for yoga mats. 390 search for pregnancy yoga. The market as a whole is strongest in January, with 111,000 searches a month, drops to 72,000 by summer, and climbs back above 100,000 in September. Timing is part of the buyer persona. Someone who arrives in January has made a resolution. Someone who arrives in September is starting a new work routine.

How do 651 Google reviews become a buyer persona?

Anyone can read reviews, yet few people read all of them, and even fewer count anything from them. The process has seven steps. For each one we show what goes in, what comes out, and what it tells us about the buyer.

Filtering: how much is left?

The twelve studio profiles hold 651 reviews. Three hundred have no text, only stars. Of the remaining 351, 53 are empty praise, five and a half words on average, like “Great place!”. 298 reviews have real content, thirty words on average. In 140 of them you can tell what situation the writer is in. That is 22% of all reviews. Two in ten ratings reveal something about the buyer, which is why one studio's reviews are rarely enough, but a whole category's are.

Codebook: six questions for every review

Before using AI, you have to decide what you are asking. Social research calls this a codebook: a labelling guide that lists the possible answers to each question and describes when each one applies. We asked every text six questions, each with a closed list of answers. The codebook fits on one page. Everything depends on it, because the same text can be labelled into two different buyer personas with two different sets of questions.

Three rules made it usable. First, only what is written counts: inferring age, gender or intent is not allowed. Second, every question offers a "can't tell" answer, because without it the model guesses, and a guess looks exactly like a fact in the spreadsheet. Third, an order of precedence. If a review points to two situations, for example a pregnant woman writing about her first class, the rarer and more revealing situation wins.

What exactly does the model do?

The labelling was done by the Claude Sonnet language model. We split the 351 reviews into four batches, and every batch got the same inputs: the codebook, the reviews and a short instruction. The core of the instruction is four sentences. Read the full text. Decide the answer yourself, without keyword matching. Label only what is written. Copy the most characteristic passage word for word, in twelve words or fewer. The output is one row per review in a fixed structure, so it can be counted like a spreadsheet.

The difference from keyword search is measurable. A search for the word "first" found 33 reviews, but the model identified 20 first-time visitors, because the word often means something else, for example that someone felt at home from the first moment. With teachers it is the other way round. The word appears in 126 reviews, but the model found praise for the teacher in 186, because visitors often mention the person they go to by first name only. A keyword counts words. The model reads meaning.

How do we know a label is right?

We ran three checks. The first is formal: a script checks that every review got its labels, that every code is valid, and that every quote can be found letter for letter in the original text. One quote failed because of a single lowercase letter and went back for correction. The second is the model's own flag: it marked 46 reviews, 13% of the texts, as uncertain, mostly when judging the situation. These, and the negative reviews, need a human to look at them.

The third is repetition. We ran the full labelling again, without access to the first result. The situation judgement matched for 97% of the reviews, the visit type for 97%, and the content rating for 96%. The full list of praise matched in 87%. This proves stability, not truth, since both runs could be wrong in the same way. And small numbers move: in the first run, 4 of 20 first-time visitors gave a negative review, in the second, 3 of 19. The direction is the same, the exact ratio is uncertain.

What do people praise, and what do they complain about?

186 reviews praise the teacher, that is 53% of the texts, and 98 mention the teacher by name. There are five complaints about teachers. 175 praise the space, six complain about it. With the welcome and the organisation, the ratio flips: the welcome gets 35 compliments and 12 complaints, while booking, passes and payment get 10 compliments and 8 complaints. Of the 24 negative reviews, 20 are not about the teacher. The disappointment happens at the entrance, on the phone and in the booking system.

Segmentation: who wrote it, and why did they come?

Three breakdowns gave a usable picture. The first is visit type. Of the 20 first-time visitors, 4 gave a negative review. Of the 55 regulars, none did. This is partly self-evident, since anyone who is disappointed does not come back, and that is exactly why the first visit is the riskiest moment. The second is motivation: recharging appears in 80 reviews, the body in 33, community in 27, learning in 22, and a life situation in 18. Weight loss in none. The third is language: almost a quarter of people writing in a language other than Hungarian are dissatisfied, compared with 5% of those writing in Hungarian.

What do the two sources see differently?

