Cluster 2 · How AI works

2.2 RAG explained

RAG — Retrieval Augmented Generation — is the most important technical concept for GEO outside of static training data. It determines whether an AI system can retrieve and use your content at the moment a user asks a question.

What is RAG?

When a user submits a prompt, the system first searches a data source for relevant passages. Those passages are supplied as additional context, and the model generates an answer based on both its training knowledge and the retrieved information.

One thing up front: what follows is a simplified view of one widely used RAG setup. Real systems differ per provider and combine techniques — alongside vector search also classic keyword search, query rewriting, multiple searches and filters. And not every system fetches pages live: often an existing index or cache is queried first, and only sometimes is a page fetched at that moment. The six phases show the principle, not the implementation. The example is also fictitious: the suppliers named exist, but the scores, tariffs and conditions are invented to show the mechanism.

Click through the six phases to see what happens to a concrete Belgian energy question:

1. The user asks a question

Input phase

A user types a question into an AI assistant. Unlike a search engine, the user doesn’t need to think in keywords — they write the way they would speak. This is what the system has to interpret in step 2.

Which energy supplier currently offers the best rate for a small business in Flanders?

2. The question becomes a vector

Embedding phase

The system converts the question into a mathematical representation — a long sequence of numbers that captures the meaning of the sentence. Questions with similar meaning get similar vectors, even if they use different words. This makes semantic search possible.

[ 0.184, -0.291, 0.503, 0.072, -0.118, 0.447, -0.226, 0.339, 0.061, -0.412, 0.288, 0.156, -0.073, 0.394, … + about 1500 more dimensions ]

3. The system retrieves candidates

Retrieval phase

The vector of the question is compared against vectors of thousands of documents in the database. The system surfaces the most similar candidates. At this point these are still raw matches — relevance is estimated on similarity, not on factual quality. In this example the search step runs on vectors; many systems combine that with classic search.

Engie — rates for small professionals

0.851

Luminus — SME energy contracts

0.823

VREG — comparison module 2026

0.798

TotalEnergies — pro offer Belgium

0.776

Old forum thread from 2021

0.692

4. Candidates are re-ranked

Ranking phase

A second layer of filtering. Which signals the system weighs here differs per provider — recency, authority and specificity are common candidates, not a fixed formula. A small number of passages makes it to the next phase. This is where many candidates drop out that looked relevant at similarity level.

VREG — comparison module 2026

recent + neutral

Engie — rates for small professionals

specific + recent

Luminus — SME energy contracts

specific

TotalEnergies — pro offer Belgium

commercial

Old forum thread from 2021

outdated

5. The context is assembled

Context phase

The selected fragments are packaged together with the original question into one prompt. This is what the language model finally sees: your question plus the external knowledge just retrieved. The model does not have to pull the answer from memorised knowledge — it is handed building blocks and combines them with what it already knows.

SYSTEM: Answer based on the following sources.

CONTEXT:
[1] VREG comparison module (Apr 2026): ...
[2] Engie rates for small professionals: ...
[3] Luminus SME energy contracts: ...

USER: Which energy supplier currently offers the best rate...

6. The model generates an answer

Generation phase

Only now does the language model itself come into play. It synthesises the retrieved fragments into a readable answer and — in citation-first systems like Perplexity — adds explicit source references. For GEO this is the moment that reveals whether your content was retrieved, how deeply it is described, and where it sits in the answer.

For small businesses in Flanders, Engie currently offers one of the sharpest fixed rates, followed by Luminus and TotalEnergies. [1][2][3]

The conditions vary per supplier — Engie has a fixed price for three years, Luminus offers a green option with no surcharge, TotalEnergies has a discount for multi-product customers.

How does RAG select which content gets retrieved?

RAG systems search on meaning: the user’s question is compared with passages that express the same thing, even when the words differ.

Practical consequence: content written from the user’s question — not from internal terminology — sits closer to how these systems search. That makes retrieval more likely, though what ultimately happens depends per system on more than similarity alone.

Writing RAG-friendly content

These are editorial best practices — they increase the odds of being retrieved and cited, they guarantee nothing: