What ChatGPT-User Fetched From Our Product Pages in One Week

In the week of October 5 to 11, ChatGPT-User fetched pages under our five products 108 times. Cloudflare verified each request as OpenAI's agent, and each one got a 2xx answer. (The count is a Cloudflare estimate from sampled data.)

A fetch is not a visit. An answer-time fetch happens when an agent reads your page to write an answer for a person. The person may never click. The agent reads, summarizes and moves on.

Two counts, two audiences

Search Console counts the people who arrived and the results they saw. For the five calculators' pages it shows 9,217 impressions in the week of September 28 to October 4. That is blended web search, and Google counts AI Overviews and AI Mode inside it. It is also a different week from the fetches, so I will not divide one number by the other.

Fetches count the agents that read you before anyone arrived. A team that reports only impressions and clicks cannot see that second group at all.

Every figure here names its source, its unit and its window. Each count comes from one recorder and one window.

The buyer is an agent now

Jason Lemkin makes the sharpest version of this case in SaaStr's piece on the new "Ask Your Developer". SaaStr runs 21 or more agents in production with three humans. He describes Render, which he says "got a qualified lead with no rep, no ad, and no search result."

He also describes the reverse. An agent moved on to the next name on its list in under 12 hours. The same agent also recommended building anyway. His summary: "Whichever product sits in the agent's default set at the moment of the task tends to win."

I would hold this at arm's length. It is one company's account of its own agents, and Lemkin says himself that he is not sure one of their calls was right. It is a set of stories, not a market study.

The consumer side looks slower. Among the numbers in Digiday's weekly briefing: 51 percent of U.S. digital shoppers have not bought through an AI platform like ChatGPT or Google Gemini and are not interested in doing so.

Who writes the default set

If the default set decides who wins, the next question is where a default comes from. Part of the answer is public standards. This week Anthropic said it is donating the Agent Skills standard to the Agentic AI Foundation. Simon Willison noted the move and quoted Anthropic saying the Model Context Protocol went to the same home. A format that agents can share through a neutral foundation is one place a default can start.

The other part is plain: your pages and your API have to be fast and complete when an agent reaches for them. Lemkin makes the same point about slow connectors. Agents route around them.

Meanwhile platform dashboards can change under you. Google is testing a new performance report in Business Profiles, with metrics on engagement, reviews, search trends and audience, according to Search Engine Roundtable. If you rely on a platform's dashboard, your yardstick can move under you. Logs you own do not.

What to measure on Monday

Agencies already feel the pressure from the other side. In the same Digiday briefing, a consultant says clients now look "inside their own organizations" at "what do they want to own". Measurement is a good thing to decide you own.

Here is what I would do:

• Count verified fetches by named agent, at the edge. Use your CDN logs, not a vendor's estimate of your rank, and say when a count is sampled. Keep answered requests (2xx) apart from probes and scanners that borrow an agent's name.
• Put visits next to fetches. The gap between them is the part of your audience that never shows up in analytics.
• Label coverage. Write down when each recorder went live and say so beside every count.
• Keep search clicks in the report, but call them what they are. They count people who arrived, not agents that read.

One claim in this piece is ours and not in any of the sources. The unit that tells you whether you are on an agent's shortlist is the verified fetch by a named agent, counted at the edge, with the visit count beside it as the check. A page can be read a hundred times and never produce a click.

We run this measurement on our own product pages. A team that wants it for its own domains can license the AI visibility report. The owner-and-signals data API and the content engine are also available to license. Terms are quoted on a call. You can send a licensing inquiry here.