The short version
- The original 2023 study by Aggarwal and colleagues (opens in a new tab) reported 30 to 40% gains from adding sources, quotations and statistics. That was in a simulated engine where the page was already among five sources.
- A 2025 benchmark by Puerto and colleagues (opens in a new tab) found the Statistics method lowered a page’s rank in 19 of 24 settings tested.
- In an e-commerce test by Bagga and colleagues (opens in a new tab), eleven of fifteen hand-written rewriting recipes made product listings rank worse.
- In a 252,000-trial test by Vishwakarma and colleagues (opens in a new tab), topic match, price, a recent date and list position decided which source was cited first; formatting alone had no effect.
- A 2026 review by Martinez (opens in a new tab) of 45 studies found no technique with a stable, long-term, cross-platform effect on being found.
Where did the idea that statistics and quotes work come from?
From the 2023 study that coined generative engine optimization (GEO). Aggarwal and colleagues (opens in a new tab) built a test engine using GPT-3.5 that answered questions from the top five Google results. They rewrote one source at a time and measured how much of the answer reflected it.
Their top methods, Cite Sources, Quotation Addition, and Statistics Addition, achieved a relative improvement of 30-40% on their main visibility measure. On Perplexity, tested on 200 queries with source text uploaded as files, Quotation Addition gave a 22% improvement. Keyword stuffing, an old search trick, performed 10% worse than doing nothing.
Two caveats matter. The gains measure how much of the answer drew on a page already handed to the AI, not whether the page gets found. Martinez (opens in a new tab) notes the famous “up to 40%” comes mainly from a rise from 19.3 to 27.2 on one score. It does not mean 40% more readers or 40% more chance of being retrieved. We trace the number in our guide to the 40% GEO claim.
Did later tests confirm those gains?
Mostly no: a larger benchmark found most of these tactics did nothing or made things worse. Puerto and colleagues (opens in a new tab) tested ten rewriting methods across six domains, from retail products to news, on four AI models. They measured whether a rewritten page got cited earlier in the answer.
In their words, most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking. C-SEO is their term for rewriting content for conversational search. The Statistics method decreases rankings in 19 out of 24 evaluated settings. In product recommendation tasks on Anthropic’s Haiku 3.5 model, 26 out of 30 cases show significant negative effects.
Only two methods showed real gains, and only in narrow cases: a summary placed at the top of the page and a combined content rewrite. Even those worked only for retail and video games on one model, and none worked for question answering.
| Study | Tactic tested | Result |
|---|---|---|
| Aggarwal et al. (2023) | Statistics, quotes, citations | Gains of 30 to 40% in a simulated engine |
| Puerto et al. (2025) | Same tactics and more | Mostly no effect; statistics hurt in 19 of 24 settings |
| Bagga et al. (2025) | Fifteen rewriting recipes for product listings | Four modestly matched or beat a plain rewrite; eleven did worse |
The e-commerce study by Bagga and colleagues (opens in a new tab) used 13,747 realistic consumer product queries, each paired with 10 Amazon listings. Hand-written recipes such as “sound authoritative” or “add technical terms” did not reliably beat a simple rewrite. Whether plainer, smoother prose helps is covered in our guide on readable writing and AI answers.
Why do the studies disagree?
They measure different things, and gains for one page come at another’s expense. The 2023 study measured the share of the answer’s words tied to a page. The 2025 benchmark measured the page’s citation rank. Puerto and colleagues note that the 2023 study’s own position-weighted results showed scores generally falling.
Competition matters too. In the 2023 study, adding citations raised the visibility of the fifth-ranked source by 115.1%, while the top-ranked source lost 30.3%. Puerto and colleagues found that gains shrank steadily as more sites adopted the same method. If every competitor adds statistics, nobody stands out.
What matters more than rewriting?
Being retrieved and ranked high in the first place, and matching what the buyer asked. Puerto and colleagues found that moving a document to the top of the AI’s reading list produced far greater gains than any rewriting method.
Vishwakarma and colleagues (opens in a new tab), at the software company Sprinklr, ran 252,000 head-to-head trials across six AI models. Two product pages differed in one factor at a time. Four factors won across all six models: matching the topic, stating the price, a recent date rather than an old one, and appearing first in the list. We cover those trials in our guide to why AI cites a rival’s page.
Our own data agrees on ranking. In our study of pages cited by AI Overviews, 41.7% of pages ranking 1 to 3 were cited, against 20.1% at positions 7 to 10. Position came first; the page features we measured explained little more.
