Chain of thought is when an AI model works through a problem in intermediate steps before giving its final answer, rather than jumping straight to a conclusion. It started as a prompting technique, asking the model to reason step by step, and is now built into many models that “think” before they reply.
How chain of thought works
A large language model produces text one piece at a time, and each piece is influenced by everything before it. When a model writes out its intermediate steps, those steps become part of the context it uses for the final answer. For problems with several stages, such as arithmetic, applying a set of rules or comparing options against criteria, that tends to produce more accurate results than answering in one leap.
Researchers showed in 2022 that simply including worked examples with reasoning in a prompt, or asking the model to think step by step, improved results on multi-step problems. Since then, model developers have trained reasoning models that do this automatically. At the time of writing (October 2026), most major AI assistants offer a reasoning or extended thinking mode, which may show a summary of its working or keep it hidden.
The intermediate steps use tokens, so reasoning is slower and usually dearer than a direct answer, and on simple tasks such as rewriting a sentence it adds time without improving the result.
Why it matters
Many marketing tasks people hand to AI are multi-step judgements disguised as simple requests. Sorting 500 search terms into “relevant”, “irrelevant” and “check”, working out a monthly budget split from a target cost per lead, or checking a page against a content brief all involve applying rules in order. Chain of thought, whether prompted or built in, makes those tasks more reliable.
It also helps you check the work. If a model sets out why it classified a search term as irrelevant, you can spot the faulty rule instead of guessing. One caution: research has found that the reasoning a model displays does not always reflect what actually drove its answer. Treat the visible steps as a useful explanation to check, not as proof.
Common mistakes
- Assuming a long, confident chain of reasoning means the answer is right. Models can reason fluently towards a wrong conclusion, and still hallucinate facts along the way.
- Adding “think step by step” to every prompt, including simple writing tasks, which only slows things down.
- Pasting the model’s full working into a report or a customer email when only the final answer was wanted.
- Using a reasoning model for bulk, simple jobs where a faster, cheaper model would do as well.
How to act on it
Use reasoning where a task has rules, numbers or several criteria. Write the rules out clearly in the prompt and, for repeated tasks, add two or three worked examples, which is few-shot prompting. Ask for the reasoning and the final answer in separate, labelled sections, so you can check one and use the other.
Spot-check a sample of outputs against what you would have decided yourself, and keep a person responsible for anything that affects spend or customers. If you are working out where AI can take on analysis in your marketing, and where it cannot yet be trusted, I can help with that through digital marketing strategy and consulting.