Applied research on GEO, AI citation patterns, and LLM behavior — the thinking that becomes the next tool. Published when ready, not on a schedule.
We ran one 100-query set through both Claude and Perplexity and compared every citation. They cited 793 domains between them — and agreed on 171. The overlap is 22%, and 16 of the 100 questions produced no shared source at all.
We ran 100 commercial and GEO queries through Perplexity (Sonar) and logged every citation. Reddit showed up in 89 of 100 answers. Perplexity cites a different web than a publication-first engine does.
Across seven topics the number of sources an AI answer cites barely moved — but 91% of the domains it cited belonged to a single topic. Same appetite, almost entirely different worlds.
Across 100 queries and 462 cited domains, 23 recurred in three or more answers. The recurring head is the SEO-tool ecosystem — and it isn't the short list a smaller sample named.
Across 100 commercial queries through one engine, the answer was strikingly stable: about six cited sources per answer, never a couple, never a flood.
We tested 23 schema.org types across 4,000 pages and measured the citation lift for each. Seven tags reliably moved the citation needle; the rest were noise or worse. Here's the ranked list — with the implementation that actually shipped lift, not just validated.
We tested 40 content changes across 120 queries to isolate which on-page factors increase the chance of being cited in Google's AI Overviews.
The engineering behind GEO Trace: an adapter registry that can't claim an engine it can't run, one shared call path for the live tool and the research panel, and the things this tool deliberately does not do.