Skip to content

Build & Tune an AI Twin

A twin is built from reports: behavioural, psychographic, and contextual data drawn from signals in the platform’s data layer. Tuning it means shaping those reports before they are locked in, so the twin’s responses reflect the real nuances of the audience rather than a generic average.

For the definitions and underlying concepts, see Concepts → AI Twins.


Step 1: Choose “Segments” not “Respondents”

Section titled “Step 1: Choose “Segments” not “Respondents””

When you reach the build-path selection screen after defining a segment, you will see two options:

  • Segments: proceeds to the twin-building flow where you can review and adjust mindset metrics, inputs, date range, and exclusions before the twin is built.
  • Respondents: builds respondents directly, bypassing the detailed tuning step.

Choose Segments if you want full control over the twin. Respondents are built from twins anyway; choosing the Segments path means you shape the twin first, which in turn shapes all respondents derived from it.


After the system generates an initial brief from your segment definition (this takes 30–40 seconds), it proposes values for five mindset metrics. These are percentage sliders that tell the system what kind of person you are building:

MetricWhat it controls
Brand attachmentHow loyal the person is to a specific brand versus open to switching
Emotional affinityHow emotionally driven their category decisions are
Purchase intentHow likely they are to buy in the near term
Category awarenessHow informed and active they are in the category
SkepticismHow much they doubt brand claims and resist marketing

The system proposes values based on your segment definition. Review each one critically. The right values depend on who you are trying to represent.

Examples from the GEICO auto insurance case:

  • A loyalist segment (someone who has been happy with GEICO for years) → brand attachment near 100%, high purchase intent, moderate skepticism.
  • A churned prospect (someone who left GEICO for a cheaper option) → brand attachment near 5–10% (they will pick a competitor), high category awareness (still shopping actively), high skepticism.
  • A digital-first young driver who shops mainly on price → brand attachment moderate (has some preference but cost overrides it), high emotional affinity (values status and peer perception), high category awareness (constantly comparing rates), high purchase intent (legally required to have insurance).

Do not just accept the system defaults. Think through the person you are building and adjust accordingly.


After the mindset metrics are set, the system generates a set of inputs: the data signals it will use to build the twin’s reports (search behaviour, social signals, content affinity, etc.).

You can review and adjust these using three tools:

A text field where you can add, remove, or reframe the signals the system should use. For example, if the system picked up “cheap car insurance” and “online car insurance” but you want to add or remove a specific behaviour pattern, you edit it here.

Sets the recency window for the data signals. How recent do you want the audience behaviour to be? A shorter window gives you a more current picture; a longer window gives you a more stable, trend-level picture.

Removes signals that would contaminate the twin. For example, if someone searches for auto insurance, they are likely to start seeing signals related to all types of insurance (home, life, health). The exclusions field lets you remove those adjacent categories so the twin reflects only the relevant behaviour.

Use these tools when you notice a specific gap or contamination in the proposed inputs.


Once you confirm the inputs, the platform runs the full pipeline: brief → inputs → reports → AI twin. This takes approximately 15 minutes.

When the build is complete, the twin appears in your Converse workspace. You can start a conversation immediately or run qualitative research with it.