The strongest evidence-backed tactics for AI citation, according to the original GEO research from Princeton, Georgia Tech, IIT Delhi, and the Allen Institute for AI, are adding specific statistics (up to 41 percent improvement), citing outside sources (up to 115 percent higher visibility for low-ranked pages), and including expert quotes (28 percent improvement). Keyword stuffing, by contrast, consistently hurt performance in the same study. If you've read general GEO advice that treats all nine tactics as equally worth doing, this is the study that shows they're not.
This research, published at KDD 2024 and generally referred to as the founding GEO paper, tested nine content-level tactics across roughly 10,000 queries using a benchmark called GEO-bench. It's the closest thing this field has to a controlled study, rather than an agency's anecdotal case study, which makes the actual effect sizes worth knowing in detail rather than just the headline "cite sources and add statistics" summary most secondary coverage reduces it to.
What were the nine tactics actually tested?
The study tested: adding citations, adding statistics, including expert quotes, using technical terminology, optimizing for fluency, keyword stuffing, making authority claims, ensuring unique content, and simplifying language. Not all nine were tested against the same baseline in the same way, and the effect sizes differ enough that treating this as a uniform checklist misses the point of the research.
Which tactics had the strongest measured effect?
| Tactic | Reported effect | What it actually means |
|---|---|---|
| Statistics | Up to 41% improvement (position-adjusted word count), 37% (subjective impression) | Concrete numbers embedded in a claim measurably increase how much of that passage gets used in the generated answer |
| Citing sources | Up to 115.1% higher visibility for pages starting at a low position | The strongest effect in the study, and notably it acts as an equalizer, helping low-ranked pages the most |
| Expert quotes | 28% improvement (subjective impression) | Direct quotes attributed to a named expert increase how favorably a passage reads to the evaluation model |
| Keyword stuffing | Negative effect | The one tactic tested that measurably hurt performance rather than helping it |
Why did citing sources have such a large effect, especially for low-ranked pages?
Because it functions as a trust shortcut for the model doing the evaluation. A claim with a cited source doesn't require the model to independently assess whether it's true. It can partially defer to the cited authority instead. That mechanism matters disproportionately for content that starts out less authoritative (a smaller site, a newer domain, a page without much existing traffic), since citing sources gives that content a way to borrow credibility it hasn't built independently yet. This is the same underlying logic behind why third-party corroboration matters more than your own website, just measured directly at the sentence level instead of the domain level.
Why did keyword stuffing actively hurt, when it's still common SEO advice?
Because the two systems are evaluating fundamentally different things. Classic keyword-based ranking historically rewarded (and in some crude implementations, still rewards) term frequency as a relevance signal. A generative model evaluating a passage for extraction and citation is judging fluency, coherence, and how naturally the passage reads as an answer, and repetitive keyword density actively damages all three. This is one of the clearest places where a tactic inherited from old-school SEO doesn't just fail to help GEO, it works against it.
Does this mean I should rewrite everything around these four tactics?
Not literally, but it does suggest a clear prioritization if you're deciding where to spend limited content-editing time. Given a choice between adding a vague qualifier ("many experts agree") and a specific number with a source ("a 2024 KDD study found a 41 percent improvement"), the second version is doing measurably more work by the study's own metrics. The other five tactics tested (technical terminology, fluency, authority claims, unique content, simplification) still matter for good writing generally, but the evidence for their specific effect on AI citation is weaker or more context-dependent than the top four.
How should this change how I edit an existing page?
A concrete, three-pass approach based directly on the strongest-effect tactics:
- Find every vague claim and add a specific number. "Significantly improves" becomes "improves by 41 percent." If you don't have a real number, either find one or soften the claim, since an invented statistic is worse than an honest qualifier.
- Find every unsupported assertion and add a citation. Given the size of the effect specifically for lower-authority pages, this is often the single highest-impact edit available to a newer or smaller site.
- Replace any keyword-repetitive phrasing with natural language. If a sentence reads oddly because a target phrase was forced in more than once, that's actively working against citation likelihood, not just reading badly.
Frequently asked questions
Is this study still the most current research on GEO? It's the original, most-cited foundational study (KDD 2024), and it established the discipline's academic basis. Since then, additional research (including a NeurIPS 2025 benchmark called C-SEO Bench) has found that many later, less rigorous "GEO methods" are largely ineffective, while the core tactics from the original study and traditional SEO fundamentals remain relevant. Treat this as the best-evidenced baseline, not the final word.
Does adding statistics mean I should add numbers even if they're not that meaningful? No. The effect measured is for genuine, specific, verifiable statistics. A padded or irrelevant number doesn't carry the same trust signal, and if it reads as filler, it likely hurts the fluency-related tactics the same study also measured.
Do these tactics work the same way across ChatGPT, Perplexity, and Google AI Overviews? The study's benchmark aggregated across multiple systems, but individual engines do have documented differences in retrieval behavior. Treat these tactics as a strong general baseline, and pair them with direct testing on the specific engines your audience actually uses.
Is citing sources the same thing as getting cited by sources? No, and this is worth being precise about. This study measures the effect of your content citing outside sources within itself. Being cited by others (the third-party corroboration effect) is a related but separate mechanism, working at the domain and reputation level rather than the sentence level.
Next step
Open your best-performing page and count how many specific, sourced claims it contains versus how many vague or unsupported ones. If vague claims outnumber specific ones, that's the single edit with the strongest evidence behind it from this entire study.