AI crawler behavior observed on production mirror sites
Insights
Overview
The Crawler Logs document real AI crawler behavior on production AI sites built by ROZZ. Using CloudFront access logs and User-Agent classification, we track how GPTBot, BingBot, ClaudeBot, ChatGPT-User, PerplexityBot, and other LLM bots discover, crawl, and cite GEO-optimized content. Across 143 days of observation on rozz.genymotion.com, we recorded 46,418 total LLM bot requests and 14,465 citation events—real users receiving Genymotion content in AI conversations.Totals
- Total LLM Bot Requests: 46,418
- Citation Events: 14,465
- Days of Data: 143
- Log Entries: 5
Log Entries
Entry #15 · June 29, 2026
Who actually reads your llms.txt? We logged 461,328 requests to find outA 90-day window of 461,328 bot requests across our production AI-native sites. llms.txt and the JSON feeds drew 0.97% of traffic combined. The bots that answer users ignored the AI-native layer hardest (ChatGPT-User 0.14%, ClaudeBot 0.02%). The sitemap outdrew llms.txt 67 to 1.
Production AI-native sites · 90-day window · 461,328 requests across 31,935 log files · 20+ crawler classes
- The AI-native add-on layer is a rounding error:
llms.txt0.30%, JSON feeds 0.67%; no crawler spends more than 2.4% on it - Citation bots ignore it hardest: ChatGPT-User 0.14%, ClaudeBot 0.02% (zero
llms.txtfetches in 90 days) - The index crawlers actually use is
sitemap.xml(20.4% of traffic), which outdrawsllms.txt67×
Entry #14 · May 19, 2026
What AI bots read, what they ignore, and what an AI site is actually for. Part 1: Citation bots21 days of clean bot logs. ChatGPT-User and Claude-User made 1,517 content fetches on Genymotion’s AI site. Q&As behave as demand-driven answers (61% corpus coverage). Pages behave as supply-driven source citations (24% coverage, but 79% of Q&A sessions also fetched at least one cited source).
rozz.genymotion.com · Apr 29 – May 19, 2026 · 21 days · 545 pages + 262 Q&As + 16 topics + 2 runbooks
- Q&A coverage 2.5× higher than pages; Q&As show a real long tail (top-10 = 41% of fetches)
- 76% of /pages corpus never fetched directly — but pages get pulled as Q&A source citations
- ClaudeBot: 209 sitemap.xml fetches, 0 llms.txt fetches — AI-native discovery isn’t the channel yet
Entry #13 · May 5, 2026
Two channels, two audiences. The hobbyists are on your website. The buyers are on the AI.Three weeks after shipping enterprise content on Genymotion’s AI site, we compared two 14-day windows. The on-site chatbot’s buyer rate stayed flat (7.2% → 8.1%). The AI site’s buyer/professional content reads jumped from 35.9% to 45.0%. Two channels, two audiences.
rozz.genymotion.com + genymotion.com chatbot · Mar 11–24 vs. Apr 21–May 4, 2026 · 856 ChatGPT-User content reads (14 days)
- Chatbot buyer rate flat at 7–8% across months—website still draws hobbyists
- AI site buyer/professional reads: 264 (35.9%) → 385 (45.0%); free-user reads dropped from 39.9% to 31.5%
- Claude-User now 26 hits / 14 days—24 from Claude Code; one 19-fetch procurement pattern in 70 seconds
Entry #12 · Apr 22, 2026
AI Site Structure Matters More than We ThoughtGenymotion’s AI site has 16 topic pages. In one week of logs, AI platforms asked for 61 more that no longer exist—1,001 requests to retired topic URLs. Clustering-algorithm iterations kept improving the taxonomy but broke every external cache that learned the old map.
rozz.genymotion.com · Apr 15–22, 2026 · 3,484 LLM bot requests this week
- 61 retired topic URLs drew 1,001 requests in 7 days; ChatGPT-User contributed 387
- Live topic hubs also crawled every week—LLM systems treat the topic layer as durable structure
- Fix: canonical topics + URL registry + 61 301-redirect mappings to current equivalents
Entry #11 · Apr 14, 2026
Your customers are asking AI about you right now. You just can’t see it. We can.People evaluate products through ChatGPT, Claude, and Perplexity now—and the company whose product is being discussed usually has no idea it happened. 667 chatbot conversations and 2,500+ reconstructed ChatGPT sessions show who’s actually asking, what they want, and where the conversations lead.
