---
title: "Do LLMs Use E-E-A-T? What ChatGPT and Gemini Apply"
date: "2025-01-07"
author: "Flavio Longato"
categories: ["GEO"]
url: "https://www.longato.ch/does-llms-like-chatgpt-gemini-use-eeat/"
---

**No**, large language models (LLMs) do **not** explicitly use E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) to evaluate or rank information in the way Google does for search results.

What it *does* apply are alignment criteria such as *helpful, honest, harmless* (RLHF or Constitutional-AI rules). Those criteria overlap with E-E-A-T’s “trust” dimension, but they are not the same thing.

 &lt;figure class=&quot;wp-block-table&quot;&gt;&lt;table class=&quot;has-fixed-layout&quot;&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Phase&lt;/th&gt;&lt;th&gt;Method&lt;/th&gt;&lt;th&gt;E-E-A-T relevance&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;**Training**&lt;/td&gt;&lt;td&gt;Predict next word from massive public text data&lt;/td&gt;&lt;td&gt;No E-E-A-T logic involved&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;**Inference**&lt;/td&gt;&lt;td&gt;Generate answers based on prompt + internal weights&lt;/td&gt;&lt;td&gt;Doesn’t score credibility like Google does&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;**RAG pipelines**&lt;/td&gt;&lt;td&gt;Retrieve text chunks, then generate answer&lt;/td&gt;&lt;td&gt;Retrieval system may favor E-E-A-T-like content&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;**Reinforcement tuning (RLHF)**&lt;/td&gt;&lt;td&gt;Human raters judge helpfulness, honesty, etc.&lt;/td&gt;&lt;td&gt;“Trustworthiness” overlaps but isn’t E-E-A-T scoring&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;

&lt;/figure&gt;### Where E-E-A-T *does* influence LLM output indirectly:

- **Source selection in RAG (Retrieval-Augmented Generation)**: When an LLM is used in tools like Perplexity, ChatGPT with browsing, or Claude with documents, it pulls content from sources that search-like algorithms rank. These upstream systems **do use** E-E-A-T-style heuristics (authority, backlinks, freshness).
- **LLMs trained on curated corpora** (e.g. academic, government, or high-authority sites) inherit some quality biases. Pages with high E-E-A-T are more likely to be in those datasets.
- **Answer citation models** prefer concise, clearly authored, fact-dense paragraphs — a byproduct of E-E-A-T optimization.
 
What is RAG?
------------

### RAG = **Retrieval-Augmented Generation**

In plain terms, it’s a two-step architecture that lets a language model *look things up* before it writes the answer—so it can stay factual even when its frozen training data is out of date.

### How RAG works

 &lt;figure class=&quot;wp-block-table&quot;&gt;&lt;table class=&quot;has-fixed-layout&quot;&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Stage&lt;/th&gt;&lt;th&gt;What happens&lt;/th&gt;&lt;th&gt;Typical tech&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;**Retrieve**&lt;/td&gt;&lt;td&gt;A search component (vector DB, BM25, hybrid) finds the most relevant text chunks for the user’s query.&lt;/td&gt;&lt;td&gt;FAISS, Pinecone, Weaviate, Elasticsearch k-NN&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;**Augment**&lt;/td&gt;&lt;td&gt;Those chunks are concatenated with the original prompt to form an expanded context window.&lt;/td&gt;&lt;td&gt;Prompt templating&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;**Generate**&lt;/td&gt;&lt;td&gt;A language model reads that *prompt + evidence* and produces the final answer (summary, chat reply, SQL, etc.).&lt;/td&gt;&lt;td&gt;GPT-4o, Claude 3, Llama-3-70B-Instruct&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;

&lt;/figure&gt;### RAG = **Retrieval-Augmented Generation**

In plain terms, it’s a two-step architecture that lets a language model *look things up* before it writes the answer—so it can stay factual even when its frozen training data is out of date.

**Retrieval-Augmented Generation turns an LLM into a live, citation-backed knowledge assistant.** Mastering the retrieve-rank-generate loop—and the guardrails around it—is now core to any production-grade LLM application.

What is RLHF?
-------------

**RLHF = Reinforcement Learning from Human Feedback**. It’s the training phase in which a language model learns *how* it should answer—“helpful, honest, harmless”-style—by optimizing against scores that come from humans rather than a fixed algorithm.

Large language models that are trained only to predict the next token often:

1. **Produce unsafe or off-topic content.**
2. **Ignore user intent** (rambling or hallucinating).
3. **Lack a consistent “voice.”**
 
RLHF adds a final alignment layer so the model’s behaviour matches human expectations of usefulness and safety.

 &lt;figure class=&quot;wp-block-table&quot;&gt;&lt;table class=&quot;has-fixed-layout&quot;&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Step&lt;/th&gt;&lt;th&gt;What happens&lt;/th&gt;&lt;th&gt;Key artefact&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;**1. Supervised fine-tuning (SFT)**&lt;/td&gt;&lt;td&gt;Humans write exemplar dialogues. The base model is fine-tuned on these *instruction–response* pairs.&lt;/td&gt;&lt;td&gt;*SFT model*&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;**2. Reward-model training**&lt;/td&gt;&lt;td&gt;Annotators read pairs of candidate answers from the SFT model and pick the better one. A smaller “reward model” is trained to predict those preferences.&lt;/td&gt;&lt;td&gt;*Reward model (RM)*&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;**3. Policy optimisation**&lt;/td&gt;&lt;td&gt;Using RL (usually Proximal Policy Optimisation, PPO) the SFT model is updated. It generates answers, the RM scores them, and gradients push the policy toward higher-scoring outputs.&lt;/td&gt;&lt;td&gt;*RLHF-aligned model*&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;

&lt;/figure&gt;