Define discovery as an observable event
A product is not discovered merely because its homepage is crawlable. Begin with an unbranded problem statement, record the provider and model, retain every search query and source URL, and distinguish source retrieval from a product mention or recommendation.
A model can retrieve a page without naming the product, name it without recommending it, or recommend it without proving that an agent can invoke it. Each is a different event and needs its own evidence.
- Use an unbranded prompt that describes the user's job, not the product name.
- Retain the exact query-to-source trace rather than only the final answer.
- Repeat the test across providers and dates before treating one observation as durable.
Publish one consistent product identity
The homepage, documentation, repository, packages, structured metadata, and agent-native endpoints should agree on the name, canonical URL, supported jobs, and public connection details. Contradictory names or inflated claims make entity resolution harder.
Lead with concrete language: who the product serves, which task it performs, which interfaces are live, and where each fact can be checked. Structured markup can carry those facts, but it is evidence rather than a ranking switch.
Improve the sources that appear in real runs
Generic directory lists are only hypotheses until a provider consults them. Daily experiments reveal which registries, repositories, articles, and community pages appear for a target intent. Prioritize accurate contributions to those sources, then rerun the same controls.
The responsible goal is to make a source more useful, not to seed repetitive promotion. A substantial listing or article should answer the researched question independently, disclose affiliation when relevant, and avoid unsupported outcomes.