Products

Explore live avatars
How it Works?PricingContact Us

Customer experience3 min read

Measure your AI agent: funnels, friction, drop-off

Launching an agent is the easy part. Six measures that tell you whether it helps, and how to turn each number into a fix.

A laptop screen showing performance analytics graphs
Photo by Luke Chesser on Unsplash
On this page
  1. Six measures that matter
  2. Name your funnel steps
  3. From a number to a fix
  4. Compare fairly
  5. Keep it private
  6. A weekly rhythm
  7. Frequently asked questions

Key takeaways

  • Define the outcome before launch, then track adoption, completion, handoffs and gaps against it.
  • Friction signals such as rage clicks, dead clicks and abandoned forms show where the page itself is the problem.
  • Analytics should respect privacy: record page paths, not what visitors type.

An AI agent that nobody opens, or that opens and does not help, looks identical to a working one until you measure it. The aim of measurement is simple: find the one thing to fix this week.

Six measures that matter

MeasureWhat it tells youWhat to do if it is low or high
AdoptionThe share of visitors who open the agent and who start a callLow: move the entry point, change the invitation, test the page
Funnel completionHow many visitors reach each step you named and where they are lostA step that loses half your visitors is the first thing to rewrite
Handoff rate and reasonHow often a person is needed and whyGroup by reason, then fix the largest group
Lead capture rateShare of conversations that leave contact details (paid plans)Low: ask earlier, or at the moment of value
Unanswered questionsWhat visitors ask that the knowledge base cannot answerWrite the missing document
FrictionRage clicks, dead clicks, idle or abandoned formsFix the page, not the agent

Name your funnel steps

Avarole Insights shows a funnel for the milestones you define with one HTML attribute or one line of code. A typical sign-up funnel might be: viewed pricing, opened the agent, started a call, clicked sign-up, completed sign-up. Read it per agent, over the last 7, 30 or 90 days.

From a number to a fix

  • A step that loses half your visitors is the first thing to rewrite.
  • A page full of dead clicks usually hides a button that does not respond.
  • A form that people abandon is a form with one field too many.
  • A question the agent keeps answering is one your page should answer.

Compare fairly

To learn whether the agent changes behaviour, compare like with like. Use similar pages with and without the agent, or the same pages across two periods that do not include a campaign or a holiday. Run long enough to cover your weekly cycle. Seasonality, ad spend and traffic mix can hide or invent an effect, so change one thing at a time.

Keep it private

  • Only page paths are recorded, never query strings or what visitors type.
  • A random visitor id lives in the visitor's own browser.
  • It stays off when the browser sends Do Not Track or the page opts out.
  • You can switch it off per agent whenever you like.
  • Events are kept for 180 days.

A weekly rhythm

  1. Open Insights and read the funnel for the last 30 days.
  2. Pick the biggest drop and read ten transcripts from that step.
  3. Ship one change: a page edit, a new document or a new action.
  4. Write down the number you expect to move, and check it next week.

Frequently asked questions

Does Insights work in mobile apps?

Yes. The React Native SDK sends the same events from your app, tagged as mobile, so one view covers web and app.

Will analytics slow my pages?

Events travel in small batches and a failed send is dropped, so analytics do not slow down or break a page.

Is visitor data stored?

Only page paths and the actions the agent took, never what visitors type or click. Details are on the Insights page.

See where visitors leave.

Avarole Insights shows the funnel, the exit pages and the friction, with privacy built in.

Keep reading

Related articles

  • A smiling woman waving at a laptop during a video call

    Guides3 min read

    AI avatar vs chatbot: when a face wins

    Where a face-to-face agent earns its place, where a text box is enough, and how to measure the difference on your own site.

    27 October 2026Read

  • A customer support team wearing headsets at their desks

    Customer experience3 min read

    Human handoff: when an AI agent steps aside

    The best AI agents know when to stop. A practical guide to handoff triggers, context and what happens when nobody is online.

    24 September 2026Read