Quality > Knowledge
What the assistant lacked to answer shoppers, from the AI judge's labels. Content gap share, answer rate, the five gap types with where to fix each, the topics behind them, and who can close capability gaps.
The Knowledge tab of the Quality report shows what your assistant couldn't answer, and where you can fix it. Every figure counts the labels the AI judge wrote (see Conversation labeling). Open Reports, then Quality, then Knowledge, or go to /dashboard/sites/{siteId}/reports/quality/knowledge.
Beta
The tab is marked Beta. The labels are judged by a model and their validation isn't finished. Figures are ranked by how many conversations hold a gap, nothing more: no revenue, impact or priority is estimated.
The tab has the same header, filter bar, coverage line, 10,000 cap and 30-conversation minimum as Quality > Conversations. Read that section first. The short version:
- The coverage line says how many labeled conversations the figures rest on, out of how many match your filters.
- A page reads at most the 10,000 most recent labeled conversations and says so when it stops there.
- A figure whose own count is under 30 conversations shows no percentage.
- With no labeled conversation, the page shows "No labeled conversation yet".
The two tiles
- Conversations with a content gap: the share of labeled conversations where the assistant lacked a product, information, a policy or a capability. It shows as "38 of 412" under the percentage. Under 30 labeled conversations it shows no percentage.
- Answer rate: conversations where the assistant answered, out of those where it answered or didn't. A conversation with no real need is in neither count. The sample is the number of conversations judged.
With Compare to previous period on, each tile shows its change in percentage points, never colored good or bad.
What the assistant lacked
A table ranks the five gap types, most frequent first. Each row shows the conversation count, its share of the labeled conversations and a bar. A conversation can hold several gaps, so the shares don't add up to 100%.
| Gap type | Where to fix |
|---|---|
| Product not in catalog | Connectors |
| Missing product info | Connectors, Pages |
| Inconsistent product info | Connectors, Pages |
| Missing policy or FAQ | Pages |
| Missing capability | Tools |
Each gap type opens the Conversations list filtered on the matching Issues value (the gap type), with your period and filters carried over.
Where to fix links
The Where to fix column links to the Connectors, Pages and Tools sections of the site. A link shows only if you may open that section:
- Connectors needs the permission to view connectors.
- Pages needs the permission to view site pages.
- Tools needs the permission to view agents, the same as in the sidebar.
Without the permission, the link is left out, not greyed. A row can have no link at all.
Under 30 labeled conversations the table is replaced by "Not enough labeled conversations yet".
Topics behind the gaps
Three bar charts count conversations by topic, most frequent first. A topic nobody held isn't listed.
- Product info missing about: size or fit, specs or performance, materials, compatibility, use or care, variants, other.
- Policy missing about: delivery, returns, payment, warranty, promotions or loyalty, store or company, other.
- Capability needed: product search, compatibility check, size advice, product comparison, stock check, order tracking, order change, warranty claim, account or loyalty, site help, coupon check, cart or checkout, handoff to a person, other.
A topic can't be opened as a conversation list, because the Conversations list has no filter for it. Each chart shows an empty state when no conversation holds that gap, and "Not enough labeled conversations yet" under 30 labeled conversations.
Who can close the capability gaps
Among the conversations that needed a capability the assistant didn't have, two counts:
- Enable a tool on your assistant: the capability exists, but the agent version that answered didn't have it switched on. If you can view agents, an Open Tools link goes to the site's Tools.
- Not available in AISA yet: no tool provides the capability today. Nothing to change on your side, and no link.
If no labeled conversation needed a missing capability, the card reads "No missing capability". Unlike the tables and charts above, these two counts have no 30-conversation minimum: they are counts, not percentages.
What it doesn't tell you
- Only labeled conversations. A conversation not yet labeled, or one whose shopper objected to measurement, isn't in any figure.
- Counts of judge labels. A gap type says the judge found a gap in the conversation. Open the conversations behind it before changing a connector or a page.
- No cost or revenue. The ranking is by frequency alone.
Quality > Conversations
How labeled conversations ended and which quality criteria failed, from the AI judge's labels. The resolution rate, the resolution ring, trust failures, pass rate by criterion, and the minimum sample behind every percentage.
Measure what the assistant adds with a holdout test
Serve half of a site's sessions no assistant, then compare how often each group adds to the cart. How to pick a duration, start, extend and stop a test, what a control session sees, and how to read the verdict.