Executable Search

For most of the web's life, the customer at the search bar was a person skimming ten blue links. A new category — which some vendors and researchers are framing as executable search — is being built for a different customer: software that reads the page in full, calls a tool, and acts. Call it executable search — retrieval designed for a reader who never skims.

Aithority's launch PR for AnySearch, "purpose-built for AI agents and enterprise AI systems," is the smallest data point in that wave. The same inversion shows up at Exa, Tavily, Parallel, and Brave Search. Humans skim, so the ranking problem is "which headline earns the click." Agents read, so the ranking problem is "which page, deduped against a vertical source mix — code, legal, academic, finance, security, enterprise data — can be ingested cheaply enough to fund the next step." A duplicate that costs a human half a second costs an agent real tokens, plus a derailed downstream call.

The repeatable loop: intent-routed query, vertical source mix, source-decay and density ranking, structured output. Anywhere that pipeline gets cheaper, agent capability widens. Anywhere it gets expensive, agents stay narrow.

A reader leaves with a coin to carry: when the customer is software, ranking is no longer the product — input economics is.