What is enterprise search?
Enterprise search is the practice of finding material across the systems an organisation already operates. It returns a ranked list of documents for a query, leaving the reader to open them and locate what they needed. An answering layer is measured against this category, since it is what the organisation has and what users compare it to.
Why a ranked list of documents fails the question actually asked
Ask a claims handler what they think of the search facility they already have and the answer is usually that they stopped using it. The reason is structural, because a ranked list answers the question of where material might be, when the handler needed to know whether a particular exclusion applies to a particular policy. They then ask a colleague, a route that is faster and leaves no record. Any answering layer inherits that history, so it is judged against a tool users have already given up on.
- A document list transfers the remaining work back to the person asking.
- Relevance ranking is evaluated on clicks, which rewards findability over correctness.
- Users who abandon search ask colleagues instead, and that route has no audit trail.
- Coverage gaps are invisible, since a query returns results whatever is missing.
How enterprise search works
- Connectors enumerate content across the repositories in scope.
- Documents are indexed with their metadata and their permission data attached.
- A query is analysed and matched against the index for candidate documents.
- Candidates are filtered by the asker's entitlements and then ranked for return.
- Click behaviour is recorded and used to tune the ranking over time.
- An answering layer then composes a response from the passages inside those documents.
How enterprise search is evaluated
Enterprise search is assessed on whether the person obtained what they came for rather than on click-through. Examine the following with real users on real tasks:
- share of tasks completed without recourse to a colleague
- position of the correct document in the ranked list
- proportion of the corpus actually reachable through the index
- time from query to the answer the user needed
- abandonment rate after the first page of results
Limitations, risks and common failure modes
- Relevance tuning favours clicks on results rather than tasks completed.
- Repositories outside the index are assumed empty by everyone using the tool.
- Permissions are applied at the document level while previews disclose the content.
- Historic low usage becomes the baseline an answering layer is credited against.
- Duplicate copies of one document occupy the whole first page of results.