Key point. Could not confirm for live AI engines. The GEO paper tested edits on a lab engine that wrote every answer from the same five search results. It measured how much of the answer a page got, not whether ChatGPT or Google picks the page at all.
Verdict: could not confirm, for live AI engines. The paper is careful research, and its result holds for what it tested. The claim that travels with it says more than the paper measured: that adding quotes, statistics and citations to a page gets it cited more by the AI engines your buyers use. The paper did not test that, and Elementera has no measurement of its own that settles it.
What did the GEO paper actually test?
The GEO paper tested nine ways of editing a web page on a search engine its authors built for the study. For each question, the engine fetched the top five Google results and had an OpenAI model, gpt-3.5-turbo, write the answer from them (Aggarwal and others, GEO, arXiv 2311.09735).
The authors then edited one of the five pages and measured how much of the new answer was attributed to it. Their main measure counts the words in sentences that cite the page, weighted by where the citation appears. A second measure asks a model to rate the page's impression on the answer.
| What the claim says | What the paper measured |
|---|---|
| Live AI engines (ChatGPT, Google AI Overviews) cite your page more | A study engine built on gpt-3.5-turbo, plus a smaller test on Perplexity |
| Your page gets picked as a source | The page was already one of the five sources; the question was how much of the answer it got |
| Citations | Words attributed to the page, weighted by position, and a model's rating |
| Any question your buyers ask | A benchmark of 10,000 questions from nine sets, including real search queries and questions written by a model (GEO-bench, arXiv 2311.09735) |
What did the paper find?
On its benchmark, three of the nine edits did best: adding quotations, adding statistics and citing sources. Making the text more fluent and easier to read also helped. Stuffing the page with keywords gave little or no gain.
The authors also report that the gains varied by subject area, and they call for methods suited to each domain (arXiv 2311.09735). In a smaller test on Perplexity, they uploaded the source pages as files because Perplexity does not let a user choose the sources by address. Adding quotations did best there too, and keyword stuffing did worse than the unedited page.
Where does the claim come from, and what do the docs say now?
The claim comes from the GEO paper, first posted in November 2023 and presented at KDD 2024. Google's documentation for its own AI features asks for no special step of this kind.
| Source | What it says | Published | Read |
|---|---|---|---|
| Research paper: Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, GEO: Generative Engine Optimization | Tested nine page edits on a study engine and a benchmark of its own; quotations, statistics and cited sources did best. | November 16, 2023 (revised June 28, 2024; KDD 2024) | October 22, 2026 |
| Current documentation: Google Search Central, AI features and your website | "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." | updated December 10, 2025 | October 22, 2026 |
| The common claim | "Add quotes, statistics and citations and AI engines will cite you a fixed percentage more." | repeated widely | October 22, 2026 |
What the common claim gets wrong. It turns a lab result about pages that were already sources into a promise about being picked by live engines. Being picked comes first: an engine has to fetch, index and choose your page before any edit to its wording can matter.
How far does the result reach?
The result reaches pages that an answer engine already reads for a question. For those, the paper is good evidence that a quoted expert, a sourced figure and a named source give the engine more to attribute to the page.
It does not reach the step before: whether ChatGPT, Gemini or Google's AI Overviews choose your page from everything they could read. It also does not show that any figure helps. Our own test found that an invented two-decimal figure could fool a simple automated novelty score, which is why we read "add statistics" as "add true, sourced statistics" (precise statistics make content look new to AI).
What should you do instead?
Start with the step the paper did not test, and then use its findings on pages that are already read.
- Check that AI crawlers can read the page: the AI crawlers directory lists the user agents and what each one does.
- Answer the buyer's question in the first lines of the page, under a heading that asks it.
- Quote a named person only when they said something worth quoting, and name them.
- Add a figure only when it is true and you can name its source; drop any figure you cannot source.
- Link the source of every figure and quotation.
- Ask your buyers' questions in the engines they use, more than once, before and after a change (why one check is not enough).
Work with us
Book a call about Originum or the Supervised plan, our managed AEO service, in which Elementera does the work for you: Book a call. Or start with Lenz: its free report shows how ChatGPT and Gemini answer your buyers' questions: Get your free AI visibility report.
More claims we tested.