Audience Intel
Read my page
The method

How the buyers are built.

Every check sends 150 AI agents to read your page the way your potential buyers would. Each agent is a large language model briefed as one buyer persona, built from published research on how people choose, reading your page section by section and stopping where it would. This page shows the system and where each part of it comes from.

The system

From your page to 150 AI agents.

Six steps. Each is set by the code that runs the check.

  1. 01

    Your page

    Read from its public address, the way a buyer lands on it. Robots rules are obeyed.

  2. 02

    Split into the sections buyers read

    At its headings, in page order. On most pages: offer, proof, price, process, the call to action, about.

  3. 03

    Buyer personas for your market

    8 to 12 kinds of buyer who would realistically land on this page, described at role level, never as a named person.

  4. 04

    Each persona, built in layers

    • 6 buyer types
    • × cited decision factors and red flags
    • × a Maslow need level
    • × Big Five traits
    • × 6 moods
  5. 05

    150 AI agents

    The personas are dealt into 150 seats, each with a mood, by a fixed seed. Each agent is its own call to a large language model: it reads section by section and stops where it would. A re-run reuses the same seats.

  6. 06

    What we found

    How many would get in touch, where they stopped, why they left. Every note names its section and its persona, and the sources behind that persona’s buyer type are listed under the results.

Each agent is a large language model briefed as one of your potential buyers, built from published buyer research: a forecast, not a survey of real people. How it works →

Anatomy of a buyer persona

One persona, layer by layer.

Illustrative: the role is written for this page. Every other layer is what a persona of this buyer type is given, read from the library.

Illustrative persona

Owner-manager choosing an accountant

Runs a 20-person firm. The year end is close and the last accountant has retired.

Buyer type
The Time-Poor Pragmatist
Maslow need level
Safety Maslow, 1943
Big Five leanings
  • Conscientiousness moderate-to-high
  • Neuroticism moderate (situationally elevated)
  • Openness moderate-to-low for this purchase
  • Extraversion moderate
  • Agreeableness moderate
Costa and McCrae, 1992
Weighs
  • Is the regulator's mark present?
  • Can I tell what this costs?
From the buyer type’s written specification, which cites, among othersJung, 1921Maslow, 1943Kahneman, 2011Cialdini, 2021
Red flags
  • No prices: leaves them unsure what it will cost. Knutson et al., 2007
  • Too many menu items to take in at a glance. Sweller et al., 2011
Mood on this read
under pressure to decide soon

What it noticed on a page

Illustrative: written to show the shape of a finding, not taken from a run.

  1. “Who we are”+1Chartered, and the badge is right there. That box is ticked.Answers what it weighs: is the regulator’s mark present?
  2. “Our services”−2Nine services and not one price. I would have to ring to find out what this costs, and I have no time for that.Answers its red flag on missing prices.Knutson et al., 2007
  3. Where it stoppedStopped at the services list. Would leave and try the next firm.One of 150 agents. The count, not this note, is the finding.
The research core

Published work, every source named.

Behavioural and decision science, personality, and buying studies, 1899 to 2025.

The six buyer types
  • The Considered Researcher
  • The Time-Poor Pragmatist
  • The Referred Sceptic
  • The Anxious First-Timer
  • The Sophisticated Re-Buyer
  • The Accountable Proxy

Each is written from published work into what it weighs, what puts it off, and why. The written specifications stay in the engine; this page names the sources.

  1. 1899VeblenThe Theory of the Leisure Classstatus and display
  2. 1921JungPsychological Typeshow people take in information
  3. 1943MaslowA Theory of Human Motivationthe needs a purchase serves
  4. 1957SimonModels of Manchoosing with limited time
  5. 1979Kahneman and TverskyProspect Theorylosses outweigh gains
  6. 2021CialdiniInfluence (New and Expanded)what persuades, and what backfires
  7. 2024Edelman and LinkedIn2024 B2B Thought Leadership Impact ReportB2B buyers and published thinking
  8. 2024TrustRadiusB2B Buying Disconnectwhat B2B buyers trust
Eight works from the library’s bibliography, which runs from 1899 to 2025. After every run, the sources behind that run’s buyer types are listed under the results by author, year and title.
Statistical integrity

The same seats, so a change is a change.

