Audience research
Part of How to run AI marketing tools audience research that guides campaigns
How to build AI marketing tools buyer personas from first-party data
How to build AI marketing tools buyer personas from evidence: the fields that matter, a before and after table, and the compliance checks to run first.
What to take away
A buyer persona is a written profile of the person you want to reach, covering the job they do, the decision they make and the evidence behind both.
- Start from data you already holdenquiry records, sales notes and converting search terms.
- Treat AI marketing tools as sorting and summarising aids, not as sources of motivation.
- Write the trigger, the objection and the proof each persona needs, then score the evidence for every claim.
- Check advertising and data protection rules before persona logic drives live campaigns.
- Review each profile quarterly, because a persona nobody tests stops being true.
Where standard personas go wrong
Most personas are written from the inside out. Someone decides what a finance manager cares about, adds a stock photograph and a name, and the document is never opened again. The result describes a mood, not a decision.
Weak persona becomes budget line
- Thin data grouped by job title
- Confident summary reads as good data
- Weak persona becomes targeting rule
- Targeting rule becomes budget line
- Copy answers neither researcher nor budget holder
Tools make this quicker. A model can group a contact list by job title in minutes, which is useful. Grouping is not explaining, though, and a confident summary of thin data reads like a confident summary of good data.
That gap matters once segments are built automatically. A weak persona becomes a targeting rule, then a budget line. A grouping built on job titles alone will merge a researcher and a budget holder into one profile. They ask different questions and stall at different points, so the copy ends up answering neither.
Evidence worth collecting
Four sources carry most of the weight: what people did on your site, what they said in sales conversations, what they search for, and what competitors pay to say. The first two are yours alone and are usually the most reliable. Record the reason a deal stalled and how long approval took, because neither detail appears in an analytics report.
Evidence sources: strength and limits
Source
- On-site behaviour
- Yours alone
- Sales conversations
- Yours alone
- Search demand
- Weaker signal
- Competitor ads
- Public
Strength
- On-site behaviour
- Most reliable
- Sales conversations
- Most reliable
- Search demand
- Curiosity, not intent
- Competitor ads
- What they pay to say
Watch out
- On-site behaviour
- Sales conversations
- Search demand
- Competitor ads
The companion guide to audience research sets out sampling and interview design, and shows how to avoid collecting more personal data than the purpose requires when transcripts and call recordings feed a model.
Search demand is a weaker signal than its volume suggests. A busy broad term can mean curiosity, not purchase intent, so pair it with a commercial signal such as a pricing page visit.
The build, step by step
- Agree the decision.Write down what the person is choosing — which AI marketing tool to adopt, at what tier, and who signs the contract. The decision, not the job title, is what the persona has to support.
- Gather the first-party evidence.Export site behaviour, enquiry records and converting search terms from your own analytics, then add sales notes and call recordings. Transcription tools such as Otter.ai, Gong and Fathom turn a recorded call into text you can search and tag.
- Group the contacts.HubSpot, Salesforce Einstein and Braze cluster an existing list into provisional segments, and ChatGPT, Claude or Google Gemini will summarise the same export if you would rather not load it into a platform. Grouping is mechanical and settles nothing about motivation.
- Code the objections.Read the transcripts and tag why deals stalled, in the words the buyer used.
- Theme the survey.Dovetail, or the AI features in Typeform and SurveyMonkey, sorts open-text answers into themes in minutes. Keep the raw answers beside the summary, because the theme labels are the part a model gets wrong.
- Draft the profile.Write it in ChatGPT, Claude or Jasper, then check every line against a source in your evidence file and cut what you cannot source.
- Score and review.Mark each field high, medium or low confidence, and revisit the profile quarterly or after a pricing change.
What AI buyers raise that others do not
Buyers of an AI marketing tool arrive with a different set of questions from the average B2B prospect. The trigger is usually a bottleneck — reporting that takes a week, a campaign calendar nobody can keep up with — rather than a new budget line.
The objections repeat, so the persona should carry an answer to each one:
- Integration:will it read from and write to the CRM the team already runs?
- Data residency:where is the input and output data stored, and who can reach it?
- Model accuracy:how is the output evaluated, and who checks it before it reaches a customer?
- Seat-based pricing:what happens to the cost as the team grows?
- Security review:how long does procurement take, and what evidence does it expect?
A profile that leaves these unanswered stalls in the same place every time.
Rebuilding the persona: before and after
The table uses illustrative examples for a fictional mid-sized B2B supplier — a marketing manager weighing up an AI tool — and not a named client.
Before
- Job title
- Marketing manager
- Goal
- Wants better results
- Trigger
- Looking for tools
- Objection
- Price
- Evidence
- Team assumption
- Owner
- Unassigned
After
- Job title
- Marketing manager at a mid-sized B2B supplier, owns the paid budget
- Goal
- Needs to show cost per enquiry fell this quarter
- Trigger
- A new finance director asked for a channel breakdown
- Objection
- Cannot justify another subscription before the annual renewal
- Evidence
- Enquiry records and two call transcripts from the last quarter
- Owner
- Demand generation lead, reviewed each quarter
What changes in a campaign
Once the profile is written down, the automated parts of a campaign have something to check against. Segment names should follow the trigger rather than the job title. Copy should answer the stated objection, and landing pages should supply the proof point.
Compliance sits alongside this. Automated variants can breach the CAP Code — the UK Code of Non-broadcast Advertising and Direct & Promotional Marketing — without anyone writing a deliberately misleading line, and the Advertising Standards Authority (ASA) enforces it. Read the sanctions available against non-compliant advertisers before widening an audience. Persona fields that touch on health or financial difficulty raise the risk again.
Agree one measure per persona before launch, such as cost per qualified enquiry, and let the result settle whether the profile was accurate.
If you are still deciding how wide a segment should be, the notes on target audience sizing cover how to size and split an audience in England, which is the step underneath persona work.
Data protection and procurement checks
Personas come from personal data unless you aggregate carefully first. The governing law is the UK GDPR and the Data Protection Act 2018, with the Privacy and Electronic Communications Regulations 2003 covering email and cookies. The Information Commissioner's Office (ICO) enforces all three, and its decision notices show enforcement following poor handling of marketing data.
Delete any field the documented purpose cannot justify. Record the lawful basis for each purpose in the ICO's lawful basis guidance, and run a data protection impact assessment before anything relies on automated profiling, which Article 35 of the UK GDPR expects for large-scale profiling. Set a retention period for each field when you create it and hold it in a retention schedule, because deleting a column later is harder than never collecting it. You can also check a supplier on the ICO's public register of fee payers before you sign.
Supplier questions deserve the same treatment. The NIST AI standards programme is a sensible reference when you ask a vendor how a model was evaluated and tested, and it helps you compare answers rather than accept them.
Common questions
Can a tool build personas without human input?
It can group and label whatever you feed it. It cannot supply the sales call detail or the objection that ended a deal, so a person still has to verify each claim.
How many personas does a team need?
Enough to cover distinct decisions, not distinct job titles. If two profiles would receive the same copy and the same offer, they are one persona.
How often should personas be reviewed?
Quarterly is a workable rhythm, and sooner after a pricing change or a new competitor enters the segment.
