Most teams enable the widget, watch the question count rise, and stop. The useful number is the gap between what users ask and what the docs already say.

What does the Kapa AI dashboard actually show?

Four views matter for support work.

• Question volume over time, split by answered and unresolved
• Top questions, grouped by intent
• Answers with no citation, which usually means a knowledge gap
• Thumbs-down ratings with the conversation attached

Unresolved questions are the gold. Each one is a documentation task you did not know you had.

How do you turn unresolved questions into doc fixes?

Export the unresolved list weekly. Sort by frequency, then by how much revenue sits behind the asking users. A question asked 40 times by enterprise accounts outranks one asked 200 times by trial users.

For each top question, either add a section to the indexed page or, better, rewrite the existing section so the answer appears in the first 100 words. Kapa’s retrieval favours pages where the answer is stated early and in plain language.

How long does it take to see ticket movement?

Realistically six to ten weeks. The first two weeks are noisy because the assistant is still being tuned and users are testing it. Weeks three to six show the deflection curve starting to bend as you fix the top unresolved questions. If nothing moves by week ten, the problem is usually that the widget sits behind a login wall or a menu nobody opens.

What does the analytics not tell you?

It cannot see questions users gave up on and never typed. It also cannot distinguish a user who got a great answer from one who got a good-enough answer and left. Pair the dashboard with a short in-widget survey, or you will overestimate impact. And since pricing is per answer, a rising question count is a rising bill, which makes the deflection number you are trying to raise also a cost number you are trying to control.

Who should not run this workflow

Teams with fewer than fifty support tickets a month will not get statistically meaningful patterns, and the tuning effort costs more than it saves. In that case, keep the widget for user convenience and skip the weekly analytics ritual.

Key Takeaways

• Unresolved questions with no citation are the most actionable view in the Kapa AI dashboard.
• Prioritise doc fixes by question frequency weighted by account value.
• Expect six to ten weeks before ticket deflection is visible.
• Analytics cannot measure users who never asked, so add an in-widget survey.
• Under fifty tickets a month, skip the analytics workflow and just run the widget.

The pitch is simple: users ask in plain language, Kapa cites the exact doc section instead of hallucinating an answer.

What sources can Kapa AI index?

Kapa connects to a handful of source types and re-crawls them on a schedule you set.

• Public documentation sites by URL crawl
• Notion workspaces
• Zendesk and Intercom help centres
• Slack community archives, if you grant access

A typical docs crawl of 500 pages finishes indexing in under an hour. Expect to re-index after every major docs release, since stale answers are worse than no answers.

How do you embed the assistant?

Two options. Drop a script tag before the closing body element to get a floating bubble on your own docs, or use the hosted Q&A page Kapa generates for you. The widget version needs a single project ID.

<script src="https://widget.kapa.ai/kapa-widget.js" data-project-id="YOUR_ID"></script>

Styling is limited to colour, position, and the welcome message. If your design system demands more, plan for a custom build against their API.

Is Kapa AI free?

There is a free demo tier for evaluation, but production use is paid and priced per monthly answer rather than per seat. That structure rewards products with low volume and punishes viral ones, because a spike in questions is a spike in cost. Get a quote before you ship the widget to a large audience.

Where does Kapa AI struggle?

It answers only what is in the indexed corpus. Ask about a feature you shipped yesterday and it will either say it does not know or pull a nearby, wrong section. Multi-hop questions, like “which of the three plans includes SSO and how do I enable it”, often need two prompts. And if your docs are contradictory, Kapa surfaces both versions without flagging the conflict.

The right fit is a product with well-maintained docs and a support inbox full of the same twenty questions. The wrong fit is a product whose docs live in engineers’ heads.

Key Takeaways

• Kapa AI answers from your indexed docs and links to the source.
• Supported sources include URL crawls, Notion, Zendesk, and Intercom.
• The widget embeds with a single script tag and a project ID.
• Pricing is per answer, so high-traffic sites should model cost before launch.
• It cannot answer beyond its corpus and struggles with multi-step questions.