Most GEO advice on the internet is one of three things: a thinly-veiled SEO playbook with “AI” prepended, an unsourced agency listicle, or pure speculation. We’re going to do something different here: take the tactics that have measured, peer-reviewed evidence behind them - primarily from the Princeton GEO paper (Aggarwal et al., 2023) - and lay out exactly how to apply each one.

These five tactics, in the Princeton benchmark, produced visibility improvements in generative-engine answers of up to 40% when applied in combination.

1. Statistics Addition

The tactic: Insert specific numerical data - measured percentages, study findings, hard counts, dates - wherever a vague claim could be quantified.

Why it works: LLMs are calibrated to prefer specific, verifiable claims over generalities. When the model is generating an answer, content with concrete numbers is statistically more likely to be selected and quoted directly.

Princeton’s finding: Statistics Addition was the strongest single tactic in the “Law & Government” domain and produced consistent 30–40% gains across factual content (Aggarwal et al., 2023).

What to do:

  • Replace “many users” with “58% of users.”
  • Replace “recent research shows” with “a 2024 McKinsey study of 3,500 executives found.”
  • Audit your existing content for every “many,” “some,” “often,” “frequently” - and replace each one with a number, sourced.

What to avoid:

  • Made-up statistics. LLMs are increasingly good at flagging unverifiable claims, and Perplexity’s quality-gate reranker explicitly filters for source corroboration (ZipTie, 2025).

2. Quotation Addition

The tactic: Embed direct, attributed quotes from authoritative figures, experts, or named studies.

Why it works: Quotations function as embedded micro-citations. When an LLM is summarizing a page, it tends to preserve attributed quotations intact - meaning your page becomes the route through which the cited authority reaches the user.

Princeton’s finding: Quotation Addition produced its largest gains in historical and opinion-based content and was one of the three top-performing tactics overall.

What to do:

  • For every major claim, find a credible source who has said something similar and quote them with attribution.
  • Use blockquote formatting in your content - semantic HTML cues help LLMs recognize the quoted material.
  • Include quote attribution inline: “quote” - Source Name, Affiliation, Year.

What to avoid:

  • Bare claims with no attribution. LLMs penalize “naked” assertions in their citation behavior.

3. Citation Addition (Outbound)

The tactic: Cite your sources inline. Link out to authoritative references.

Why it works: Counterintuitively for traditional SEO - which historically discouraged outbound links - LLMs reward content that itself behaves like a citable source. A page that cites authoritative references is identified by retrieval systems as part of a credible knowledge cluster.

Princeton’s finding: Citation Addition produced 30–40% improvements in factual domains, particularly for queries seeking factual information.

What to do:

  • Link to primary sources: academic papers, government data, major institutions.
  • Use descriptive anchor text - not “click here.” LLMs use anchor text to understand the relationship between your content and the cited source.
  • Place citations adjacent to the specific claim they support, not in a footer block.

What to avoid:

  • Citing your own content for everything. LLMs detect self-citation patterns and discount them.
  • Linking to low-authority or commercial sources for factual claims.

4. Authoritative Voice

The tactic: Write with definitive, expert-toned language. Avoid hedging and weasel words where the data supports a firm claim.

Why it works: LLMs are trained on vast corpora where authority correlates with definitive language. When the model is choosing which sentence to extract for an answer, it preferentially selects the most direct, declarative statement.

Princeton’s finding: Authoritative Voice produced its largest gains in historical and reference content, where users implicitly expect a confident, expert framing.

What to do:

  • “The data shows X” instead of “the data may suggest that X could be the case.”
  • Lead paragraphs with the core finding, not throat-clearing.
  • Where the evidence is strong, write like an expert who knows it. Where it’s not, hedge honestly - but don’t hedge by default.

What to avoid:

  • False confidence. If a claim isn’t well-supported, hedging is correct. LLMs are increasingly cross-referencing claims against multiple sources, and overclaiming damages your citation rate over time.

5. Fluency Optimization

The tactic: Make your prose readable, scannable, and well-structured. Use clear headings, short paragraphs, and logical flow.

Why it works: LLMs preferentially extract from well-structured passages. The retrieval-augmented generation (RAG) pipelines used by Perplexity, ChatGPT, and others rely on chunking - splitting documents into ~200–500 token passages - and a well-structured passage with clear topic sentences chunks more cleanly than a wall of unstructured text.

Princeton’s finding: Fluency Optimization was a consistent, moderate-gain tactic across most domains - not the biggest single lever, but a high-floor one.

What to do:

  • Use H2/H3 headings every 200–400 words.
  • Lead paragraphs with topic sentences.
  • Keep paragraphs to 3–5 sentences.
  • Use bulleted lists for enumerable content. Bullet items are highly extractable for AI answers.

What to avoid:

  • Walls of text. They chunk poorly and get cited less.
  • Headings that are too clever to be parsed (e.g., puns instead of clear topic descriptors).

Three things the data does NOT support

Honesty over hype. Some widely-circulated GEO claims are not yet supported by peer-reviewed evidence:

  • “50–150 word chunks get 2.3x more citations” - agency-sourced, no published methodology.
  • “Pages with 15+ Knowledge Graph entities get 4.8x more AI Overview inclusion” - agency-sourced, not peer-reviewed.
  • “0.334 correlation between brand search volume and citation rate” - agency-sourced, methodology unclear.

These claims may turn out to be true - and the directional logic behind them is sensible - but if you’re presenting them to stakeholders, present them as industry-reported observations, not established science.

The tactics in this article, by contrast, all trace to either the Princeton benchmark or to published technical analyses of production AI systems. They are the evidence-backed foundation. Everything else is execution on top.

The one tactic that actively hurts you

The Princeton paper also tested Keyword Stuffing - the classic SEO behavior of repeating target keywords throughout a page.

It produced negative gains. Visibility in generative-engine answers decreased when keyword stuffing was applied.

This is the cleanest signal in the entire paper: the optimization surface has inverted. What helped you on Google in 2015 actively hurts you in ChatGPT in 2026. Brands still doing keyword density audits are spending money to become less visible in AI search.

Putting it together

If you implement the five tactics above on your highest-priority pages and remove any keyword-stuffing legacy, the Princeton benchmark predicts a measurable lift in citation rate across major AI agents.

That lift, however, only becomes a business outcome if you can measure it - daily, across every agent, against your competitors. Cite rate is the new rank. The tools to measure it are different. The discipline to act on it is different. But the work - get cited, stay cited, expand cited - is the entire ballgame.

Primary source for this article:

Supporting sources: