Entry #13 · May 5, 2026
Three weeks ago we ended an article with a question. The question asked whether AI-mediated conversations about Genymotion could shift toward enterprise. The question also asked whether the ratio could shift toward enterprise. We made changes to the AI site and we just looked at the data.
The traffic on the chatbot on the website hasn’t shifted. The traffic on the AI site has shifted.
Buyer/professional content reads went from 35.9% to 45.0%. Buyer/professional content reads increased by +9.1pp. The chatbot stayed flat at 7–8% buyer rate. Two channels, two audiences.
Where we left off
In article #11 we shared traffic numbers. On the Genymotion website, the on-site chatbot recorded 667 conversations in March. Fewer than 20 of those conversations showed clear commercial intent. Fewer than 20 of those conversations were under 5%.
The rest of the conversations were people trying to run TikTok on their PC. The rest of the conversations included people trying to set their device language to Chinese. The rest of the conversations included people playing Minecraft on Steam.
The marketing team targeted CI/CD teams, mobile security testers, and cloud-deployment buyers. Those audiences were less than 5% of the conversations the chatbot recorded.
That same article also showed a different pattern on the AI side. A pricing evaluation from Madrid involved eleven ChatGPT-User fetches across five rounds. The pricing evaluation ended on the full pricing breakdown.
A late-night deliberation from the US involved macOS compatibility checks. The macOS compatibility checks came from multiple continents. Those sessions were buyer-pattern sessions. Those buyer-pattern sessions happened in ChatGPT.
The website never recorded those buyer-pattern sessions.
We asked at the end of that article whether the AI site could push the ratio. We asked whether AI platforms could recommend Genymotion to buyers. We asked whether AI platforms could do that recommendation instead of just describing Genymotion to hobbyists.
We can now answer with three more weeks of data.
We compared a 14-day window in mid-March with a 14-day window from April 21 to May 4.
The comparison applied to the chatbot on genymotion.com.
The comparison also applied to the AI site at rozz.genymotion.com.
Chatbot rate: flat
The Rozz chatbot now classifies every conversation with a buying_intent flag.
The buying_intent flag is a yes/no value.
The buying_intent definition is broader than the manual “clear commercial intent” used in March.
Applied consistently before and after, it gives a consistent measure.
| 14-day window | Conversations | Classified | buying_intent='yes' | Rate |
| --- | --- | --- | --- | --- |
| Mar 11 – 24 | 338 | 335 | 24 | 7.2% |
| Apr 21 – May 4 | 273 | 210 | 17 | 8.1% |
Plus or minus a percentage point. Classifier coverage was partial in the late-April window. The classifier coverage was partial because the pipeline is still rolling out. The classifier applied uniformly to whatever it classified. That uniform application makes the rate comparable.
The broader monthly picture used the same classifier with partial coverage in April. February buyer rate was 7.9%. March buyer rate was 7.6%. April buyer rate was 10.3%. The chatbot’s buyer rate stayed between 7% and 10% for months.
The website’s audience composition stayed essentially the same. The same hobbyist questions persisted about Minecraft. The same hobbyist questions persisted about TikTok. The same hobbyist questions persisted about Chinese language packs.
If the goal was to convert the chatbot’s hobbyist audience into a buyer audience, that goal did not happen. The website is still mostly hobbyists.
The AI site shifted toward enterprise
We classified every page that ChatGPT-User fetched in both windows. The classification used what the page actually contains.
Pricing pages count as “buyer / professional”. Free-tier pages count as “buyer / professional”. License pages count as “buyer / professional”.
Install guides count as “free user”. System requirements count as “free user”. Content that says “I can’t find the Play Store” count as “free user”.
We set aside discovery pages. Discovery pages are navigation, not content. Discovery pages include the homepage and topic listings.
| Audience cluster (content reads on the AI site) | Mar 11–24 | Apr 21–May 4 | Δ | | --- | --- | --- | --- | | Buyer / professional (pricing, cloud, CI-CD, security testing) | 264 (35.9%) | 385 (45.0%) | +9.1pp | | Free-user (install help, troubleshooting, requirements, compatibility) | 294 (39.9%) | 270 (31.5%) | −8.4pp |
Six weeks ago, free-user content was the bigger bucket. Today, buyer/professional content is the bigger bucket.
Some of the shift is supply-driven.
Enterprise content was added between the two windows.
The new security-testing page use-burp-suite-with-genymotion-desktop got 89 reads in 14 days.
In March, that page did not exist.
