Lesson 02·GEO Tactics·~10 min·why → tactics
The citable passage
Lesson 1 showed how engines pick sources — they retrieve pages, then quote specific passages from them. This is where you build a passage they can actually lift, verify, and attribute — using the edits research says matter (and honestly flagging the ones it now doubts).
In ~10 minutes you'll be able to
- Name the content edits that measurably lift AI citation (and the one that backfires).
- Rewrite a flat paragraph into a citable answer block, step by step.
- Hold the honest caveat: where the evidence is strong vs. hyped.
Short on time? Watch the 60-second version.
The whole skill, up front:
To make a paragraph citable, do five things: front-load a 40–80 word answer, name the entity explicitly, add a specific number with its source, add an attributed — and real — expert quote, and write it plainly — no keyword stuffing. In the foundational GEO study, edits like these lifted visibility by roughly 30–40%, while keyword stuffing lowered it.[1] A 2025 benchmark is far less bullish (see §5) — so treat this as evidence packaging, not a magic hack.
01The tactic leaderboard (what actually works)
In 2024, Aggarwal et al. (Princeton + Georgia Tech / IIT-Delhi) ran the foundational controlled experiment on GEO: 9 content edits, tested across a 10,000-query benchmark, measured by Position-Adjusted Word Count (how much of the answer is sourced from your page, weighted toward the top) and Subjective Impression (how prominently you're featured).[1] Here's how the edits ranked — these are the relative lift on the Word component of that metric; the paper's own headline for its top methods is a more conservative ~30–40%:[2]
GEO tactic leaderboard
relative lift on the Word sub-metric of PAWC (baseline→method) · Aggarwal et al. 2024 · paper's top-method headline: ~30–40% PAWC / 15–30% SI
⚠ Directional, and contested. A 2025 benchmark (C-SEO Bench, NeurIPS) re-ran these same edits across more models and tasks and found them mostly ineffective — sometimes negative — with retrieval position mattering more than the rewrite. We handle this head-on in §5.[5]
Three things to take from this:
1. Quotes, stats, and citations win — and they require almost no new content. You're not writing more; you're making what you have verifiable.[1]
2. Keyword stuffing actively hurts. The reflex that defined a decade of SEO is the one move that lowers AI visibility (−8.7% here, −10% in their real-world Perplexity test).[1] This is the clearest proof that GEO ≠ SEO-as-usual.
3. In that study, the biggest wins went to pages that weren't already #1. Buried in the paper: Cite Sources gave a +115% lift to a page ranked ~5th, while the top-ranked page actually lost ~30%.[3] Encouraging if you're the underdog — but newer work finds the source's retrieval position usually matters more than the rewrite (§5), so earn the ranking too, don't just polish the prose.[5]
02The recipe: a citable answer block
Every winning tactic above folds into one repeatable shape. When you want a passage cited, build it like this:
Never invent quotes, statistics, or sources. A weak real source beats a polished fake one. Fabricated attribution destroys the exact trust GEO depends on — and creates real legal and reputational risk. This whole lesson is about packaging real evidence, never manufacturing it. If the proof doesn't exist yet, go earn it (run the survey, collect the reviews), then cite it.
03Rewrite Lab — turn a dead paragraph into a citable one
Here's a paragraph a sandwich shop might actually publish. Apply each edit and watch the citability climb. This is the exact move you'll make on your own pages.
Interactive · your feedback loop
Rewrite Lab
Click "Apply next edit" to rewrite the passage step by step.
"We make the best sandwiches in town and locals love us. Come try our famous subs — you won't be disappointed!"
"Dom's Deli, on 5th Street in Austin, makes hot Italian subs. Locals love us — come try them!"
"Dom's Deli, on 5th Street in Austin, is rated 4.8★ from 1,200+ Google reviews (2026). Known for its hot Italian sub."
"Dom's Deli (5th St, Austin) is rated 4.8★ across 1,200+ Google reviews and was named "the best Italian sub in Austin" by Austin Eater's 2026 guide."
04Spot the citable one
2 quick calls
Click an answer for instant feedback.
Q1You can only make one edit to a paragraph today. Which buys the most AI visibility?
Q2Which line is most citable for "best deli near me"?
05The honest caveat
How much to trust this
The leaderboard comes from one foundational study (Aggarwal et al., 2024), run on a simulated engine — GPT-3.5 with the top-5 Google results as context — and its headline numbers are derived from a single sub-metric.[4] Treat the exact percentages as directional, not gospel.
And there is serious counter-evidence. C-SEO Bench (NeurIPS 2025) re-tested these same content edits across more models, tasks, and domains and found them "largely ineffective" — often lowering a document's ranking — while the retrieval position of the source mattered far more than any rewrite.[5] So treat GEO copy edits as a second layer, not a replacement for crawlability, authority, ranking, and being genuinely useful.
The durable lesson survives both papers: GEO is evidence packaging, not magic. Make the claim clear, name the entity, add real proof, attribute it, and write it so a machine can lift it without guessing — and still earn the ranking that gets you retrieved. When advising someone, sell that principle, cite the studies on both sides, and let your own before/after tests be the proof.
Sources
- Aggarwal, Murahari, et al., "GEO: Generative Engine Optimization", KDD 2024 — up to ~40% visibility lift; quotations/statistics/citations win; keyword stuffing −8.7% (−10% on Perplexity). Full text: arxiv.org/html/2311.09735v3. ↩
- Exact per-tactic figures (v3 Table 2): Quotation +42.6% · Statistics +32.8% · Fluency +28.7% · Cite Sources +27.7% · Keyword Stuffing −8.7% (Position-Adjusted Word Count), reproduced at TurboAudit, GEO 2026 Guide. ↩
- Ranked-position finding: Cite Sources gave a 115.1% lift to a rank-5 page while the top-ranked page lost ~30.3% (Aggarwal et al., ranked-position analysis). ↩
- SandboxSEO, "GEO Paper: A Critical Review" — limits of the study (2024, simulated engine, sparse replication). Read for the skeptic's view. ↩
- Parameter Lab, "C-SEO Bench: Does Conversational SEO Work?", NeurIPS 2025 Datasets & Benchmarks Track — re-tested the Aggarwal content edits and found most "largely ineffective," frequently negative for ranking; retrieval position dominates. The key counter-evidence to the leaderboard. code/data. ↩
— end of lesson 2 —