Gaitkit started selling on October 21. In the ten days to October 30, 223 visitors came, 104 of them actually used the site, 20 tried a download, 11 met the prompt for Pro files, and nobody bought. The numbers say the problem is how many people arrive, not the paywall: ChatGPT sent people who used the site, X sent people who mostly left, and bots far outnumbered both.
How we count
The numbers come from our own analytics, which we built to be cookieless. A visitor is a hash that changes every day, so a person who returns on another day counts again; these are visitor-days, not people. The IP address is never stored. The 223 include our own visits from Türkiye while we tested, which is also a country where checkout is not open yet.
On October 29, 25 "visitors" from the US opened the homepage in clusters of three to seven within the same minute and did nothing else: no search, no click on a model, no second page. Earlier visitors who landed on the homepage the same way used it about two times in three. These were automated browsers, so the dashboard now separates engaged visitors, who opened a second page or used one, from everyone else. Over the ten days that is 104 of 223.
Where people came from, and what they did
The useful comparison is not how many people each source sent, but what those people did next. We grouped visitors by where their first page came from:
- X sent 58. Ten used the site. None downloaded anything. Short videos of models walking get views; the people they bring are browsing, not looking for a model.
- ChatGPT sent 49. Forty-four used the site, and ten downloaded a model. They land deep: on the pump jack, the pistol, the slime, a cargo pile, the City Street kit. Every model page ChatGPT linked to was one of the free models. One visitor came back to the same page four times in an hour before downloading; another downloaded the praying mantis and then read the licence.
- Google sent about 20, and twelve used the site. Pages for single models (a runner, a pig, a pistol) started to bring clicks from search in the last few days.
ChatGPT recommending us for free models is good news and a warning. It brings people with a real need, but they are looking for the free file, and so far they take it and go. The fox and the other free models are the front door; nothing yet walks people from the free file to the rest of the kit.
Bot addresses outnumber visitors fifty to one
Over the same ten days, scrapers made 11,876 requests for model files from 11,636 different addresses, sweeping through 586 different files. They rotate addresses, so no single one looks busy; what gives them away is a single ten-minute window in which anonymous requests touch eight or more different files. A person downloading by hand fetches one to four. The download gate refused 11,342 of those requests. The bots count separately, so they do not inflate the download numbers or the funnel.
Our MCP server, which lets AI agents search the catalogue, saw the same split. Directory crawlers and health checkers called it thousands of times. Real agent sessions made 24 tool calls: 21 searches from 9 callers, two requests for a model's code and one for a variant.
What we changed, and what we expect
Two things shipped during these ten days because of these numbers:
- After a free download, a short welcome shows the rest of that model's kit and what one plan unlocks, once per visit, with a one-click "Not now". Since October 28, six visitors outside Türkiye have seen it. Four closed it, and none clicked through yet.
- Search fixes from real queries. Someone searched for a "human boy" seven times and got a grown-up villager, because "boy" was read only as a style setting. Now it also finds the child model.
At this volume, zero sales is not a verdict. A store that converts one or two in a hundred visitors would expect zero to three sales from 223 visits, and a lot of ours were us, bots or people from X who never meant to buy. The next step is more of the right visitors: posts in the communities where people build with three.js and game engines, and more pages that answer the questions ChatGPT and Google are already sending people to us for.
We will publish the next numbers the same way, including if they are still zero.
Pump Jack
Pistol
Fox