In searches, the biggest group is people arriving on a recommendation, with 26,170 searches a month. The reviews contain eight such texts, because people who came on a recommendation rarely say so. The over-fifties account for 1,900 searches a month and three reviews: they practise at home, or they don't write reviews. For people looking to recharge it is the reverse: in searches they are scattered, in reviews they are the loudest. Neither source is enough on its own. Without the biggest studio, the number of pregnant women drops from 18 to 5, and the number of people looking to recharge from 47 to 21, so this is the step where you have to check for sample bias.

Where can AI get it wrong?

AI can write Anna without any of this, in ten seconds, and more polished than any workshop. If you ask it for a buyer persona without data, it hands back the stereotype from its training material in a confident voice. Researchers at the Nielsen Norman Group found that synthetic users enthusiastically agree with everything and treat every need as equally important, so their answers are only worth treating as assumptions to be tested. The invented buyer is polite. The real one is not, and that is exactly what makes them useful.

There is evidence from the other direction too. Researchers at Stanford conducted two-hour interviews with 1,052 people, then built AI agents from them. The agents built from interviews reproduced the participants' survey answers with 83% accuracy, measured against how consistently the people themselves repeated their own answers two weeks later. Agents built from demographic data alone reached 74%. So Anna is not enough for the machine either. A model is only as good as the real material it gets, which is why every claim in the buyer personas below comes with a number or a quote.

Our Annas: six and a half buyers who emerge from the data

The buyer personas below include no age or income, because the data does not split along those lines. Each one describes a situation, and the same person can move through several of them over the years. Anything shown with a number or a quote, we measured. The suggested channel, message and format are assumptions derived from it, to be tested in a campaign.

1. Anna, whose colleague sent her
Situation	She comes on a recommendation, and her decision was made before she searched. She searches for a studio or teacher by name.
What we measured	Branded search: 26,170 a month, 29% of all searches. 53% of reviews praise the teacher, 28% by name.
In her words	“It was my first time here, on a friend's recommendation.”
Where we reach her	Nowhere with cold advertising. She reaches us through existing visitors, the teacher's own social media page and the Google profile.
How we speak to her	With reassurance. She is looking for the teacher's name and timetable, so that is what she should see first.
In what format	A teacher page with the timetable and two-click booking, a bring-a-friend session included in the pass, a review request after class.
When	All year round.

2. Anna, who closes her laptop at seven in the evening
Situation	She comes after work to put the day down. Exercise is secondary.
What we measured	Yin, nidra, stress relief, morning and evening yoga together account for about 3,000 searches a month. In 23% of reviews, recharging is the stated reason.
In her words	“I arrived stressed and I leave feeling like I've had a full night's sleep.”
Where we reach her	In local search and on the map, with the district name. With social ads in the late afternoon.
How we speak to her	In her own words: switch off, slow down, calm down. There is no point talking to her about poses and styles.
In what format	A short video of the relaxation at the end of class, an evening timetable, online booking with no waiting.
When	Weekday evenings, peaking in September and January.

3. Anna, who stands at the back
Situation	She is getting ready for the first yoga class of her life. She is afraid of sticking out.
What we measured	Beginner searches account for about 2,700 a month, 3,980 in January. Of 20 first-time reviewers, 4 gave a negative review. Of 55 regulars, none did.
In her words	“I didn't know what the protocol was.”
Where we reach her	In search, on beginner terms, and on video platforms, where she tries it at home first.
How we speak to her	We take away her fear. In one visitor's words: don't worry about anything, just come down for a class.
In what format	A first-class page that tells her which door to use, what to bring and where to stand. A beginner course with a fixed group, a message after the first class.
When	In January and September, and the campaign has to start before then.

4. Anna, who is expecting a baby
Situation	Pregnant or with a newborn. She is looking for safety and time for herself, in a nine-month window.
What we measured	Pregnancy yoga and mum-and-baby yoga together account for about 1,900 searches a month, often with a district name. 90 a month for the online version.
In her words	“I started coming when I was pregnant, then moved on to mum-and-baby yoga and later hatha yoga too.”
Where we reach her	In local search by district, on the map. Through doulas and antenatal classes. The doula shows up in the reviews too.
How we speak to her	With an experienced teacher and calm. The word performance puts her off.
In what format	A district page, a pass that can be paused, an online alternative. After the birth, an email with the next step.
When	Steady all year. You have to respond fast, because her time is limited.