rozz.genymotion.com + Genymotion.com chatbot · Mar 1 – Apr 14, 2026 · 3,830 March AI site fetches, 1,323 this week
- Two data sources reveal invisible AI-mediated evaluations: chatbot intent on the site + ChatGPT session reconstruction via CloudFront logs
- 95% of chatbot conversations are hobbyists/individual users; enterprise-buyer goals represent <5%—not the audience marketing targets
- Reconstructed sessions show live pricing evaluations, cross-continent macOS checks, and recurring VirtualBox bug reports no one else sees
Entry #10 · Apr 8, 2026
Three Citation Bots. Ten Articles. What We Learned Building an AI Site.Three AI platforms—ChatGPT-User, Claude-User, Perplexity-User—now retrieve content from the same AI site during live user sessions. 9,250 citation requests across 90 days. Three months ago there were zero.
rozz.genymotion.com · Jan 8 – Apr 8, 2026 · 22,679 total LLM bot requests (90-day cumulative)
- Perplexity-User first appearance (Apr 5)—all 3 major citation pipelines now active
- 376 commits across 90 days; structural fixes (sitemaps, robots.txt, topic taxonomy) drove every breakthrough
- Q&A pages drive 66–75% of citations; CLI runbooks open a developer-tool sales channel
Entry #9 · Mar 31, 2026
Selling in Claude Code: From Pricing to Implementation in 10 Seconds with an AI SiteA developer asked Claude Code how much Genymotion costs. Ten seconds later, it was showing them how to set it up. Both answers came from our AI site. 14 Claude-User requests in 6 days, 12 from Claude Code.
rozz.genymotion.com · Mar 24–31, 2026 · 1,612 total LLM bot requests
- Claude-User: first-ever appearance—14 requests, 12 from Claude Code terminal sessions
- 10 seconds from pricing Q&A to CLI runbook—evaluation and implementation in one session
- Claude pipeline complete: ClaudeBot crawl → 5 days → Claude-User live retrieval
Entry #8 · Mar 24, 2026
And now we have it: ClaudeBot just made 958 requests.ClaudeBot made 958 requests in one week, up from 123 the week before. 503 GEO pages and 162 Q&A pages—the largest ClaudeBot crawl since December. Six hours after deploying per-topic sitemaps, it came back.
rozz.genymotion.com · Mar 17–24, 2026 · 2,446 total LLM bot requests
- ClaudeBot 8x increase (123 → 958) triggered the day per-topic sitemapindex was deployed
- March 20: 577 ClaudeBot requests in a single day—largest since December
- Every major AI crawler has now completed a deep indexing event on the AI site
Entry #7 · Mar 17, 2026
We changed one page. PerplexityBot went from 42 requests to 511.PerplexityBot made 511 requests in one week, up from 42 the week before. It crawled 172 Q&A pages and 256 GEO pages—more content in 7 days than in its entire prior history on the site combined.
rozz.genymotion.com · Mar 10–17, 2026 · 2,532 total LLM bot requests
- PerplexityBot 12x increase (42 → 511) triggered the day after index page redesign
- 84% of PerplexityBot requests hit content pages—Q&A and GEO pages, no homepage
- ClaudeBot: 0 Q&A pages, 0 Claude-SearchBot traffic—still in discovery mode after 3 weeks
Entry #6 · Mar 10, 2026
What the AI Site reveals about AI-mediated discoveryChatGPT-User made 681 visits in one week. By grouping visits into sessions using IP hashes and timing, we reconstructed 168 sessions showing how users navigate AI-mediated discovery.
