426 Visits in One Day, Then Silence: How Meta's AI Crawler Reads in Batches

Day one, and a visitor who was not browsing

km.idaeo.ai's self-built observation instrument went live at 2026-08-10 07:52 UTC. On its very first day, a visitor with unusual behavior showed up: Meta's AI crawler, under the UA name meta-externalagent.

Why build your own instrument? Because ordinary traffic reports cannot show you what AI crawlers read. They do not click ads, fill forms, or leave comments; they come and go. Only logs at the server and edge layers tell you who is reading your content, and how.

This visitor was decisive. It visited 426 times, all concentrated in a single day. The last hit landed at 14:10 UTC, and then it went quiet. Like a customer who walks into your shop on opening day, photographs the entire store — menu, walls and all — and leaves. At least within the window we can see, it has not come back. That shape — come once, sweep everything, disappear — is exactly what this article profiles.

What it read: news fact cards, plus the author page

Its reading list, top entries: homepage, 29 times; the news fact card insight/ank-2026-07-28-003, 28 times; ank-2026-07-16-001 and ank-2026-07-16-003, 25 times each; the /insight list page, 24 times; the /tag/ank tag page, 24 times; favicon, 14 times; the author page (Takenouchi Rin), 10 times.

A term first: a news fact card is a content format on this site that breaks one news item into a structured card with sources and timestamps — one card states one fact clearly. For machines, the format is easy to digest: clear boundaries, checkable sources, no need to dig the point out of a long article.

Two signals hide in this list. First, its staple diet is exactly these fact cards, not random page-flipping. List and tag pages also rank high, suggesting it read the catalog first, then fetched cards one by one — the path of a planned purchase, not window shopping. Second, it made a point of visiting the author page. Like a publishing editor screening manuscripts: judge the piece first, then go back and check who wrote it and what else they wrote. Reading “content plus author” together is a typical outline of training-corpus curation — picking material, and noting where the material came from.

Almost every time, it walked into the kitchen

Now look at how it used the cache. In the CF edge-cache window starting 08:00Z on 08-10, the distribution was: dynamic 359, miss 46, expired 27, and hit only 1.

In plain terms: a hit means “take the ready-made copy off the shelf”; dynamic, miss, and expired all mean “ask the kitchen to cook it fresh” — the request really went back to the origin server and read the content in full. Almost everything went back to origin, which means it was not passing by to say hello; it was seriously carrying things home.

This has practical meaning for site owners. The cost of cache-hit traffic is mostly absorbed by the edge layer; origin-bound traffic is what consumes your origin server's compute and bandwidth. In other words, this style of crawling leaves real load, not splash. Small sites especially should know: the weight a batch crawler's single-day burst puts on your origin, and the weight it puts on the cache, are two different things.

Compared with other crawlers, a different animal

A previous article recorded this site's cumulative ledger: ClaudeBot crawled 6.44M times and GPTBot 5.62M times (source: https://km.idaeo.ai/ai/ai-crawler-marketing). That pattern is a slow, steady stream — coming every day, turning page after page.

There is also a control group on tempo. The same article measured: after a new page goes live, ClaudeBot's median first crawl arrives within 4.2 hours; OAI-SearchBot's median takes 332.8 hours, which is 13.9 days. One is the neighbor around the corner; the other is a distant relative.

What Meta showed this time is a third rhythm: not daily, not slow, but a single-day burst sweep followed by silence. Like a wholesaler stocking up: haul away the whole batch at once, with no retail-style daily rounds. This shape usually corresponds to a “batch collection” way of working — a scheduled job starts, sweeps, and clocks out — rather than the real-time lookup where a user asks a question and the machine fetches one page on the spot. We cannot verify from outside what the data is used for; we can only record honestly: the shape looks like stocking up, not price-checking.

A necessary reminder: the cumulative numbers above come from an older instrument with a much longer observation window, while Meta's 426 visits come from the new instrument's first day. Different instruments, different windows — the magnitudes are not directly comparable. This section compares the shape of behavior, not its size.

The boundaries of this profile, stated up front

• Single-site observation: only km.idaeo.ai. On other sites, Meta may behave entirely differently.
• The instrument has run for only 1.5 days: the sample is very young. “It stopped afterwards” only means nothing more was seen within the window, not that it will not return.
• The cache-distribution window starts at 08:00Z on 08-10 and does not fully overlap the visit-count window; the two sets of numbers cannot be added to or subtracted from each other.
• Identification relies on the UA: we identify by the UA name meta-externalagent, with no reverse-IP verification. UAs can be forged.

If you see this pattern on your own site

• A single-day burst followed by silence is most likely batch collection, not continuous surveillance. Do not rush to conclude you are being targeted.
• What it reads is more informative than how many times it comes. Content pages plus the author page is a curation signal; homepage and favicon only is usually just scouting.
• If you want to state a position, your robots.txt is the first stop. Work out what it has read from you before deciding to welcome or decline.
• If you want evidence, install your instrument early. Crawler behavior evidence evaporates fast: had we gone live one day later, these 426 visits would not exist in any record.