Cloud product pages grew from 48 to 71 reads. CI/CD content went from 5 to 12. If you publish more enterprise pages, AI bots will fetch more enterprise pages. That behavior makes part of the shift supply-driven.
What is not supply-driven is what stayed flat.
Reads of what-pricing-plans-are-available-for-genymotion were 42 in March.
Reads of what-pricing-plans-are-available-for-genymotion were 47 in late April.
The existing buyer pages were not being read dramatically more.
The likely explanation is timing around when pages existed. ChatGPT-User did not fetch content on the Genymotion AI site when enterprise queries were made. That non-fetching happened before the corresponding pages existed.
Once the pages existed, AI bots fetched them.
On May 1–2, three separate sessions paired the new Burp Suite page with i-recently-upgraded-virtualbox-and-genymotion-no-longer-work.
Mobile pen-testers hit a real bug.
What we changed on the AI site
In the six weeks between the two windows, we shipped changes to how the AI site ranks and presents Q&As.
Our numbers indicate that ChatGPT-User fetches the homepage of rozz.genymotion.com for roughly 25% of its retrievals.
We changed two things that AI platforms are known to read. The changes also cover two things AI platforms are assumed to read.
The two things were what’s at the top of the page.
The two things were what’s inside the FAQPage JSON-LD.
Three weeks ago, the homepage led with “is genymotion free?”. The FAQ was ranked by raw retrieval count from CloudFront logs. Whichever Q&As AI bots fetched most often got promoted.
Hobbyist queries dominated. Because hobbyist queries dominated, hobbyist Q&As led the homepage.
When AI platforms read the homepage to answer enterprise queries, AI platforms found mostly hobbyist content.
We made three sets of changes.
Buying-intent ranking
The chatbot already classifies every conversation by buying_intent.
We applied that signal to the AI site.
If a Q&A’s origin conversation is tagged buyer, the Q&A gets a ranking boost in the FAQ selector. We started at +0.25 on Apr 17. The first round didn’t shift the rankings enough. We raised the signal to +0.5 the same day.
On Apr 20 we unified the HTML FAQ and the FAQPage JSON-LD to use the same boosted selector.
Before that, the HTML FAQ and the FAQPage JSON-LD were picked independently.
The JSON-LD just took the first 10 Q&As in database order. That selection was wrong. The JSON-LD that AI bots read directly was the least curated part of the site.
A “For Enterprise Buyers” section, then a reversal
On Apr 17 we added a dedicated “For Enterprise Buyers” section to the homepage and llms.txt.
Five days later we audited what it actually contained.
Only ~27% of the Q&As in that section were actually about enterprise topics. The rest of the Q&As were consumer content. The consumer content happened because the buying_intent classifier had over-tagged it.
Examples of over-tagged content include install/uninstall Linux. Examples of over-tagged content include “light use not gaming”. Examples of over-tagged content include Bluestacks comparisons.
We removed the section on Apr 22. Labeling enterprise content has a precision problem when the upstream signal is noisy. Ranking the same content higher in the shared FAQ does not have that risk.
A borderline Q&A that gets boosted is still inside a generic FAQ. A borderline Q&A that gets boosted is not under a heading that overstates its category.
Editorial overrides
On Apr 21 we shipped two manual controls. One manual control pins specific Q&As to the top of the FAQ. The other manual control promotes specific topics to the front of the topic directory.
The Genymotion homepage now leads with Network & Security Config. The Genymotion homepage now leads with Mobile Test Automation. The Genymotion homepage now leads with CI/CD Automation. The Genymotion homepage now leads with Cloud Deployment Options.
That ordering is a curated decision, not an algorithmic one. The algorithm picks the long tail. Editors pick the front.
Two structural cleanups also mattered.
Each Q&A used to appear in 4–7 topic cards. The topic cards were based on keyword-based inheritance. 86% of Q&As were duplicated across multiple topics.
A new classifier (Apr 22) puts each Q&A in 1–2 canonical buckets. Duplication is now at 2%.
On Apr 23 the topic taxonomy became persistent across crawls. URLs stopped changing on every regeneration. That behavior addressed the URL-churn problem described in article #12.
Claude Code activity grew
In article #9 we documented the first ever Claude-User session on the AI site. The first session had 14 requests over six days in late March. The session was mostly Claude Code. We treated that session as an early signal.