5. Anna, who starts with a chair in January
Situation	Over fifty or close to retirement, looking for gentle exercise. She often starts at home.
What we measured	Chair yoga and senior yoga together account for 1,900 searches a month, 5,360 in January. The word free appears in 180 searches a month, the 28-day challenge in 100.
In her words	“I was looking for something that keeps me in shape without straining my joints.”
Where we reach her	In search and on video platforms. Family is a channel too: one grandmother brought along her daughter and her granddaughter.
How we speak to her	Joints, balance, safety. A photo of someone practising on a chair instead of flexible young people.
In what format	A free 28-day video series in exchange for a sign-up, then a morning senior class at the studio.
When	It has to be ready by late December. By April, a sixth of January's interest remains.

6. Anna, who wants to teach
Situation	She has done yoga for years and wants to change careers or find a side job. She weighs it up for months.
What we measured	Training-related searches account for 1,530 a month, 2,100 in January. The term for yoga teacher training alone accounts for 880.
In her words	“I'm glad I took the plunge.”
Where we reach her	Among our own visitors, by email and in the studio. In search, on training terms.
How we speak to her	With facts: duration, fee, certificate, curriculum. With reviews from graduates.
In what format	A detailed training page with a downloadable curriculum, an open information evening, an email sequence over several weeks, a one-to-one conversation.
When	For a September start, the decision begins in spring. For a January start, in autumn.

And a half: Anne, who doesn't speak Hungarian
Situation	A foreigner living in Budapest, or passing through. Invisible in Hungarian search data.
What we measured	11% of reviews are in English, and almost a quarter of people writing in a language other than Hungarian are dissatisfied. The term yoga budapest gets 210 searches a month from Hungary, yoga near me gets 480.
In her words	“I wish more English speaking teachers could be available.”
Where we reach her	On the map and in English-language search.
How we speak to her	In English, and honestly. A class advertised as bilingual but taught in Hungarian brings a negative review.
In what format	An English page, English booking, the language marked in the timetable.
When	All year round.

What can you do with a buyer persona like this?

The first consequence is that a studio needs six landing pages. The same room and the same timetable greet the person who is afraid, the person who is tired and the person who wants to teach with a different first sentence. The ad headline becomes the buyer's own sentence, almost word for word. It is the cheapest copywriting there is, because the sentences are already in the reviews. That way the buyer personas turn directly into ad copy, email subject lines and the opening line of a video.

The second is the calendar. By late December the chair yoga and beginner campaigns have to be ready, because the wave comes in January and is gone by April. People interested in teacher training need to be reached months before the course starts. Pregnant women arrive all year round. An annual marketing plan that says the same thing to everyone every month hits at most one of the six situations, and only by chance.

The third is the order of things. Ads bring in new visitors, and new visitors arrive exactly where most bad experiences happen: the entrance, the reception desk, the booking system. Fix that first. Then you can spend. Here the buyer persona goes beyond marketing, because it tells you what the receptionist needs to learn and what belongs in the confirmation email.

The fourth is measurement. Every Anna has a search group, a page and a goal, so every quarter you can see which assumption held up. The ones that didn't get rewritten. Nobody could ever disprove the Anna born in a workshop. These ones can be.

Does Anna exist after all?

The woman in the photo could be any of the six. Maybe her colleague sent her, maybe she closed her laptop at seven in the evening, and maybe next year she will be the one teaching the class. Her age does not tell you that. Her search does, and so does what she writes after class. Anna, the 34-year-old marketing manager, still doesn't exist. The situations do, and real people find themselves in them.

Anna: buyer persona, measured version MEASURED
Age	we don't know, and it isn't what decides
Situation	one of the six
What she searches for	we know from 2,757 terms
What she says	we know from 351 reviews
Valid	until the next quarterly update
Source	measured data, verbatim quotes

Methodology and sources

Search data: Google Ads average monthly search volume via the DataForSEO database, Hungary, Hungarian language, September 2026. Reviews: public Google reviews of 12 yoga studios in Budapest, 651 reviews, of which 351 have text. The reviews were labelled by the Claude Sonnet language model against a codebook (labelling guide) fixed in advance, in four batches, without keyword matching. A script checked that the quotes match the originals word for word, and a second, independent run repeated the full labelling. The two runs matched on the situation judgement for 97% of the reviews. Human review was done on a spot-check basis. The photo is an illustration, and the person in it is not the fictional figure described in the article.

Curious who your real buyers are?

Let's talk about how your own searches and customer reviews can become a usable buyer persona, campaign plan and ad copy.

Request a consultation

Anna is 34, upper-middle-class, has two kids, and doesn't exist: buyer personas in the age of AI, built from 2,757 searches and 351 customer reviews