This is a walkthrough of a 90-day content restructuring project on 12 existing, already-ranking articles, with no new content published and no link building involved. The only variable we changed was how the content was structured for extraction. Citation rate against a fixed 40-query panel went from roughly 1 in 10 checks returning a citation to our site, to roughly 3 in 10 — a 3x improvement.
Starting point
We picked 12 articles that already ranked on page one of Google for their target keyword, on the theory (later confirmed) that GEO amplifies existing authority rather than creating it from nothing. Each article was well-written by traditional SEO standards: solid keyword coverage, decent length, reasonable internal linking. None had FAQ sections or FAQPage schema. Most opened each section with a sentence or two of context before getting to the actual point.
We built a panel of 40 realistic queries related to the articles' topics and ran each one weekly against ChatGPT, Claude, Perplexity, and Google AI Overviews, logging whether our domain appeared among the cited sources.
Baseline citation rate: roughly 10% (about 4 of 40 queries per week returned a citation to one of our 12 articles, averaged over a two-week baseline period).
What we changed
1. Answer-first restructuring (the biggest lever)
For every H2 and H3 section, we rewrote the opening one to two sentences to directly answer the question implied by the heading, moving supporting context and caveats after it. A typical before/after:
- Before: "There are a number of factors that influence how search engines evaluate page speed, and it's worth understanding the broader context before diving into specific metrics..."
- After: "Largest Contentful Paint (LCP) under 2.5 seconds is the current Core Web Vitals threshold for a 'good' rating. The factors that push it past that threshold are usually..."
This alone, isolated in a partial rollout to 6 of the 12 articles before touching the rest, produced roughly half of the eventual total lift.
2. Added FAQ sections with FAQPage schema
Each article gained a 3-5 question FAQ section built from the actual queries in our test panel and from "People also ask" data, marked up with FAQPage JSON-LD. This was the lowest-effort, second-highest-impact change — most articles took under 30 minutes to add.
3. Tightened entity references
We replaced vague references ("the tool," "this metric") with consistent, specific names throughout each article, and made sure every tool, metric, or concept mentioned was spelled out in full at least once per major section rather than relying on a single early definition.
4. Added a summary block and speakable schema
Each article got a 2-3 sentence "key takeaways" style summary block near the top, marked as speakable alongside the article's main summary paragraph.
Results
| Metric | Baseline | Day 90 |
|---|---|---|
| Citation rate (panel of 40 queries) | ~10% | ~30% |
| AI-referral sessions/month | 340 | 1,180 |
| Average time to edit per article | — | ~6 hours |
Citation rate improved gradually rather than jumping immediately — the first visible movement in the query panel came 3-4 weeks after the first batch of edits went live, consistent with normal re-crawl and re-indexing latency plus whatever caching layer sits in front of each AI system's retrieval step.
What we'd do differently
We under-invested in tracking which specific sentences were getting quoted versus just linked — later spot-checks showed the AI systems were often quoting our FAQ answers almost verbatim, which suggests FAQ quality deserves even more editing time than we gave it in this round. We're also watching whether the effect decays if a competitor copies the same structural pattern — this case study reflects a first-mover advantage on structure, not necessarily a permanent one.
