Just a few years ago ChatGPT was a very specific kind of tool. Today, it’s blossomed. ChatGPT is now a more generalised service - people use it for all manner of tasks. Increasingly, that includes as a search engine.
Instead of Googling “best digital marketing company Australia”, you might just as well ask ChatGPT “what’s the best digital marketing agency for my business?”.
More and more of us are using ChatGPT this way. That means two things: 1) users expect more direct, personalised answers, and 2) businesses must adapt their digital marketing strategies.
In order to really hack GEO - that’s Generative Engine Optimisation - businesses need to fully understand how generative engines like ChatGPT and Gemini really work. So let’s find out!
How LLMs “Retrieve” Information
The first thing to understand is that generative engines don’t retrieve data the way normal search engines do.
LLMs are far more complex. What happens when you ask ChatGPT a question? Here’s a simple breakdown:
Embedding - text is converted into high-dimensional vectors (embeddings) that capture semantic meaning. Similar ideas end up with similar vectors.
Vector search - embeddings are compared against millions of vectors using approximate nearest-neighbour search.
Chunking - source content is split into “chunks”. The engine assesses these chunks for relevance and credibility.
This all makes GEO very different from SEO. It’s not about keywords. It’s about semantic matches.
How Models Cite and Summarise
Generative engines aren’t perfect. They don’t explicitly retrieve information. Rather, they mimic citation patterns.
When a model cites something, it’s based on:
Pattern learning - models learn the statistical structure of citations (e.g., APA, URLs).
Associative recall - models look for specific facts often paired with source patterns.
The model then summarises information in line with the query using:
Salience - the model internally estimates which tokens or concepts are most “informative” based on attention weights.
Abstraction - it compresses meaning into fewer tokens by predicting the most likely “high-information” continuation.
Hallucination
We’ve all been there: ChatGPT spits out a “fact” that simply isn’t true. These “hallucinations” are a problem; they emerge from the model’s next-token prediction nature.
Here’s why:
Missing info - even when the engine lacks a fact it must still produce a continuation. So, it basically invents something.
Overfitting to patterns - models sometimes borrow structures from similar patterns (even when the context is different).
Weak retrieval - retrieval might fail, but the model will still reply based on prior statistical associations and guesswork.
What Every Digital Marketer Needs to Know About Generative Engines
Looking to optimise your business for generative engines? It’s crucial to understand that models don’t choose citations through factual lookup. They follow patterns and retrieval quality. Here’s what that means in practice:
Alignment with Structured Data
LLMs love structured data. No messy prose. No fluff. Semantic meanings are easier to capture when the source is chunked and well-labeled. Formats that work include:
Tables
Bullet lists
Step-by-step instructions
Glossary-style definitions
Tight key–value patterns
Schema-rich data
How Citations Are Influenced
We know that generative engines cite information through pattern learning and associative recall. But what exactly influences this process when it comes to the source text? Here’s a breakdown of aspects you should consider:
Presence of retrieval - the model is more likely to cite what it sees in the prompt and the structure and markup of documents the system has injected.
Formatting cues - models usually mimic citation formats and reference structures it sees in the source text.
Prominence & repetition - LLMs love attributions that repeatedly appear paired with source patterns seen across the web.
Query pressure - models will always try to answer the query. So, if that query says “cite sources”, for instance, it will - hallucinatory or not.
How “Answer Quality” Is Evaluated
LLMs can’t access “objective truth” any more than we can. Instead, they use model-internal priors and instruction-tuning reinforcement. This is extremely important for GEO. Optimisers must understand that models use:
Reward models - models recognise helpfulness, accuracy, completeness, clarity, and groundedness.
Internal coherence signals - models prefer answers that match common QA patterns and maintain logical structure.
Retrieval alignment - answers that match retrieved docs equal a high score. Poor retrieval leads to hallucinations.
Compression efficiency - models favour source text that gives important info early on and compresses more meaning into fewer words.
Confidence priors - just like you, models prefer answers that are structured, authoritative, clean, and validated.
Final Thoughts
Generative engines work very differently to traditional search engines like Google. Some retrieval aspects are similar (Google’s E-E-A-T - Experience, Expertise, Authoritativeness, Trustworthiness), so you should still aim to follow these principles in all content.
But other factors are different. Generative engines don’t use keywords. They use semantic meanings and a combination of complex internal training patterns.
Still have questions? Reach out to Salt & Fuessel. Our GEO experts are here to help your business succeed in the emerging generative engine landscape!
