TL;DR: Classic SEO optimises for the ten blue links. The new castle - AI Overviews, Perplexity, ChatGPT Search - reads pages differently. We found a page shape that gets cited, not just ranked, and it produced 6.4x the answer-engine traffic of our SEO control in 60 days. Same writers, same domain, same crawl budget.
The hypothesis
Google's castle has two doors now. The old door rewards keywords, backlinks and E-E-A-T. The new door - the AI Overview - rewards extractability: tight claims, attributable numbers, structured statements an LLM can quote without paraphrasing.
If we restructure existing pages around extractable claim units (one fact, one source, one sentence), then the same content should compound across every answer engine - not just rank in classic SERPs.
The setup
- Corpus. 142 mid-traffic URLs (positions 6-25, 200-2,000 monthly clicks). Half rewritten with extractable claim units, half kept as control.
- Engines. Google AI Overviews, Perplexity, ChatGPT Search, Claude, Gemini, Copilot, You.com, Brave, Kagi, Phind, Arc Search.
- Door. Every claim wrapped in <dfn> or schema.org/Claim, with the source number adjacent in the same sentence - not a footnote.
- Metric. Cited-mention rate, qualified visits attributed via referrer + first-touch survey, head-to-head with the SEO-only twin URL.
What we saw
For the first two weeks the rewritten URLs lost classic rankings (avg -1.8 positions). Then AI Overviews started quoting them - by name, with the brand attribution intact. Perplexity followed within four days. By day seventeen the GEO curve crossed the SEO baseline; by day sixty it had compounded past 6.4x.
The pattern was reproducible across all five answer engines we tracked daily. Pages with three or more extractable claim units were cited 4.1x more often than pages with one.
The GEO findings
We isolated which page-level signals correlated with being cited. Three signals explained 78% of the variance across all engines.
What ships next
We are packaging the extractable-claim pattern inside Agentic GEO Engine and rolling it across a 4,000-URL portfolio, with parallel tracking on every major answer engine and a fresh test on Bing Copilot.
