Meta's Muse Lifted Cloudflare 10%; Cloudflare Blocks Agents Sept. 15

On Tuesday, September 8, Meta opened its personal AI agent, Muse, to users in the United States.

What separates Muse from other chat products is that it has a browser of its own. It runs inside a separate virtual machine Meta calls Muse Secure VM. Where a service has no open API, the agent visits the website like any user, fills in the form, books the ticket, negotiates the bill.

The product launched with a free tier. The two paid plans cost $20 and $100 a month.

The next day two stocks pulled away from the tape. Meta closed up 6.55% at $653.69, and Cloudflare rose about 10%.

The same day the S&P 500 fell 0.48% and the Nasdaq Composite fell 0.64%.

Why Cloudflare Sits Inside This Story

Cloudflare does not produce content. It is a network company that stands in front of websites, filters every incoming request and passes it to the origin. The answer to "who just arrived" is held there.

The company reported second-quarter results on August 6: revenue up 36% to $696.1 million. Next to that figure sat a threshold.

For the first time, more than half the traffic on its network was not human.

Co-founder and chief executive Matthew Prince had announced the crossing in May: "non-human traffic has now passed human traffic online"

Prince's own guidance had put that crossover in the second half of 2027. It arrived more than a year early.

The market read Muse as a traffic story.

Cloudflare reads the same traffic as an identity problem.

What One Visit Costs in Fetches

Cloudflare has been publishing a figure it calls the crawl-to-refer ratio. It counts how many pages an AI company's crawler pulls from websites, and how many visitors it sends back in return.

Here is the June 2026 measurement as reported.

The gap comes out of how the products work. The Google crawler indexes the page and then sends the user to it; in products that put the answer in place of the page, that second step never happens.

On a site that runs on advertising, this ratio goes straight into the arithmetic.

This arithmetic sits apart from the copyright argument. The right to use content in training is a separate question; what is at stake here is whether the page covers its own costs.

What Actually Changes on September 15

Cloudflare set a date on July 1. From September 15, "mixed-use" crawlers are blocked by default on pages carrying ads.

A mixed-use crawler performs three separate jobs under a single identity: indexing for search, fetching a page at a user's request, and bulk crawling to train a model.

That is the company's objection. The large providers do all three under the same user agent, the one-line identity a crawler gives the server. So a site owner cannot say no to training and yes to agents.

Prince's sentence on July 1 is the rule's stated reason: "we must go further and act faster so that a sustainable ecosystem can emerge"

Where Muse falls matters. In an earlier Cloudflare measurement, 79% of AI crawling was training, 17% search and 3.2% user-initiated. Agents live in that small slice — and that is the slice now growing.

Pinning the Rally on a Single Headline

A caution belongs here.

Two different explanations for Cloudflare's September 9 move are in circulation. One points at Muse: if agent traffic is about to grow, Cloudflare is the network it travels through.

The other points at a security product announced the same day — a service that uses an OpenAI model to find software vulnerabilities and propose fixes.

Even the closing price differs between accounts: $311.95 in one, $314.18 in another.

The sources do not tie causality to one headline. It is equally possible the market priced both announcements on the same day. More: Investor Psychology

The index over the same window is the comparison that matters.

What Is Left

Cloudflare has been building toward this for a year. On July 1, 2025 it launched Pay Per Crawl. When a crawler reached a site that asked to be paid, the server returned HTTP 402, the code that means payment required.

A year later, on July 1, 2026, the model changed. The company dropped per-fetch charging in favour of paying when the content actually appears in an answer.

The reason was a measurement: more than half of bot crawl traffic was re-fetching pages that had not changed. A fetch turns out to be a poor proxy for value.

The first partners are Ceramic.ai and You.com. No pricing was disclosed.

This piece draws on press accounts of Meta's September 8 Muse launch, Cloudflare's July 1, 2026 crawler policy announcement, the company's second-quarter results and the September 9 US close. The crawl-to-refer ratios are third-party reporting of Cloudflare Radar's June 2026 data; the 79%/17%/3.2% split comes from an earlier company publication. Accounts of what drove the September 9 rally diverge, and the closing price differs between sources.

Correction (September 13, 2026): prose tightened for rhythm; figures and sources unchanged.