Why LLMs cite some sources and not others
Large language models learn from their training data, but in practice they also search in real time: Perplexity, ChatGPT with browsing, Gemini. What decides whether they cite a source is the perceived authority of the domain, how clear and specific the answer is, the format of the content, and the presence of structured data that says plainly who the entity is and what it does.
What matters is not how much content there is but how much useful information each paragraph carries. A 600-word article that answers directly in the first three sentences gets cited more than a 3,000-word article with no clear structure.
The 7 GEO techniques that work
These are the levers with the greatest impact on visibility in AI engines, in order of return:
- A direct answer in the first two or three sentences. Models extract the most informative part of the text. If the article opens with storytelling or context, the useful answer never gets cited.
- FAQPage schema. Implement schema.org/FAQPage with explicit questions and answers. It's the format LLMs process best for extracting definitions.
- A well-defined Organization schema. Use schema.org/Organization so models can work out who you are, what you do and why you have any authority on the topic.
- Semantic lists and tables. Content in `
- ` and `` is more citable than dense paragraphs. Lists faked with dashes in plain text do not send the same signal.
- Headings written as questions. "What is X?" as an H2, followed by the definition in the very first paragraph.
- Data and statistics with sources. Specific figures with attribution get cited more than generic claims, for a simple reason: a model can attribute "43% according to X", but it cannot attribute "many companies".
- Canonical, readable URLs. Models with browsing have to identify the source URL. Keep URLs clean, with no parameters and the keywords in the path.
GEO vs. SEO: what changes and what stays the same
GEO doesn't replace SEO, it extends it. Ranking on Google is still a priority, because many models with real-time search lean on sources that already rank well. What changes is the format of the content: instead of optimizing for keywords and backlinks, you optimize for extraction and attribution.
Good GEO content has everything a good Google snippet has (a direct answer, kept short, backed by data) plus explicit structured data and a clearly identified authoring entity.
How to measure whether your GEO strategy is working
Unlike SEO, there is no official index of visibility in language models. The practical metrics are these: regular manual tests on ChatGPT, Gemini and Perplexity using your business's key queries; brand mention monitoring on social listening tools; and referral traffic from AI platforms, which starts showing up in Google Analytics with chatgpt.com or perplexity.ai as the referrer.
At Elevatec we run a monthly test protocol: 15 target queries across 4 AI engines, logged in a tracking sheet with the date, the exact query, the engine and the result. It's the closest thing to a rank tracker for GEO that exists today.
Frequently asked questions
Results in language models come more slowly than in SEO, because they depend on re-indexing or on the model's training cycle. With a well-executed content strategy, the first results usually show up within 2 to 4 months. Improvements in engines that search in real time (Perplexity, ChatGPT with browsing) can appear within weeks.
Both. The first priority is normally the content you already have: adding FAQPage schema, rewriting opening paragraphs so they answer directly, and adding specific figures with attribution. New content then reinforces the entity's authority on the topic.
GEO is especially powerful for small companies with narrow expertise. A niche covered properly by a small company is more likely to be cited than generic content from a big brand. Specificity and topical authority matter more than company size.