A Request Is Not a Visit

DataDome published its report on AI traffic for the second quarter of 2026 last week, and one number has been sitting with me. Its network processed 17.7 billion requests from AI agents during the quarter. Meta accounted for 9.1 billion of those requests and sent almost no real visitors back to the sites that served them.

I have looked at website traffic dashboards for enough years that the word traffic still suggests people to me. Somebody searched for something, followed a link, opened a page, and gave a small amount of their attention in exchange for whatever they found there. We always knew the number was imperfect, of course, but it was close enough to something human that we could use it as a rough measure of interest.

That connection now feels much less certain, and I keep wondering whether traffic has become one of those familiar measurements whose meaning changed while the dashboard stayed exactly the same.

An AI crawler is automated software that retrieves pages for model training, indexing, or a current answer. Referral traffic describes visits that arrive after someone clicks through to the source. We continue to place both activities inside the broad category of web traffic even though one consumes infrastructure and the other may create an audience, a customer, or a relationship.

Traffic no longer reliably means attention. Sometimes it only means that a machine asked the server for another copy of something it had already seen. Cloudflare says more than half of AI crawler traffic is spent fetching pages that have not changed, which makes the word traffic feel even stranger because there may be no discovery, no reader, and no new information involved. If no person arrived, no relationship began, and no new information was needed, what exactly was the traffic number measuring?

The map is not the territory.

Alfred Korzybski

I keep coming back to that line because a traffic dashboard is a map, and for a long time it was a useful one. The map told us where attention probably came from and where it probably went. Now the same mark on the map can represent a person reading an article, a training crawler collecting language, an answer engine looking for a citation, or an agent gathering information before it acts somewhere else. Those are not small variations of the same event. They are different events with different costs, permissions, and possible returns.

I wrote recently that the old web contract no longer describes the traffic. I am not sure I carried that thought far enough. If the contract has changed, then the measurements built around that contract have changed too, even if they still arrive in familiar charts with familiar labels.

This also complicates the growing focus on search engine optimization, answer engine optimization, and generative engine optimization. Visibility inside an answer may be valuable even when it produces no click, but a referral counter cannot tell us whether an article was cited, whether it changed a decision, whether an agent used it to complete a task, or whether the person receiving the answer ever knew where the information came from. We are trying to understand a new form of influence with instruments designed to count visits.

I do not think the answer is simply to charge every bot or block every crawler. Some requests create real value, some are part of legitimate work being done for a person, and some may eventually become more valuable than a page view ever was. The problem is that the server cannot reliably tell the difference, and neither can the publisher looking at the report afterward.

Maybe the next generation of web analytics begins by asking who made the request, what they intended to do, what they were permitted to use, and whether any value moved in either direction. That sounds more complicated than counting visits because it is describing something more complicated than a visit.

A request is not a visit, and the larger the traffic number becomes, the more careful we should be about pretending that it tells us who was actually there.