AEM Sites Playbook, Chapter 13: Site search, relevance and content discovery
Can customers find the right results in the right locale and access context?
November 10, 2026
Specify what a useful result looks like
Use this chapter when customers cannot find content, locale results are wrong or a search implementation is being selected. Start with representative customer queries and the task each should support. Include synonyms used by customers, product/service names, incomplete queries, no-result cases and content that must not appear.
Distinguish author-console search, a repository query and the public site's search service. They can share underlying content but have different users, access requirements and relevance goals. A fast repository query is not proof that the customer sees useful results.
Business owners nominate important queries and expected outcomes. Content owners identify missing terminology, duplicate pages and outdated material. Engineers and search specialists propose the retrieval, ranking, filtering and freshness implementation. Keep a shared acceptance collection so future content or index changes can be compared with the same cases.
Compare the operating consequences
Assess a repository-backed approach, an external search service or another supported capability against locale handling, access filtering, facets, ranking, pagination, update/delete handling, operational ownership and cost. Confirm how each receives updates and removes withdrawn content. A managed search service can move responsibility; it does not remove the need for a content/index contract.
For multiple brands or languages, include isolation and language-sensitive relevance. In the Cloud Service localized full-text discussion, the author challenged recommendations about index overlap and reported a support-advised selection policy. Treat that as evidence to inspect index selection and platform constraints; it does not validate one per-language configuration for every installation. Use query/indexing practices and managed indexing guidance for the target design.
Adobe also documents a Content AI Search component. Evaluate its current support, entitlement, indexed content, access model and answer quality before choosing it. Its existence does not establish suitability for the customer's corpus or permission-sensitive information. Retain traceable examples of answers/results and a way to report incorrect output.
Test both retrieval and change handling
Run the acceptance collection with the intended visitor identities and locales. Verify relevant content appears, excluded content stays excluded, navigation/pagination remains usable and snippets or highlighted terms do not disclose content the user cannot access. Test empty, long and unusual inputs through the supported interface.
Publish a controlled update, then verify when the result changes. Withdraw the item and check that the chosen search route removes or replaces it as agreed. Repeat with a moved page and an updated asset/reference. Record which system owns each freshness delay before adjusting an AEM cache or index unrelated to that search route.
For the hypothetical campaign launch, include an old product name and its replacement. The content team may need a synonym or redirect/content correction as well as a technical change. Review no-result queries and failed searches with content owners; a better retrieval engine cannot supply information that nobody has authored.
Separate relevance, isolation and freshness
Community discussions about multilingual search include tested suggestions, challenged index-selection assumptions and reported support guidance. The example leaves target-environment validation open. Use a representative query collection across languages and content scopes, rather than one successful keyword, to judge the design.
Scroll horizontally if needed.
| Requirement | Owner must define | Acceptance evidence |
|---|---|---|
| Relevance | Which results help a particular customer task? | Judged representative queries and expected/excluded results |
| Locale/brand scope | Which content belongs in this experience? | Positive and negative cases across boundaries |
| Permission isolation | Which identity may see each result and snippet? | End-to-end identity-sensitive tests |
| Freshness | When must changes and withdrawal appear? | Observed publication/update/delete timing |
| Usability | How should zero results, pagination and filtering behave? | Complete interaction tests on realistic result sets |
| Operation | Who repairs ingestion, ranking or outages? | Alert, reconciliation and support procedures |
Do not merge a relevance failure with a repository-performance diagnosis. A technically correct query can still return the wrong vocabulary, stale content or results users cannot understand.
Worked scenario of multilingual product search
All queries, rankings and outcomes in this scenario are hypothetical. Two brands operate English and French sites. Customers use both current and former product names. Some supporting documents are restricted to partners. The sponsor wants one improved search capability while regional owners require correct language and brand boundaries.
The content and search owners create a small acceptance collection before choosing a service:
Scroll horizontally if needed.
| Query case | Expected behavior | Content that must not appear |
|---|---|---|
| Current product name in French | Appropriate French product content is useful and prominent | Unrelated brand or unintended language results |
| Former product name | Customer can reach the approved replacement | Withdrawn offer presented as current |
| Common task phrase | Useful service/help content appears | Internal operational content |
| Partner-document term as an anonymous visitor | Only permitted public results/snippets | Restricted document title, excerpt or answer |
| No useful match | Clear recovery/navigation option | Fabricated certainty or a broken result page |
The team compares a supported repository-backed design with an external service and, where applicable, a currently documented AI-assisted option. Each candidate receives the same fixtures and permissions requirements. Product support and entitlement are checked separately from result quality; Content AI Search documentation does not validate a particular customer's answers.
Suppose one candidate ranks French results well but exposes a restricted title in an anonymous snippet. Another keeps isolation but fails former-name queries because content/synonym ownership is undefined. The first is held for the mandatory exposure failure; the second remains a candidate with a specific content/relevance gap. Faster response alone cannot decide between them.
Test changes through the ingestion and removal path
Publish a revised product name and observe the intended search route. Then withdraw a test item and inspect result titles, links, snippets, cached answers and pagination as relevant. Distinguish the source content state from the search system's indexed state.
Scroll horizontally if needed.
| Observation | Investigate | Avoid assuming |
|---|---|---|
| New page absent | Source scope, publication, ingestion and query filters | That an AEM cache purge repairs the search index |
| Wrong language | Locale mapping, content and query/analyzer behavior | That a path filter alone supplies linguistic relevance |
| Old item remains | Deletion/withdrawal contract and ingestion history | That a missing source page removes every cached result |
| Restricted excerpt appears | Index exposure, access model and result rendering | That clicking the protected target is the only access check needed |
| Poor result despite correct retrieval | Content quality, ranking and customer terminology | That more indexing capacity solves the task |
For each failure, assign the owner of the next checkpoint. External search needs a data-exchange contract and reconciliation, just as another integration does. Repository-backed search needs supported query/index evidence without overlapping assumptions borrowed from a local trial.
Operate relevance as a content responsibility
Review no-result and unhelpful-query patterns with content owners and the appropriate privacy/data controls. Decide whether the problem needs new content, corrected terminology, a ranking adjustment or a product change. Retain the query collection across releases so improvements can be compared against prior behavior.
The acceptance decision combines usefulness, appropriate exclusion, freshness and support readiness. Record the tested languages, identities and content range, plus explicit gaps. Expanding to another brand or language triggers another representative check rather than inheriting a universal search-pass status.
Retain relevance evidence
Complete W13 with customer queries, expected/excluded results, locale/identity, result quality, update timing and the accountable owner. Business owners accept usefulness, security owners review appropriate exclusion and practitioners preserve the repeatable collection. Re-run it after changes to content structure, indexing, ranking, filters or the search product. Route performance analysis to chapter 12 and external indexing contracts to chapter 14.
W13 Search acceptance collection
Use with chapter 13. Keep an entry for each important query and access/locale scenario.
Scroll horizontally if needed.
| Record | Entry |
|---|---|
| Customer task, query and expected relevant result | [enter] |
| Locale/brand/identity and results that must be excluded | [enter] |
| Search route/service, ranking/filter owner and content source | [enter] |
| Zero-result, pagination, facets and snippet/access behavior | [enter] |
| Publication/update/delete timing and observed result | [enter] |
| Relevance judgment, error evidence and remediation owner | [enter] |