rozz.genymotion.com · Mar 3–10, 2026 · 681 ChatGPT-User visits
- 4.6 pages fetched per turn—ChatGPT-User verifies across multiple sources, not just one
- 28% of sessions hit only the index page and stopped—led to index redesign
- 30% of sessions are multi-turn: we can reconstruct actual ChatGPT conversations
Entry #5 · Mar 3, 2026
The platforms that crawl your AI site the most cite you the most (except one)We tested 24 queries across four AI platforms. ChatGPT cites Genymotion 83% of the time. Claude, 21%. Perplexity, 17%. Gemini, 4%. The first three track with crawl volume. Gemini doesn’t crawl the AI site at all.
rozz.genymotion.com · Feb 24 – Mar 3, 2026 · 24 queries × 4 platforms
- ChatGPT: 83% citation rate, up from 14% before the AI site launched
- ClaudeBot jumped 24x (21 to 505 requests)—reads topic taxonomy, not individual pages
- Gemini recommends Genymotion in most queries but doesn’t link—different pipeline entirely
Entry #4 · Feb 24, 2026
Bing just found Genymotion: 1,556 BingBot requests and what it meansBingBot made 1,556 requests this week—more than any other bot, including ChatGPT-User. What started as a ChatGPT story is now happening across six platforms.
rozz.genymotion.com · Feb 17–24, 2026 · 3,188 total requests
- BingBot: 1,556 requests—largest single bot category, surpassing ChatGPT-User
- ChatGPT-User citations: 1,329, up from 1,077 the prior week (+23%)
- Six platforms now crawling: OpenAI, Microsoft, Anthropic, Meta, ByteDance, Perplexity
Entry #3 · Feb 17, 2026
3x Week-Over-Week: What Sustained ChatGPT Citation Growth Looks LikeChatGPT citations hit 1,077 this week—3x the 345 reported last week. Six-week trajectory: 42 → 345 → 1,077. Q&A pages drive 66% of all citations.
rozz.genymotion.com · Feb 10–17, 2026 · 2,606 total requests
- 1,077 ChatGPT citations in 7 days—3x the previous week's 345
- Q&A pages remain the citation driver: 66% of all ChatGPT requests
- PerplexityBot activity up 5x: 33 requests vs 7 the previous week
Entry #2 · Feb 10, 2026
16x Citation Growth in 7 Days: What ChatGPT Users Are Actually AskingChatGPT citations grew from 7 to 116 in one week. 345 citation events, 161 unique sessions, and 75% of requests hitting Q&A pages—not traditional content.
rozz.genymotion.com · Feb 2–9, 2026 · 2,195 total requests
- 16x daily citation growth: from 7 on Feb 2 to 116 on Feb 9
- 75% of ChatGPT requests landed on Q&A pages, not traditional content
- Q&A pages cited 10x more than traditional GEO content pages
Entry #1 · Feb 3, 2026
547 Requests in One Day: What Happens When GPTBot Discovers Your Mirror SiteGPTBot made 547 requests in a single day—47% of all training bot activity in 30 days. Three weeks later, ChatGPT users were receiving content in their conversations.
rozz.genymotion.com · Jan 3–Feb 2, 2026 · 1,280 total requests
- GPTBot made 547 requests on January 7—47% of 30-day training activity in one day
- 42 citation events recorded; concentrated on high-intent pages (requirements, compatibility)
- ~3 weeks from major crawl to first ChatGPT citations
Methodology
Data source: CloudFront access logs for rozz.genymotion.com. Bot classification based on User-Agent strings. Training bots (GPTBot, ClaudeBot) are distinguished from search index bots (OAI-SearchBot) and citation events (ChatGPT-User). Citation events represent real user conversations where ChatGPT retrieved and cited mirror site content. All data is from a single production mirror site; results may vary by domain, content volume, and vertical.Get Your Own Crawler Logs
ROZZ builds mirror sites with full CloudFront logging. See exactly which AI bots crawl your content, when they crawl it, and when citations begin.
$997/month | Mirror site + crawler analytics
- Book a call
- See how it works
- rozz@rozz.site
Author
Author: Adrien Schmidt, CEO, ROZZ
Serial tech entrepreneur with 10+ years experience building AI systems including Aristotle (conversational AI analytics) and products for eBay and Cartier.