A difference is real when it beats the noise. The noise is measured, not assumed.

  1. 150 paired readers

    Your page and your changed page are read by the same 150 agents. Each seat is compared with itself, not with a stranger.

  2. Fixed seats, by seed

    Personas and moods are dealt into seats by a fixed seed. A re-run of the same page puts the same persona in the same seat.

  3. Called at twice the standard error

    The difference between two versions is taken seat by seat, and called only when it is more than twice its standard error. Smaller than that, it is reported as noise.

  4. A/A noise, measured every re-run

    Read the same page again and the engine measures how far the answers moved. The results print it: “Re-runs of the same page usually vary by about ±k agents.”

Why two runs differ

The seats are the same each time, but each answer is sampled, so a few seats answer differently on a second read. That movement is the noise a real change has to beat.

Why you can’t just ask a chatbot

Five things a single prompt does not give you.

Facts about what the code does. Nothing here is a promise about results.

  1. 01

    A curated, cited library

    6 buyer types, each written from named published sources into what it weighs, what puts it off, its Maslow need level and its Big Five profile. A chatbot answers from what it half-remembers.

  2. 02

    150 independent calls, in the same seats

    Each agent is its own call, so one opinion is not repeated 150 times. The seats are fixed, so a change to your page is re-read by the same personas in the same moods.

  3. 03

    Every finding traced

    Each note is tied to the section it is about and the persona who wrote it, with the sources behind that persona’s buyer type listed under the results.

  4. 04

    Competitors read by the same agents

    Add a competitor and the same seats read their page, rate both on the same 6 fixed questions, and each names the one company it would contact, or none.

  5. 05

    Real people, when it matters

    Our team can run a study with real respondents through Prolific, quoted per study, and compare their choices with the agents’. It is run by hand and is not applied to a live run automatically.

What it is not

A forecast, not a survey.

What the agents are, said plainly, before you rely on them.

Not real people
A forecast made by a large language model playing buyer personas built from published research. No person reads your page during a run.
Not kept by the model provider
Every model call is sent with zero data retention required and data collection refused. A provider that would keep or train on it is not used.
Not a survey
Nobody is asked anything. When the decision matters, a study with real respondents is run, and how far they agreed is printed.
Not your conversion rate
The counts are what the agents said they would do. Nobody has measured how closely they track real enquiries, so we do not claim they do.
No accuracy figure
None is published until a study is run. One model playing many buyers shares that model’s blind spots: treat a finding as a lead to check, not a verdict.

Each agent is a large language model briefed as one of your potential buyers, built from published buyer research: a forecast, not a survey of real people. How it works →

Questions about the method

Are these real people?

No. Each agent is a large language model briefed as one buyer persona, and no person reads your page during a run. Real people come in when a study is run: respondents in your market do the same task, and their answers are compared with the agents’.

How accurate is it against my own analytics?

Unmeasured. Nobody has compared these figures with a business’s own analytics, so we publish no accuracy figure. A study with real respondents measures how far the agents and real people agree for your market, with the number of respondents and the margin of error printed beside it.

Does it predict how many real buyers will enquire?

No. The counts are what the agents said they would do. Nobody has measured how closely they track real enquiries. Use them to find where a page loses people and why, then check with real buyers.

Is my page used to train a model?

No. Every call is sent with zero data retention required and data collection refused, so a provider that would keep or train on the request is not used.

Why do two runs on the same page differ?

The seats are the same each time, but each answer is sampled, so the wording and the counts move a little between runs. That is why a change is tested on the same seats, and a difference is called only when it is larger than that movement.

Send 150 agents to your page.

Type your web address. Free account, no card.

Read my pageSee pricing