What Is Chain-of-Thought in LLMs?

Chain-of-thought (CoT) is the practice — and now built-in capability — of having a large language model reason through a problem in explicit intermediate steps before stating its answer. Introduced as a prompting technique by Wei et al. (Google, 2022), it began as the phrase "let's think step by step" and evolved into the architectural foundation of reasoning models like OpenAI's o1 and DeepSeek-R1.

From prompt trick to model architecture

The 2022 finding was that step-by-step prompting unlocked reasoning that direct questioning missed — but only in large models, an emergent-scale effect. By 2024, labs stopped relying on users to ask for reasoning and trained models to generate extended chains of thought natively, rewarded via reinforcement learning for chains that reach correct answers. Today's reasoning models produce thousands of internal thinking tokens; some engines display a summarized trace, others hide it entirely.

Why reasoning traces are GEO intelligence

When an engine shows its thinking on a commercial query, it is publishing its decision process. A trace for "best project management tool for a remote agency" might reveal steps like: identify agency-specific requirements → check client-access features → compare pricing at 20 seats → verify integration with common agency stacks. Every step is an intermediate query, and each one is a retrieval opportunity. Content mapped to those intermediate questions — a pricing-at-scale page, an integrations matrix, a client-portal explainer — enters the reasoning chain where a generic "why choose us" page cannot.

This is the observable layer of query fan-out: fan-out happens in the search calls, chain-of-thought shows the logic driving them. Reading traces across your category's key prompts is one of the cheapest research methods in GEO — it converts the engine's own reasoning into a content backlog, a practice covered in the GEO optimization guide.

Caveats worth knowing

Displayed traces are often summaries, not the raw chain, and research has shown stated reasoning does not always match the computation underneath. Treat traces as directional evidence about criteria and sub-questions — strong enough to prioritize content, too weak to reverse-engineer a ranking formula.