Is adding real evidence different from adding statistics?
Yes: relevant, verifiable evidence helps; inserted numbers for their own sake may not. In the Vishwakarma trials, claims without supporting evidence such as tests or certifications lost out. That is real substance, not decoration.
Martinez puts it plainly: the criterion is not to add numbers, but to provide relevant, verifiable, dated and properly attributed evidence. The automated “Statistics” rewrites in these benchmarks were generated by an AI model, so they may not reflect what real, sourced data does.
There is also a trust risk. Chu and colleagues (opens in a new tab) scanned 10,095 pages that Google Search and Gemini retrieved for real queries. They estimated 8.90% were optimized for AI engines, rising to 16.36% among pages modified in 2026. On those pages, 69.34% of citations were rated low on verifiability.
Can automated rewriting tools do better?
In lab tests, yes, but those tests are simulations. Wu and colleagues (opens in a new tab) built a system that learns what AI engines prefer and rewrites pages to match. It reported an average improvement of 35.99% on visibility measures, using simulated engines given five candidate documents each.
Bagga and colleagues found an automatically tuned rewriting instruction beat every hand-written recipe. Its rewrites turned marketing prose into clear, labeled product facts. Neither study tested live ChatGPT or Google results.
What should you do about it?
Spend on substance and findability, not on stylistic rewrites.
- Stop paying for “add stats and quotes” rewrites. The strongest benchmark found these did nothing or harmed ranking in most settings.
- Keep classic search work. Ranking high enough to be retrieved mattered more than any rewrite in the controlled tests and in our own data.
- Answer the actual question. Topic match was a gatekeeper in all six models tested. Write pages that address what buyers ask, in their terms.
- Publish facts buyers need. State prices, specifications and dates clearly. These were among the strongest factors in head-to-head trials.
- Use evidence you can stand behind. Cite real tests, data and sources. Fabricated or vague statistics risk trust if a reader or an AI checks them.
If you want help deciding where content work will pay off, see our generative engine optimization service.
What does the research not tell us yet?
No study yet shows any content tactic lifting citations on live AI search over time.
- Almost every test is simulated. The engines were built by researchers from fixed sets of documents, not live ChatGPT or Google results.
- The rewrites were machine-made. “Add statistics” meant an AI inserting statistical elements, not a company publishing its own research.
- English and narrow domains. The 2025 benchmark used English only, across six domains.
- Vendor research. The 252,000-trial study comes from a software company and tested only two pages at a time with brand names removed.
- No long-term evidence. Martinez found no reviewed technique with a stable, long-term, cross-platform effect on being found, or on clicks and sales.
Frequently asked questions
Does adding statistics to content help with AI search?
Not as a rewriting trick. In one benchmark, automated statistics insertion lowered rankings in 19 of 24 settings, though relevant, verifiable evidence may still help.
What is the “40% GEO improvement” people quote?
It comes from the 2023 GEO study’s simulated engine. It describes a larger share of the answer drawn from a page already given to the AI, not 40% more traffic.
Is GEO just SEO under a new name?
Not entirely, but ranking still matters most. In controlled tests, moving a page to the top of the AI’s reading list beat every rewriting method.
Do quotes and citations make AI trust my page more?
The evidence is mixed. Quotes helped in the 2023 simulated test but not in a 2025 benchmark; claims backed by real evidence did better in a 2026 head-to-head test.
Sources
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande (2023), GEO: Generative Engine Optimization (opens in a new tab), arXiv:2311.09735.
- Puerto, Gubri, Green, Oh and Yun (2025), C-SEO Bench: Does Conversational SEO Work? (opens in a new tab), arXiv:2506.11097.
- Bagga, Farias, Korkotashvili, Peng and Wu (2025), E-GEO: A Testbed for Generative Engine Optimization in E-Commerce (opens in a new tab), arXiv:2511.20867.
- Vishwakarma, Kumar and Jamidar (2026), What Gets Cited: Competitive GEO in AI Answer Engines (opens in a new tab), arXiv:2605.25517.
- Wu, Zhong, Kim and Xiong (2025), What Generative Search Engines Like and How to Optimize Web Content Cooperatively (opens in a new tab), arXiv:2510.11438.
- Chu, Leng, Li, Shen, Shen and Zhang (2026), GEO-Flag: Detecting and Measuring GEO-Optimized Web Content (opens in a new tab), arXiv:2608.16824.
- Martinez (2026), Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026) (opens in a new tab), arXiv:2607.14035.
- Underneath (2026), What pages cited by AI Overviews have in common