Ask users why

When to ask users

Use this page when an analytics tool has already shown what happened, and the remaining question is why those people did it. Funnel drop-off, churn, a stalled activation, a spike in regenerations, and a guide that people dismiss are signals. They become a UserCall investigation when the chart cannot name the cause.

Connect once at https://mcp.usercall.co, sign in, and start with list_studies, then create_study only if none already asks the question. Setup for the connector lives in the MCP docs. Read the page that matches the tool that produced the signal before you write the study.

The order

1. Signal. Record the metric, the time window, and who was affected. Keep the numbers the analytics tool returned.
2. Unknown. Write the question those numbers cannot answer, in one sentence a person can respond to.
3. list_studies, then create_study. Reuse a study that already asks this question. Otherwise call create_study. Put the question in key_research_goal and the signal in business_context.
4. simulate_interview. Omit simulation_id to start. The start returns immediately. Call the same tool again with that id to read the transcript. review_study is optional and does not edit the guide.
5. A person shares the link, or a paused trigger. Hand interview_link to a person. You cannot send it. Or, after list_trigger_events has seen the event, call create_research_trigger. It stays paused. A person opens activation_url. You cannot activate it.
6. Evidence. Poll get_study_status until status is complete. running and analyzing are not findings. Then call get_study_results with format summary. Place the themes next to the original signal.

Hard rules

• There is no share tool. There is no activate tool.
• You cannot send the interview link. You cannot turn a trigger on.
• Never treat an in-progress result as a finding.
• list_studies before create_study. list_trigger_events before any trigger.
• One active agent study. Five simulations per account per UTC day.
• A 402 returns checkout_url. A 409 returns activation_url. Hand that URL to a person.
• Trigger filters are exact property or trait matches only. Counts, sequences, absence, time windows, and not-equals are rejected.

Start with create_study

create_study is the entry. key_research_goal is the question the interviewer must learn. business_context is the product, the signal, the window, and who to talk to. target_interviews is how many conversations you need before you will trust a pattern. duration_minutes defaults to 12.

create_study({
  key_research_goal: "Why did these users leave at this step? Ask what they expected, what stopped them, and what they did instead.",
  business_context: "Name the product, the analytics tool, the metric, the dates, the counts, and the cohort to interview.",
  target_interviews: 8,
  duration_minutes: 12
})

The call returns study_id and interview_link. One active agent study is allowed per account. If credits are short, the tool returns checkout_url. Hand that URL to a person. Do not send interview_link yourself. The next tool is simulate_interview.

After status is complete, get_study_results with format: "summary" returns themes, insights, and risks. Ask for format: "full" only when you need a quote. A summary is the evidence. A transcript is the backup. Empty themes while the study is still running are not a finding.

Investigation pages

• PostHog: a funnel, path, or replay shows the drop and still cannot say why.
• Amplitude: a cohort or retention curve names who left and still cannot say why.
• Pendo: adoption, a guide, or a thin poll shows the stall and still cannot say why.
• AI products: regenerate, thumbs-down, or abandon shows the failure and still cannot say why.

Skill

The investigate-user-behavior skill runs this same loop beside a PostHog, Amplitude, or Pendo MCP. It lives in this repository at skills/investigate-user-behavior/SKILL.md. An agent that can read this page can follow the loop without the skill installed.