Google's own tooling now audits a file that does nothing for a coach, and a controlled experiment finally puts a number on the disappearing click.

The Lighthouse changelog line is verbatim `ard-schema: add Agent Resource Discovery audit`, with a companion entry adding a `.well-known ai-catalog` test fixture. The audit looks for an `Agentmap:` line in robots.txt, a link tag or HTTP Link header with an `ai-catalog` relation, and falls back to `/.well-known/ai-catalog.json`. The Agentic Resource Discovery spec itself is at v0.91, dated 26 August 2026, status "Proposal", authored by Junjie Bu of Google, R.V. Guha of Microsoft and Shaun Smith of Hugging Face. Its scope is stated plainly: "LLMs increasingly rely on external capabilities, MCP tools, A2A agents, skills, and other callable services." It describes callable capabilities. It does not describe web pages, and it contains no language at all about whether an assistant should cite or recommend a business.

Checked against the live repository this morning: ten remote branches, none named for perplexity, sonar or agent; `main` still at `20b1172` dated 22 August; `src/adapters.ts` line 14 still reads `perplexity: "sonar"` and line 91 still calls `https://api.perplexity.ai/chat/completions`. The migration doc is unchanged, still "Sonar will be supported until September 27, 2026", mapping `sonar` to the `fast` preset on a differently shaped endpoint at `https://api.perplexity.ai/v1/agent`. The changelog has not moved.

One correction worth recording. A web search this morning returned a confident claim that sonar moves on 25 September while sonar-pro and sonar-reasoning-pro stop on the 27th. No primary source says that. The Perplexity forum announcement of 13 August says "Sonar endpoints are fully available today and retire on September 27, in 45 days", the migration doc gives one date for all variants, and the third-party page that the search cited says "Sonar tiers retire on the same date." The staged-shutdown rumour in yesterday's watchlist is weaker today, not stronger.

A preregistered field experiment with a browser extension on real Google users, recruited between 17 and 19 March 2026, N=1,100 who made at least one search in the 7-day period, with three conditions: "(1) No AI Search, which hides AI Overviews and AI Mode; (2) Current Search, which makes no changes; and (3) AI Mode Search, which redirects all searches to AI Mode." The results: "Assignment to AI Mode Search significantly reduced click-through rate by -18.8 percentage points [95% CI: -22.2pp, -15.3pp; p<0.001]". Trust fell too: "significantly decreased trust in information on Google by -0.34 points on a 7-point scale". So did perceived usefulness, at -0.59 standard deviations, and satisfaction, at -0.73.