The 11 Levers That Actually Drive AI Citations (Backed by Princeton Research)
Most "GEO advice" online is hand-wavy. The Princeton GEO study (KDD 2024) ran controlled experiments on 9 optimization techniques across Perplexity. Here are the 11 levers we built into Inxy, with their measured citation lift.
Why a controlled study matters
The "GEO" / AI search optimization advice on most marketing blogs in 2025 was vibes-based: "write for entities", "use schema", "be authoritative". None of it had a control group. None of it could tell you which lever moved the needle vs which was placebo.
The Princeton GEO study (Aggarwal et al., KDD 2024) changed that. They ran A/B-style tests across hundreds of queries on Perplexity.ai, comparing variants of the same source content with one optimization applied at a time. They published the citation-rate lift for each.
The 9 levers Princeton measured
| Lever | Citation lift | Notes |
|---|---|---|
| Add citations to authoritative sources | +40% | The single biggest lever. Quote sources by name + link. |
| Add specific statistics | +37% | Numbers > prose. "9.5 Mohs hardness" beats "very durable". |
| Add expert quotations | +30% | Named expert + title. Drops the LLM's suspicion of generated text. |
| Use authoritative tone | +25% | First-person plural, declarative claims, citations. |
| Improve clarity | +20% | Short sentences. One idea per paragraph. Active voice. |
| Use technical terms | +18% | Domain-specific vocabulary signals expertise. |
| Diversify vocabulary | +15% | Avoid same phrase twice in 100 words. |
| Optimize fluency | +15-30% | No grammar errors. No awkward AI tells. |
| Keyword stuffing | -10% | Actively hurts citation rate. Don't. |
Two findings stand out. First, citing sources alone gave +40% — more than statistics, quotations, or any other content technique. AI search engines preferentially cite content that itself cites well. Second, keyword stuffing was the only lever that NEGATIVELY affected citation rate. The traditional SEO playbook of repeating your target keyword 8 times per article actively reduces AI visibility.
These results are for Perplexity. ChatGPT and Claude likely behave similarly (their training corpora overlap), but the lift sizes may differ. Inxy's recommendations engine adjusts per-source based on observed citation patterns.
The 2 levers Princeton didn't cover (because they didn't exist yet)
The Princeton study was 2024 work. Two important AI search optimization vectors emerged in 2025-2026 that aren't in the original list:
- · Lever 10: Agent-readable files (AGENTS.md, pricing.md) — Anthropic's Claude Computer Use shipped in late 2024; OpenAI's GPT-5 agent capabilities followed. AI shopping agents read these structured files BEFORE rendering JS. Stores without them get filtered.
- · Lever 11: Knowledge Graph entity reinforcement — Brands with strong Wikidata QIDs, Google Knowledge Graph entries, and Crunchbase profiles get more CONSISTENT mentions in ChatGPT/Claude responses. Not necessarily more citations, but the citations are more accurate.
How Inxy automates each
| # | Lever | Inxy worker |
|---|---|---|
| 1 | Cite sources | Citable content generator (§C of spec 11) |
| 2 | Add statistics | Citable content generator (§C) |
| 3 | Expert quotations | Quote injector (§C — productized 2026-05) |
| 4 | Authoritative tone | Brand DNA layer |
| 5 | Clarity | Phase 4 product SEO worker |
| 6 | Technical terms | Brand DNA layer |
| 7 | Vocabulary diversity | Phase 4 product SEO worker |
| 8 | Fluency | Phase 4 product SEO worker |
| 9 | Anti-keyword-stuffing | Schema audit guardrail (§A) |
| 10 | AGENTS.md / pricing.md | Agent file builder (§H — productized 2026-05) |
| 11 | Knowledge graph | Entity audit (§I — productized 2026-05) |
Reading the levers as a map, not a checklist
The Princeton study confirmed what good editors knew intuitively: cited, specific, expert-led, well-written prose gets cited more than padded keyword-stuffed marketing copy. The 11 levers aren't an arbitrary list — they're basically "produce content that's genuinely useful, structured, and citable" broken into measurable components.
Which is, ironically, the same advice good content writers have given for 30 years. The difference now is that the AI search engines reading that content have specific, measurable preferences. Inxy operationalizes those preferences into per-shop weekly recommendations.
Want Inxy to do this for your store?
Connect Shopify + GSC + GA4 in 60 seconds. The first weekly Recommendations Engine output lands within a day. 14-day free trial, no credit card required.