Chain of Thought (CoT)

Chain of thought (CoT) is a sequence of intermediate reasoning steps that a language model produces before giving its final answer. It started in 2022 as a prompting technique and is now built into the reasoning models behind most AI assistants and AI search engines.

The technique was introduced in the paper Chain-of-Thought Prompting Elicits Reasoning in Large Language Models by Jason Wei and colleagues at Google Research in January 2022. They showed that giving a model a few worked examples with step-by-step solutions sharply improved its accuracy on math word problems and multi-step reasoning. A few months later, Takeshi Kojima and colleagues from the University of Tokyo and Google showed in Large Language Models are Zero-Shot Reasoners that simply adding 'Let's think step by step' to a prompt had a similar effect without examples. With InstructGPT (text-davinci-002), that one line raised accuracy on the MultiArith benchmark from 17.7% to 78.7% and on GSM8K from 10.4% to 40.7%.

Researchers then built variants on the idea:

• few-shot CoT, which shows worked examples
• zero-shot CoT, which only asks for step-by-step reasoning
• self-consistency, which samples several reasoning paths and picks the most common answer
• tree of thoughts, which explores and scores branching lines of reasoning

The larger shift came when labs began training models to reason this way on their own. Reasoning models, starting with OpenAI's o1 in 2024 and followed by DeepSeek-R1 and others, are trained with reinforcement learning to produce long internal chains of thought before answering, spending extra test-time compute on harder problems. By late 2026 this is the default: GPT-6 in ChatGPT can interleave thinking with answering and start responding before its reasoning is finished, and Anthropic's Claude Opus 5.5 runs with adaptive thinking that cannot be switched off. Most products show a summary of the reasoning, not the raw chain. Explicitly asking a reasoning model to 'think step by step' adds little, because it already does.

The chain of thought is also a safety tool. In July 2025 researchers from OpenAI, Anthropic, Google DeepMind and others published a position paper arguing that reading a model's chain of thought is a useful but fragile way to detect intent to misbehave, and that developers should avoid training choices that make it less legible.

For AI search and GEO, chain of thought shows up as the planning step. Before searching, a reasoning model works out what it needs to know and breaks the question into parts, which drives the query fan-out it then runs. While it reads results, it checks claims against each other and decides which sources support which statement. Content that states facts explicitly, shows how a conclusion follows from evidence, and answers one clear sub-question per section is easier to use in that process. A page that buries its key number in a long narrative forces the model to infer it, and the model may cite a source where the fact is stated plainly instead. Teams can see this effect in LLM citations: reasoning-heavy answers tend to cite the specific passage that supplied each step.