Fifteen Minutes of the Open Web

For every hour a person spends looking for information online, fifteen minutes are spent on the open web. Cloudflare published that in July 2026, in a report on how bot traffic has changed. It is an easy sentence to read past. Forty-five minutes of that hour now happen somewhere else — inside a summary, inside an assistant, inside a product that read the web on your behalf.

This is the sort of claim that gets repeated until it stops meaning anything, so here are two measurements, both public, both read at the source, and both with their limits stated.

What happens when a machine answers first

Pew Research followed 900 US adults across 68,879 Google searches. On the 12,593 searches that produced an AI summary, people clicked a traditional result in 8% of visits. Without a summary, 15%. Roughly half.

They did not click into the summary instead. Only 1% of visits to a page carrying an AI summary produced a click on a link inside it. And they were more likely to stop altogether: sessions ended on 26% of pages with a summary, against 16% without.

The limit worth stating plainly: the browsing data covers March 2025, and the search results were collected between 7 and 17 April 2025. That is old enough to matter. AI summary coverage has expanded since, which makes the effect more likely understated than overstated — but "likely" is doing real work in that sentence, and I would rather flag it than let it slide past.

The crawl changed its purpose

The second measurement is about who is reading, not who is clicking. As of June 2026, 52% of crawler requests were for AI training, up from 22% in spring 2025. Cloudflare describes pure search crawling as a small and declining share of what bots now do. Some of the most heavily crawled categories saw human traffic fall by as much as 40% in under a year.

Put the two together and the shape is clear. Your pages are still being read. They are being read more than ever. They are increasingly being read by something that will not visit, will not convert, and will not necessarily say where the answer came from.

The answer the market is selling

The response on offer is more content. More pages, faster and cheaper, so that you are present wherever the summary gets assembled.

There is a measurement against that too, and it is the one I handle most carefully. Graphite found that 49.9% of newly published articles in the first quarter of 2026 were AI-generated. That figure travels as "half the internet is AI slop", which is not what it says: it counts articles, not the internet, and detection of machine-written text is itself unreliable, which caps the confidence of any measurement of this kind — including that one.

But the part of the study that almost never travels is the part that matters here. That content largely does not appear in Google or ChatGPT. The authors suggest the share has plateaued precisely because it does not perform.

So the strategy being sold is to add volume to a contracting channel, using a method whose own measurement says it does not surface. That is not a moral complaint about AI writing. It is an arithmetic one.

What I think this actually means

Distinctiveness and discovery used to be separate problems, owned by separate people, funded from separate budgets. I do not think they are separate any more.

If a machine reads your page and produces a summary of it, the question stops being whether someone found you. It becomes whether anything in the page survives compression. Generic material summarises perfectly — that is roughly what makes it generic. A specific claim, a stated method, a number with a denominator you can inspect, a result that failed: those are harder to absorb without attribution, because there is no way to state them without stating where they came from.

I am not offering that as a growth tactic. I am not certain it works, and nobody telling you they are certain has measured it either. It is the position this laboratory is built on, which is a different thing from a proven one.

What I do not know

I do not know whether the Pew effect has grown or shrunk since April 2025, and I have not found a 2026 replication. I do not know how much of that 40% decline in heavily crawled categories is attributable to summaries rather than to everything else that happened in the same year. I do not know whether being cited by an assistant produces anything commercially useful, and I have not seen a credible measurement of it.

What I do know is the two numbers at the top, because I opened both reports and read the method. That is the entire standard here. When either is replaced by better evidence, this page will say so.