A study, a protocol fight, and a browser shutdown all landed on the same question this weekend: what happens to search when most of what it reads was written by a machine. Here is what moved.
The Feedback Loop: AI Search Starts Eating Its Own Output
The most uncomfortable finding of the window is not that AI writes the web. It is that AI search appears to prefer what AI wrote.
- Study: AI Search Engines Reward Synthetic Text, Then Consume It In a controlled 2026 Web Conference experiment, machine-written pages made up two-thirds of the pool but produced more than 80 percent of what the answer engine cited. Accuracy held near 68 to 70 percent the whole time, so no dashboard would have flagged it. Pair your citation count with a source-diversity check before you call it a win.
The Agent Stack: Two Protocols, Only One That Does Anything
The agentic web is splitting into an identity bet and a capability bet, and the industry has piled onto the one with the weaker evidence.
- Llms.txt Spreads by Default While WebMCP Waits for Deliberate Setup Slobodan Manic notes a WordPress plugin on more than three million sites generates llms.txt by default, publishing a self-description nobody reviewed, while Google’s John Mueller has called the file purely speculative. WebMCP actually lets an agent search, price, or book. Neither gets you discovered: that still runs on HTML.
The Surfaces Keep Moving Under You
Two reminders that the surfaces you measure are not stable ground.
- OpenAI Retires ChatGPT Atlas, Merges It Into One Desktop App The standalone AI browser shuts down Aug. 9, with browsing, ChatGPT Work, and Codex folding into a single desktop app. The packaging changed, the stakes did not: a question answered inside ChatGPT is still a query that never reaches a results page. Track the unified app, not Atlas.
- Google Business Profiles Show Empty Review Dashboards, Again Owners are seeing “You have no reviews yet” while the public listing still shows the real count, one case at 916 reviews. It is a separate fault from the confirmed bug earlier this month. If the public count has not moved, the dashboard is wrong, not your review history.
Today’s Quick Hits
- Google Details Timeline For Fixing Canonicalization Duplicates Updated documentation says pages can stay clustered as duplicates for up to two weeks after you fix the content, faster when the difference is clear and significant. Wait the full window before assuming the fix failed.