Method
Our method: GEO built on published research
Every recommendation rests on four published studies of how AI assistants pick their sources: GEO (KDD 2024), GEO-16 (2025), AutoGEO (2025) and the study on manipulating language models (2024). This page sets out what each found, what we do because of it, and what it does not prove.
What did the GEO study find?
The GEO study, published at KDD 2024, tested nine ways to edit text on 10,000 queries and measured how much each source appeared in the answer1. The main results against unedited text:
| Method | Change in visibility |
|---|---|
| Add quotations from credible sources | +41% |
| Add statistics | +30% |
| Improve fluency | +28% |
| Cite sources | +27% |
| Keyword stuffing | −8% |
What we do: every key page has a quote from your expert, specific numbers with sources, at least two credible sources, short clear sentences and no repeated keywords.
What is the GEO-16 framework?
GEO-16 scores a page on 16 pillars, from metadata and freshness to internal links and transparency. The researchers analysed more than 1,700 citations from Brave, Google AI Overviews and Perplexity, and found that pages with a score of 0.70 and at least 12 pillars were cited 78% of the time2. The pillars most linked to citation were freshness (r = 0.68), semantic HTML (0.65) and structured data (0.63).
- Metadata & freshness
- Semantic HTML
- Structured data
- Evidence & citations
- Authority & trust
- Internal links
- External links
- Readability
- Claim accuracy
- Direct answers
- Transparency
- Content depth
- Visuals & media
- Engagement
- Canonical & crawlability
- Topic focus
A pillar counts as hit at 2 or more
What we do: our engine computes a GEO-16 score for every page. The paper names 14 pillars and does not publish weights, so we added two from its principles (canonical and crawlability; one topic per page) and set the weights ourselves, and we say so in every report.
Which rules do AI assistants prefer?
The AutoGEO study from Carnegie Mellon University automatically extracted the rules that make Gemini, GPT and Claude prefer one text over another, and found that most are shared by all three3. The main ones:
- Start with the conclusion.
- Attribute every fact to a credible source.
- Use specific data and examples instead of general claims.
- Keep a neutral tone with no promotional language.
- Explain “how” and “why”, and give practical steps.
- One idea per paragraph, and clear headings.
- Use up-to-date information.
“Attribute all information to credible sources and cite them clearly.”
Why don’t you use tricks?
A Harvard study showed that a calculated string of tokens, added to a product description, can push it to first place in a language model’s recommendations in about 40% of 200 runs4. That is plain manipulation: Google treats hidden text as a breach of its spam policies5, and AI assistants may stop trusting the whole site. So we do the opposite: our engine scans your site and your competitors’ for hidden text, instructions aimed at AI, invisible Unicode characters and random-symbol strings.
How do you measure without fooling yourselves?
One answer means nothing, because answers change. The GEO study collected five answers per question1; we ask every question at least three times in each assistant, show the mention rate with a 95% margin of error, and compare against the same competitors every month.
What are the limits of this research?
This research has limits you should know: most of it is on English content, some on experimental engines rather than the tools people use, and GEO-16 measures correlation, not cause. Google says there are “no additional requirements” to appear in its AI features6. So we use the research to set priorities, keep to SEO basics, and measure the real result for every client.
This site was checked with our engine before it went live. Read also how ChatGPT chooses which businesses to recommend, or audit your site.
FAQs
Does GEO replace normal SEO?
No. Google says there are “no additional requirements” to appear in its AI features, and that SEO best practices still apply6. So we start with the basics: a fast site, clear pages and a complete Google profile. Then we add the research rules and measure your visibility in each assistant.
How do you measure my visibility in AI answers?
We agree 20 questions your customers ask and put them to ChatGPT, Gemini, Perplexity and Claude, each at least three times, because the answer changes between runs. Then we calculate your mention rate with a margin of error, your position among competitors and the sites the answers cite.
Can you guarantee ChatGPT will recommend me?
No, and nobody can honestly promise that, because answers change from one run to the next and from one assistant to another. What we guarantee is the work: rules proven in research, and a monthly report showing how often you are named, who is named instead and why.
Sources
- Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024), 2024. arxiv.org
- Kumar & Palkhouski, AI Answer Engine Citation Behavior: the GEO-16 framework, 2025. arxiv.org
- Wu, Zhong, Kim & Xiong, What Generative Search Engines Like and How to Optimize Web Content Cooperatively (AutoGEO), 2025. arxiv.org
- Kumar & Lakkaraju, Manipulating Large Language Models to Increase Product Visibility, 2024. arxiv.org
- Google Search Central, Spam policies for Google web search, 2026. developers.google.com
- Google Search Central, AI features and your website, updated December 2025. developers.google.com