Guide · Practitioner
Statistics, Quotes and Sources: Do They Raise Citation Rates?
In the GEO research paper's benchmark, three methods (adding citations, quotations and statistics) performed best, with a 30-40% relative improvement on a word-count-based visibility metric and 15-30% on a subjective-impression metric ([Aggarwal et al.](https://arxiv.org/abs/2311.09735)). Keyword stuffing didn't perform well. Results varied by domain and come from a research setup, so treat them as direction, not a guarantee for any engine.
Figures below are quoted from the paper (arXiv 2311.09735), accessed 30 Oct 2026. Recommendations marked "our view" are interpretation.
What did the GEO paper test?
The paper introduced Generative Engine Optimization (GEO) and GEO-bench, a benchmark of diverse queries across domains with web sources (Aggarwal et al.). It modified source content using nine methods and measured how visible each source became in generated answers.
| Method | What it changed |
|---|---|
| Authoritative | More persuasive, authoritative style |
| Statistics Addition | Quantitative statistics instead of qualitative discussion |
| Keyword Stuffing | More query keywords |
| Cite Sources | Relevant citations from credible sources |
| Quotation Addition | Relevant quotations from credible sources |
| Easy-to-Understand | Simpler language |
| Fluency Optimization | More fluent text |
| Unique Words | Added unique words |
| Technical Terms | Added technical terms |
What were the results?
From the paper (Aggarwal et al.):
- Top methods: Cite Sources, Quotation Addition and Statistics Addition, with 30-40% relative improvement on Position-Adjusted Word Count and 15-30% on Subjective Impression.
- The best methods improved on baseline by 41% and 28% on those two metrics.
- Keyword stuffing didn't perform well against baselines.
- Overall, GEO could boost visibility by up to 40%, and efficacy varied across domains.
What are the limits?
Our view:
| Limit | Why it matters |
|---|---|
| Research benchmark | Live engines retrieve and rank differently, and change often |
| Visibility metrics | Measure how much of a response draws on a source, not clicks or sales |
| Content modified by a model | Real pages need real data and quotes |
| Domain variation | What works for history may not work for product queries |
| Date | Published in 2023; engines have changed since |
No engine has published that it rewards statistics or quotes specifically. Google says its systems aim to reward helpful, reliable, people-first content (Google).
Why might these methods help?
Our view: engines that quote or cite sources (Claude says responses can include relevant quotes (Anthropic); Perplexity cites numbered sources (Perplexity)) need specific, attributable sentences. A sourced number or a named expert quote is easier to lift than general claims.
How do you add statistics honestly?
- [ ] Use only numbers from a named, verifiable source, or your own data with method stated
- [ ] Date every statistic
- [ ] Link the source next to the number
- [ ] Prefer primary sources over articles quoting them
- [ ] Never round or "adjust" numbers to sound better
- [ ] Remove stats that are out of date
Google's spam policies cover scaled content created to manipulate rankings (Google). Padding pages with unsourced numbers is low value and risky.
How do you add quotations well?
- Quote named people with relevant expertise
- Get permission for interview quotes
- Keep quotes short and specific
- Attribute clearly with role and date
Illustrative example: instead of "Experts agree onboarding matters," write "'Most churn we see happens in the first 30 days,' says [named head of customer success at your company], based on [stated dataset]." Only publish if the data exists.
Where do you find trustworthy statistics?
Our view, in order of preference:
- Your own data, with sample size, period and method stated (see Original research)
- Official statistics from government agencies and standards bodies
- Peer-reviewed or preprint research, linked to the paper itself
- Company documentation for product facts (pricing, specs, limits), dated
- Reputable surveys that publish their methodology
Avoid statistics that appear only in listicles or "stats roundup" posts without a traceable original source. If you can't find the primary source, don't use the number.
How do results vary by domain?
The paper reports that the efficacy of methods varied across domains and calls for domain-specific approaches (Aggarwal et al.). Our view: for product and B2B topics, verifiable specs, prices and benchmarks are the most natural "statistics"; for health, finance and legal topics, credentials and citations to authoritative bodies matter more. Test within your own domain rather than assuming a single best method.
How do you test it on your own pages?
- Pick 10 pages with similar traffic and topic.
- Add sourced statistics, quotes and citations to 5; leave 5 unchanged.
- Run the same prompt set across engines before and after (4–6 weeks).
- Compare citation share for changed vs unchanged pages.
| Page group | Citation share before | After |
|---|---|---|
| Changed | — | — |
| Control | — | — |
Template; fill with your data.
What doesn't follow from the paper?
- That any number helps (the paper used relevant statistics)
- That keyword density matters (keyword stuffing underperformed)
- That results transfer unchanged to every engine or domain
- That more statistics are always better; relevance matters
Checklist
- [ ] Sourced, dated statistics
- [ ] Attributed expert quotes
- [ ] Citations to primary sources
- [ ] Controlled before/after test
- [ ] No invented or padded numbers
Related
Sources
- Aggarwal et al., GEO: Generative Engine Optimization (arXiv 2311.09735)
- Google, Creating helpful content; Spam policies
- Anthropic, Claude web search; Perplexity, How does Perplexity work?
All accessed 30 Oct 2026.
Questions
Do statistics raise AI citation rates?
In the GEO paper's benchmark, adding statistics, quotations and citations were among the top methods, improving a visibility metric by 30-40% relative to baseline. Results varied by domain and were measured on a research setup, not live engines.
Does keyword stuffing work for AI engines?
The GEO paper found keyword stuffing didn't perform well compared with its baseline.
Should I add statistics to every page?
Only real, sourced statistics. Invented or unsourced numbers damage trust, and Google's guidance favours reliable, people-first content.