Our sample site’s first 7 days: 333 visitors from ChatGPT

About this case study

• Whose site: an AI-powered consumer web app that we built and launched on 21 September 2026. It’s our own product (the company that runs katman.pro), not a client. We don’t name it or its category on purpose.
• Source: the site’s own analytics database (Cloudflare D1), read-only queries, 21 September 00:00 to 28 September 00:00 UTC.
• Only humans: 528 of 1,546 sessions passed our human filter (the rule is below). The filter is cautious: about three in ten real sessions fail it.
• Also used: server-side bot counts, Search Console, and ChatGPT answers we recorded by hand.

The same numbers are in chapters 2, 10, 11 and 17 of our book.

The result in one table

Day by day

On 28 September, a technical SEO pass took the home page’s mobile Lighthouse performance score from 59 to 89, and after the 29 September audit to 98 (lab measurement).

Where the visitors came from

528 human sessions, 21–27 September 2026.

“Purchase” here means a purchase in the same session; the 72 of 102 accounts and 11 of 17 payers above are counted by the channel of the account’s first visit. Much of “Direct” is probably ChatGPT too: 79% of ChatGPT visitors sent no referrer, and we only recognised them by the utm_source=chatgpt.com tag in the link.

Who they were: visitors from 59 countries, led by the US, France, India, Germany, Nigeria, Spain and Brazil; 71% on mobile. ChatGPT visitors viewed 2.45 pages on average, scrolled through 57% of the pages they opened and stayed 198 seconds. Visitors from search engines viewed 2.33 pages and stayed 258 seconds.

Where they landed

The two situation pages took 195 of 333 visitors (59%). The two honest list guides, which name competitors and don’t put us first, took 24.

Where ChatGPT recommended the site, and where it didn’t

Questions asked in ChatGPT with web search on, 26–28 September 2026 (own measurement). We describe the question types and don’t give the real texts, so the site stays anonymous:

The same question can get a different answer in a different session; one run is a sample, not a rank. Perplexity didn’t name the site for any of these questions.

ChatGPT also used our own words. The site’s one sentence stood in the first paragraph, the meta description and llms.txt, in four parts: what the user brings, what the site does in measurable words, what they get back, and the promise that removes the fear. In two separate sessions on 28 September, the answers repeated all four parts, almost in our own words. An invented sentence with the same shape: “Upload your invoice. We move the line items into your accounting software in 30 seconds and flag the ones that don’t match. Your document is never stored.”

What the bots did

Counted on the server, 21–27 September 2026. Bots don’t run analytics scripts, so browser analytics can’t see them.

Turning visitors into customers

People from ChatGPT arrived with the situation they had already described, usually holding a file they wanted to work on. What we changed:

• A file drop box on the two pages ChatGPT linked to most. 52 of the first 68 generations used the visitor’s own file.
• No repeat questions: the start quiz lost about 40% of people at its first question, so visitors from ChatGPT now skip it.
• The right plan: four people who arrived with a file bought a plan that didn’t accept files, and two cancelled within an hour. Checkout now opens on the plan that matches the visitor’s intent.
• The right payment method: with 71% of visitors on mobile, 7 checkout pages opened on 23 September and 1 was paid. Google Pay was switched on that night.

What didn’t work

• Perplexity: named in 0 of 12 buyer questions; PerplexityBot made 8 requests. It also filed the site as a “template and education site”, because about 110 template pages outnumbered about 10 product pages.
• “Best X” questions: ChatGPT listed big brands every time.
• Bing and Copilot: 0 pages indexed on 24 September despite every submission, and 0 Copilot citations.
• Claude: Claude-User opened pages 122 times, and claude.ai sent 0 clicks. Being read isn’t being clicked.
• PDF books: the first ChatGPT visitor came about 8 hours after the PDFs went up, but that’s a sequence, not a cause; ChatGPT linked the feature pages, not the PDFs.
• Google organic: Search Console’s 28-day view on 28 September showed 32 clicks and 306 impressions, almost all for the brand name.

Limits of this evidence

This is one consumer AI app in one category over one week, and the site was built for the method from day one. It’s our own product, so we have an interest in the result; that’s why the numbers come with their source and filter. AI answers change between sessions, and the human filter undercounts. Read it as a worked example of the Katman method, not as a forecast for your site.

What to copy

1. One page per customer situation, with the answer in the first sentence.
2. One sentence in four parts, identical in the first paragraph, the meta description and llms.txt.
3. Text in plain HTML and AI bots allowed, from day one.
4. ChatGPT counted by utm_source=chatgpt.com, not only by referrer, and bots filtered out.

If you want your AI coding tool to do this with checks at each step, see Katman. The background is in What is GEO?.