The Princeton GEO Study: What Actually Works for AI Visibility

Most advice about AI optimization is guesswork.

In 2024, researchers at Princeton conducted the first large-scale empirical study of what actually works.

They tested various content modifications across 10,000 queries and measured impact on citation probability in AI-generated answers.

Here’s what they found.

The Study Design

Researchers modified content in specific ways and tracked whether AI systems were more or less likely to cite it.

What they tested:

  • Adding expert quotes with attribution
  • Including statistics and data
  • Adding inline citations to other sources
  • Improving readability and text flow
  • Using domain-specific terminology
  • Simplifying language
  • Adopting authoritative tone
  • Keyword stuffing (the old SEO tactic)

What they measured:

Whether the modified content was cited in AI-generated answers more or less frequently than the control.

The Results: What Works

Here’s the data, ranked by impact:

Expert Quotes: +41%

Adding expert quotes with proper attribution increased citation probability by roughly 41%.

Example of what works:

According to Dr. Sarah Chen, lead researcher at Stanford’s AI Lab, “Large language models prioritize sources that demonstrate expertise through direct quotation and attribution.”

Why it works:

AI models use quotation marks and attribution as a proxy for credibility.

When the model sees properly formatted quotes, it interprets the content as more authoritative.

Clear Statistics: +30%

Including specific statistics and data increased citation probability by about 30%.

Example of what works:

A 2024 study of 2.3 million keywords found that AI Overviews now appear on roughly 60% of informational queries, up from around 15% in early 2024.

Why it works:

Statistics signal “factual density” to the model.

The presence of specific numbers, percentages, and data points indicates the content is evidence-based rather than opinion.

Inline Citations: +30%

Adding citations to other authoritative sources increased visibility by around 30%.

Example of what works:

Research from BrightEdge and Semrush confirms this trend, showing organic CTR declining by 20-40% for queries where AI Overviews appear.

Why it works:

Citations create a chain of trust.

When your content cites authoritative sources, the AI interprets it as part of a credible information network.

Improved Readability: +22%

Making text more readable and fluent increased citations by roughly 22%.

What this means:

  • Shorter sentences
  • Clearer structure
  • Logical flow
  • Active voice
  • Concrete language

AI models evaluate text quality using perplexity scores. Natural, well-written text scores better.

Domain-Specific Terminology: +21%

Using appropriate technical language (when relevant) increased visibility by about 21%.

Example:

For a technical audience, using precise terms like “Largest Contentful Paint (LCP)” instead of “page load speed” signals expertise.

But: this depends on context. For general audiences, simpler language wins.

Simplified Language: +15%

When writing for general audiences, simpler language increased citations by around 15%.

Example:

“Your website loads too slowly” beats “Your site exhibits suboptimal rendering performance.”

Authoritative Tone: +11%

Writing with confidence and authority (without arrogance) increased visibility by roughly 11%.

What this means:

  • State facts clearly
  • Avoid hedging every claim
  • Use active voice
  • Be specific

What Hurts: Keyword Stuffing

Keyword stuffing decreased visibility by roughly 9%.

Example of what fails:

Best Dallas HVAC repair Dallas HVAC near me Dallas air conditioning Dallas AC repair emergency Dallas HVAC service Dallas Texas HVAC best Dallas air conditioner repair…

Why it fails:

Keyword stuffing degrades text quality.

AI models detect this through perplexity analysis. Unnatural, repetitive text gets downgraded.

How to Apply This to Your Content

Here’s a practical checklist based on the Princeton findings:

1. Add Expert Quotes to Every Major Topic

If you’re the expert, quote yourself with attribution.

Format:

“Same-day service isn’t just a feature for us, it’s a commitment,” says Maria Rodriguez, owner of ABC HVAC. “We’ve built our entire dispatch system around getting to customers within 2 hours of the call.”

2. Include Specific Statistics

Replace vague claims with specific data.

Before: “Most customers prefer transparent pricing.”

After: “In our 2024 customer survey of 500 homeowners, 78% said they value upfront pricing more than the lowest price.”

3. Cite Your Sources

When you reference industry data, name the source.

Format:

According to a 2024 study by BrightEdge analyzing over 1 million queries, AI Overviews now appear on roughly 50-60% of informational searches.

4. Write Clearly and Concisely

  • Short sentences
  • One idea per sentence
  • Active voice
  • Specific details

5. Match Language to Audience

For technical audiences: Use precise terminology. Define it once, then use it consistently.

For general audiences: Use plain language. Define technical terms in simple words.

6. Avoid Keyword Stuffing

Write for humans first.

Include keywords naturally where they fit the sentence.

Don’t force repetition.

What About Schema and Structure?

The Princeton study focused on text-level tactics.

Other research shows that structural signals also matter:

Schema markup (LocalBusiness, FAQPage, HowTo) helps AI systems understand what your content is about.

Heading hierarchy (H1, H2, H3) provides the scaffolding AI uses to parse your page.

Direct-answer format (40-60 word answers immediately after question headings) makes extraction easier.

These work in combination with the text-level tactics from Princeton.

Measuring Your Results

You can test these tactics yourself:

Manual method:

  1. Update one page with expert quotes, statistics, and citations.
  2. Wait 2-4 weeks for AI systems to recrawl.
  3. Ask ChatGPT, Perplexity, Claude, and Gemini questions that page should answer.
  4. See if you’re cited more often.

Automated method:

Use Surmado Signal ($25) to test across 7 platforms with multiple personas automatically.

Run it before and after changes to measure impact.

The Bottom Line

The Princeton GEO study gives us quantitative benchmarks for the first time.

What works:

  • Expert quotes: +41%
  • Statistics: +30%
  • Citations: +30%
  • Readability: +22%
  • Domain terminology (when appropriate): +21%
  • Simplified language (for general audiences): +15%
  • Authoritative tone: +11%

What hurts:

  • Keyword stuffing: -9%

AI models favor content that looks evidentiary, well-sourced, and readable.

The old SEO playbook (keyword density, exact-match phrases) actively hurts your visibility in AI answers.

Apply these findings to your most important pages this week.

Then measure the results.

Learn more: