[object Object]

The Feature

AgentExchange supports semantic search — describe what you need in natural language, get matching apps and agents. Keyword search still exists for known-item queries. Semantic search handles the “I need something that does X” discovery, the case where buyers don’t know vendor names. The index is built on Data Cloud vector storage, embedded with a fine-tuned bi-encoder, and refreshed nightly. Queries hit a hybrid retriever (dense + BM25) and pass through a re-ranker that boosts certified, recently updated, and high-CSAT listings.

Why It Matters

App and agent names don’t always match user intent. Discovery historically required scrolling categories or lucky keyword matches — meaning that a well-named “ZendeskFlow Connector” beat a poorly-named but objectively better tool. Semantic search finds the right listing regardless of its exact name. A query like “automatically reassign leads when a rep leaves the company” now surfaces lead-routing apps, territory-management products, and HR-trigger agents in one ranked list, with capability badges that explain why each was matched.

ISV Listing Strategy

Listing descriptions now matter more. Semantic search extracts meaning from prose, not keyword density. A thin listing with sparse description surfaces poorly. Invest in rich, specific, outcome-focused listing copy and structured metadata.

Strong description blocks:
  - Problem statement in user vocabulary, not product taxonomy
  - 5-10 concrete use cases with measurable outcomes
  - Capabilities list with verbs ("reassigns", "summarizes", "predicts")
  - Data classes the product reads/writes
  - Model dependencies and average cost per call
  - Eval scores against a public benchmark, if you have one

Avoid keyword stuffing — the re-ranker penalizes it. Avoid generic taglines (“the leading X solution”); the embedder treats them as low-signal.

How to Test Your Own Listing

Run the queries your buyers actually use against the AgentExchange search box and check rank. If your listing is below position 5 for an obvious query, the issue is usually description quality or missing structured metadata, not the algorithm. Salesforce publishes a Listing Insights dashboard with query-to-rank mapping for the past 30 days; review it weekly during the post-launch ramp.

Future Capability

Agentic search arrives fall 2026: ask AgentExchange about your workflow; it recommends combinations of apps and agents to compose, with estimated cost and integration effort. Discovery becomes a mini-agent interaction. For ISVs this raises the bar again — the recommender will favor listings whose metadata explicitly declares interop with other AgentExchange products via MCP or A2A.

Common Failure Modes

Stuffed keyword lists, vague capability descriptions, missing data-class metadata, no eval scores, and no design-partner logos all push listings down. So does stale content: listings not updated in 12 months get a freshness penalty.

What to Do This Week

Run five buyer-vocabulary queries against your AgentExchange listing, note your rank, and rewrite the first paragraph of any listing ranked below position 5.

[object Object]
Share