What does a product launch look like from the other side of the web? For the businesses whose sites we monitor, Muse looked like this: nothing on the day, and then, the morning after, a crawler that had been visiting a few hundred times a day began visiting thousands of times a day, and kept climbing for a week.
The numbers
Across 25 monitored sites, Meta-ExternalAgent averaged 725 hits a day between 24 August and 8 September, the day Muse launched. On 9 September it recorded 3,125. On 10 September, 6,355. On 11 September, 13,860. After a weekend dip it climbed again, to 18,234 on 15 September and a peak of 20,177 on 16 September, roughly twenty-eight times the pre-launch average. On 17 September it was falling, and on 18 September it was back to 437. Over the nine days from 9 to 17 September the crawler made 104,058 requests, which is 44 percent of every bot hit the fleet saw in the entire 31-day window of the first Index edition.
The chart on the edition page draws it day by day. The whole shape of the month's bot traffic is that one crawler's surge; the AI category as a whole moved with it, and everything else stayed flat.
What it was
Meta documents Meta-ExternalAgent as the crawler it uses to index web content and gather training data for its AI products. It reads pages. It does not call, chat, book or submit anything. In the vocabulary of the Index it is the reading layer, the same layer as a search engine's crawler or an assistant fetching a page because someone asked a question.
The distinction matters because the cost profile is different. A reading-layer surge is measured in bandwidth, cache misses and, for the smallest sites, a hosting bill. It never rings a phone. An acting-layer surge, an assistant calling five clinics for one patient, is measured in staff minutes. One is a hosting question and the other is a staffing question, and a count that mixes them tells a business nothing it can act on.
What it was not
Two things the surge was not are worth stating plainly. It was not Muse itself browsing. Meta publishes a separate user agent for on-demand fetches, the kind an assistant makes when a person asks it to look at a page right now, and that pattern recorded no hits on any monitored site in the window. This is consistent with Amazon's public complaint that Muse browses without identifying itself; whatever Muse fetched, it did not fetch as Meta's fetcher.
And it was not, on the evidence, caused by the launch. The timing is as tight as timing gets: baseline until the launch day, a sharp rise the day after, a peak eight days later, baseline again two days after that. Meta has not said what the crawler was doing. The edition records the correlation as strong and the cause as plausible, and it stops there, because a number that is published and signed should not carry a guess.
What the sites did
Nothing, which is the right answer for a reading-layer event. Every monitored site allows Meta's crawler in its robots file, along with the other AI crawlers and fetchers a business should expect, and none of them changed that during the surge. The crawler that read the site in September is the same one that will let an assistant describe the business accurately in October, and being described accurately is the beginning of every inquiry that turns into a customer.
Whether to keep allowing it is a decision each business gets to make, and the point of counting is that it can now make it on its own numbers rather than on a feeling. The channels page describes what the web instrument reads, and the layers page explains why a crawler and an agent are never added into the same total.