An AI assistant meets your content three times: when its training crawler collects it, when its search index refreshes it, and when it reads a page live while answering someone. Genymotion's AI site is fetched at all three stages, every day. ChatGPT alone opened it in 8,313 conversations in 90 days, most often to check a price. When the assistant cites a source, it cites genymotion.com.
A citation is the visible part. Behind it, an assistant meets your content three times. Its training crawler collects pages months before a model ships. Its search index refetches them every few days. And when someone asks, it opens a page live and reads it on a budget of a few seconds and a few thousand words. The same page serves all three, so it has to be a page a machine can finish: one answer per question, current prices, stated limits, and the questions phrased the way real visitors ask them. What the assistant reads is what it repeats.
Genymotion makes an Android emulator that developers, QA teams and enterprises use to test apps. The facts a buyer needs, prices, system requirements, what the free plan includes, which CI tools it plugs into, were all published. They were spread across www, docs and support, behind a JavaScript front end.
The content was not the issue. The packaging was. The main site loads 69 scripts and renders 3,249 DOM nodes, and an assistant fetching it spends most of its reading budget on framework overhead instead of on the answer.
Rozz built an AI site at rozz.genymotion.com: the same content as the main site, repackaged for AI agents. 307 clean pages and 265 Q&A pages, each with Schema.org markup, organised under 15 topics with an llms.txt file so crawlers can find their way. It renders in under 100 ms, on two scripts and 61 DOM nodes.
The Q&A pages come from real questions. Genymotion's on-site chatbot handled 3,225 sessions and 6,420 questions in the same 90 days, and the ones worth publishing become pages. When Genymotion changed its prices in June 2026, the 14 Q&As that quoted a price were re-answered and republished.
Genymotion's side of the setup was two DNS records. We build and run the rest, refreshing the content every week as new questions come in.
We log every fetch of the AI site by a recognised AI crawler. Between June 23 and September 20, 2026, the crawlers of every major assistant came back daily, and between them they fetched every page on the site.
| Stage | Crawler | Distinct URLs fetched | Days present |
|---|---|---|---|
| Collected for training | GPTBot (OpenAI) | 699 | 89 of 89 |
| Collected for training | ClaudeBot (Anthropic) | 1,023 | 89 of 89 |
| Collected for training | Meta AI, ByteSpider, Common Crawl | 686 to 1,377 each | 14 to 89 |
| Indexed for search | OAI-SearchBot (ChatGPT search) | 860 | 89 of 89 |
| Indexed for search | PerplexityBot | 773 | 31 of 89 |
| Indexed for search | Googlebot | 13,112 crawls forwarded to genymotion.com since May | daily |
| Read live while answering | ChatGPT-User, Claude-User | see § 04 | 89 of 89 |
The site has 307 pages and 265 Q&A pages. GPTBot's 699 distinct URLs cover the whole site; ClaudeBot's 1,023 includes topic hubs, the API files and pages since retired. Nothing was submitted, pinged or tuned per platform: an llms.txt file and a sitemap, and the crawlers found the rest.
The training stage is the one you cannot buy later. Every model release trains on what was published before its cutoff. A site that is fetched daily by GPTBot and ClaudeBot is in the material the next version learns from; one that is not stays invisible to it.
The third stage is the one that touches a buyer directly. We cluster live fetches into conversations, so one answer that reads three pages counts once, and we exclude scanners that spoof an assistant's user agent.
| Assistant | Conversations that opened the AI site | Trend |
|---|---|---|
| ChatGPT | 8,313 | about 92 a day, steady |
| Claude | 260 | 71 in July, 86 in August, 97 in the first three weeks of September |
One in seven of those conversations, 1,190 of 8,313, was about pricing or licensing. Next come setup and configuration, and security testing. Buyers ask the assistant what you cost before they ask you.
Where Genymotion stands in the answers. On a fixed panel of 160 buyer questions asked over the same 90 days, with no brand name in the question, ChatGPT with web search cited genymotion.com in 85 answers, 53%, and named the brand in 57.5%. Gemini cited it in 15, 9%.
That is a position, not a lift. Genymotion's AI site went live before we began tracking citations, so there is no before-and-after to report, and we do not claim one. What the data shows is where the assistant reads, and where it sends people.
It works because the content is built from real user questions, the actual problems people bring to the chatbot. A marketing page makes one pitch to everyone. A library of real questions lets the assistant find the answer that fits the person in front of it. Someone on macOS, on a small team, with a security requirement needs a different answer than someone wiring up a CI pipeline at scale, and the AI site has a specific page for each. Pricing for their team size. Compatibility with their setup. The constraint they are stuck on right now.
So the assistant is not matching a buyer to a brochure. It is matching a product to a person, on the details that decide the sale, and it is quoting the price you published this week, not the one it remembers from training.
A second site sounds like split authority. The data says otherwise. Of every citation for Genymotion, 94.8% link to genymotion.com or its docs and support subdomains. The AI site itself almost never appears as a source. The assistant reads the structured content, then credits the original.
Humans who reach an AI site page in a browser are forwarded to the matching page on genymotion.com, so the AI site never holds a visitor either.
Two DNS records on your side. We build and run the rest.