Whether a text shows up in an AI answer depends measurably on how it is built. The Princeton study on Generative Engine Optimization compared nine methods. Four work, one does harm. We use the ones that work.
When an AI system answers a question, it reads several pages and builds an answer from them. Only some of those pages are named as a source. Whether the others fed into the answer cannot be seen from the outside.
What gets cited is what can be cited. In the measurement, simply adding evidenced statistics raised visibility clearly, while promotional paragraphs contributed nothing. A statement with a source can be checked, a claim without one cannot.
This is not a guess. A research team around Princeton University measured nine editing methods across 25 subject areas and published which ones work and which do not.
Baseline without editing: 19.3 points. The measure is how much of a source's text flows into the answer and where in the answer the source is cited.
Statements from experts, studies or people involved, in their original wording, with name and origin. The strongest method measured, around 41 percent above the baseline. For this we work with what is already being said inside your company.
Concrete values instead of vague claims. Not “much faster”, but “from 4.1 to 1.3 seconds”. Every number gets a source and a collection date so it stays verifiable, in the AI answer as well as in a sales conversation.
References to sources that exist independently of you: standards, public authorities, industry associations, studies. According to the study, this method scores highest for statements, facts, and legal and government topics.
Well-written texts reach the same range as source references, here even a tenth of a point higher. Simply making the language simpler was measured separately and reaches 22.0 points. The authors conclude that generative systems also judge presentation, not just content.
The only method tested that ends up below the baseline. What long counted as a lever in classic search engine optimization works against you in this measurement. We take it out of texts instead of putting it in.
According to the study, a deliberately confident style brings no notable improvement. The authors note that generative systems are resistant to such changes in style. Evidenced content beats tone.
The study breaks down the effect by the source's search position, and there the sign flips. For a page that already ranks first, all three methods lower visibility in the AI answer: source references by 30.3 percent, quotations by 22.9 percent, statistics by 20.6 percent.
For a page in fifth place, it is the other way round. There, visibility rises by 115.1 percent with source references, by 99.7 percent with quotations and by 97.9 percent with statistics, more than doubling. The authors sum it up as: the method helps lower-ranked pages in particular.
That is why we do not apply the process to an entire website. We first check where each page stands today and start where the study shows a gain. For pages that already rank at the very top, we measure before and after instead of rebuilding blindly.
For each page we determine its current position and filter out what needs no changes. What remains are the pages whose position falls in the range where the study shows a gain.
Numbers from your business, statements from your team, sources from outside. This is the time-consuming step, and there is no shortcut. Invented numbers are out of the question, both professionally and legally.
The content is reworked, and every statement gets a source and a date. The page is then included in our ongoing measurement, so you can see whether it gets cited more often and further up.
The study measures how prominently a source appears in an AI answer. It does not measure visitors. More mentions do not automatically mean more inquiries, because a complete answer can make visiting the website unnecessary. If you use the method, you should see the mention itself as the value.
Which questions people actually ask, and where you are missing from the answers, is clarified beforehand by prompt mapping. If an answer says something wrong about you, that is a case for brand sentiment and reputation. The bigger picture is described on our GEO and LLMO support page. The do-it-yourself steps are in the Swiss GEO Framework (PDF in German).
Citation engineering is part of our GEO support, from CHF 1,800 / month. 12 months: from CHF 1,710 / month (5% discount).
Reworking content in a targeted way so that AI systems use it as a source instead of just reading it. Three changes are proven to work: your own numbers with a source, direct quotations from experts and references to solid evidence. The Princeton study measures 30 to 40 percent more visibility in AI answers for these.
On “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD 2024. Three of the six authors are affiliated with Princeton University. Nine methods were tested across 25 subject areas.
No, it does harm. In the Princeton measurement, keyword stuffing reaches 17.7 points compared with 19.3 points for unchanged content. It is the only method tested that performs worse than doing nothing.
Hardly. The study finds no notable improvement for a deliberately authoritative style and notes that generative systems are resistant to such changes in style. What works is evidenced content, not tone.
No. The study breaks down the effect by search position. For pages in first place, visibility in the AI answer drops by 20 to 30 percent; for pages in fifth place it rises by 98 to 115 percent. That is why we use the method selectively on lower-ranked pages.
Not every page benefits, some lose. We tell you which ones qualify, free and without obligation. You can work with us in English.