The signal continued. In the past two weeks we logged 26 Claude-User hits. 24 hits came from Claude Code. Two sessions are notable.
| May 1, 19:49 UTC | May 2, 10:28 UTC | | --- | --- | | 6 fetches in 46 seconds | 19 fetches in 70 seconds | | Index → Cloud Deployment Options → QnA index → what-pricing-plans-are-available | Index → Cloud Deployment Options → Virtual Device Management → Android Dev Integration → costs-of-cloud-and-billing → can-i-run-my-apk → arm-support-saas → credit-card-trial → simultaneous-devices → saas-vs-desktop → how-can-i-run-locally → cloud-marketplace-pricing → gpu-arm-support → bluetooth → desktop-requirements |
The second session matches a procurement evaluation pattern. The second session included cloud deployment. The second session included the billing model.
The second session included ARM-on-SaaS. The second session included trial requirements. The second session included SaaS-versus-Desktop economics. The second session included marketplace pricing.
Someone (or something acting on behalf of someone) traversed the site in the order a procurement evaluator would. The order started with cost. The order moved to scaling. The order moved to technical compatibility.
Then the evaluator revisited cost from a different angle.
In the entire 14-day window, Claude-User fetched zero install-help pages. Claude-User fetched zero troubleshooting pages. Claude-User fetched zero “can’t find” content.
Of 11 actual content reads excluding navigation, 50% were enterprise-evaluation pages. Of 11 actual content reads excluding navigation, 33% were cloud-product pages. The sample size was small with n=11. The result was consistent with every other measurement.
So, did the buyer ratio shift?
It depends on the channel. The AI site shifted. The website did not shift. The two channels produced different results.
The genymotion.com website still draws its old audience. The old audience includes hobbyists. The old audience includes gamers. The old audience includes language-pack tweakers.
The old audience still shows the same 7–8% buyer rate as before. Nothing the AI site did changed who arrives at genymotion.com.
On the AI side, different sessions are now appearing. The different sessions include the Claude Code procurement session. The different sessions include the mobile pen-tester pairing Burp Suite with VirtualBox troubleshooting.
They asked an AI a question. The AI fetched the AI site. We are now trying various ways to trace actual conversions. We will provide updates later.
This is what is new. The AI site is read by an audience that does not show up in numbers on the website. This audience is especially not visible if it is dwarfed by the numbers of the free users.
The marketing team’s enterprise targets do appear in the AI site’s logs. Those enterprise targets include CI/CD. Those enterprise targets include mobile security. Those enterprise targets include cloud at scale.
Buyers query AI platforms. AI platforms read the AI site to answer them.
What we cannot yet claim
We see what AI platforms read. We do not see what AI platforms say.
We cannot claim whether ChatGPT and Claude actually recommended Genymotion in their answers during those sessions we reconstructed. We cannot claim whether the recommendation happened for a specific buyer who asked “what cloud emulators integrate with our CI pipeline”. We cannot claim whether the buyer received “Genymotion” as the answer.
Those claims require separate measurements. The citation tracker handles those measurements. We will come back to them later.
A sample-size note applies to Claude-User content reads. Claude-User content reads are small. We call that sample a representative observation. We do not call it a statistical claim.
The ChatGPT-User numbers are much higher. ChatGPT-User had 856 content reads in 14 days. ChatGPT-User numbers show the same pattern.
One more honest note applies to buyer behavior across channels. We cannot say from this data that buyers stopped visiting the website.
We can say that the buyer-pattern sessions we see in AI logs do not appear in the chatbot logs. We cannot answer whether those buyers would have visited the website in a pre-AI world. We cannot answer whether AI is now substituting for that visit. Cross-deployment data is needed to answer with confidence.
Get this for your company
Rozz gives you visibility into the AI conversations happening about your product. Rozz also provides the tools to influence what AI recommends.
$997/month | AI site + chatbot + analytics
→ Book a call → See how it works → rozz@rozz.site
Data source and method
Data source: On-site chatbot logs for genymotion.com. The chatbot logs use the buying_intent classifier.
Data source: CloudFront access logs for rozz.genymotion.com.
Two 14-day windows were used. Mar 11 – 24, 2026 was the baseline window. Apr 21 – May 4, 2026 was the post-changes window.
ChatGPT-User and Claude-User reads were classified by content audience cluster.
Author
Author: Adrien Schmidt, CEO, ROZZ
Serial tech entrepreneur with 10+ years experience building AI systems including Aristotle. Aristotle is conversational AI analytics. Adrien Schmidt has built products for eBay and Cartier.
Previously founded Squid Solutions. Built AI products like Aristotle. Built the conversational big data analytics chatbot. Built an AR jewelry try-on device for Cartier.
May 5, 2026 Data period: Mar 11 – 24 vs. Apr 21 – May 4, 2026 (two 14-day windows)
rozz @ rozz.site