The Users Who Never Open Your Website

01One hundred and twenty nine people, fourteen visits

One of my products can be used from inside ChatGPT. A person connects it once, then asks the assistant for what they want, and the assistant calls the product's tools and hands back the result. By 13 October, 129 people had connected it, the first on 4 September. The obvious assumption is a familiar funnel: visitors who found the site, liked it and connected the assistant later. The database said otherwise. Only 14 of the 129 had ever used the product's website. For a growing share of users, your product is the answer it returns, not the website it lives on.

What they did instead

They used it, steadily. The main action was called 1,555 times through the assistant. A usage check was called 157 times, a list of available options 136 times and a suggestion tool 114 times, the same planning behaviour I saw from agents in a server for AI agents. The median person sent 580 characters of input that month. Only 6 had reached the free monthly cap. And not one of the 129 had ever seen the paywall on the website, because to see it, you have to be there.

02Every website metric misses them

Analytics counts the 14

Page analytics would put this group at 14 people, and an honest reading of those 14 would say the assistant channel is tiny. The server says 129 people and well over two thousand calls. If a person only meets your product through an assistant, your analytics will never know they exist. It is the same blind spot as machines reading pages without running scripts, the problem in zero click is the end of tourists, except these are not tourists. They are users, and some of them pay. Across the whole product, about 27% of everything generated now comes through the programming interface rather than the website, which includes these people and every automation built on the same keys.

The rules close the usual door

The usual answer is to sell inside the channel. That door is closed by design. OpenAI's rules for apps in ChatGPT forbid upgrade prompts, plan lists and checkout links inside the conversation. An app may say that a feature is not on a person's plan and point to a page that describes the plans, and that is all. When the assistant is the storefront, you are not allowed to put a till in it. I agree with the rule. It protects the person asking. It also means the money has to come from somewhere else.

03Build for the moments outside the chat

The one page they all see

Every one of the 129 passed through a single page on the website: the page where they create the key that connects the assistant. That is the only moment all of them were on the site, and five of the 129 have bought a plan, none of them through a prompt inside the chat. So that page is now treated as the most important page on the site, not as a settings form. It now shows a sample of the paid version's output in the visitor's own language, right where the key is made. If users visit your site once, design that once as if it were your homepage.

Measure by key, not by visit

Count people by the key or connection they use, not by sessions. Count calls by client, split assistant traffic from automation traffic, and record which tools each group calls. Give the free version real value, the strategy in free is a growth strategy, because inside an assistant the free version is the only one people ever touch. And keep the assistant version as good as the website, because a worse version would be the version most of these people judge you by. The quality of what comes back through the assistant is the whole of your marketing to people who never visit.

Tomorrow: what Google documents about how Gemini picks its sources, and a crawler name that never sends a request.