Chain of thought (CoT) describes a model generating intermediate steps, such as breaking a problem down, comparing options or doing arithmetic, before stating a conclusion. It began as a prompting technique: asking a model to 'think step by step' improved results on reasoning tasks. Newer reasoning models are trained to do this by default, often in a hidden reasoning phase the user sees only as a summary.
From prompt trick to built-in behavior
A 2022 Google Research paper showed that giving models worked examples of step-by-step reasoning improved performance on math and logic problems. Practitioners found that simply asking for step-by-step reasoning helped too. Since then, model makers have trained reasoning directly into some models, which plan and check their work before answering.
Chain of thought in AI search
In search products, reasoning often shows up as planning: the model decides what it needs to know, runs several searches, reads results, notices gaps and searches again. This connects closely to query fan-out. For a buyer's question like 'which payroll tool fits a 30-person company with staff in three US states', a reasoning model may break the question into requirements and look for evidence on each one.
That favors content that states requirements and capabilities explicitly. If your page says plainly which states, team sizes and integrations you support, a model checking those conditions one by one can confirm them.
- State concrete capabilities and limits, not just benefits.
- Cover the conditions buyers actually filter on.
- Keep facts consistent across pages so checks do not conflict.
Example
Checking requirements one at a time
A prospect asks an assistant for 'an expense tool with per diem rules, NetSuite sync and SOC 2 Type II'. The assistant's reasoning summary shows it searching for each requirement. Sablecrest, a fictional expense tool, has a security page stating its SOC 2 Type II report and an integration page for NetSuite, but per diems are mentioned only in a video. Adding a short text section on per diem rules covers the missing check.
Common mistakes
Treating visible reasoning as an exact log
Reasoning summaries shown to users are often condensed, and vendors note they may not reflect every internal step.
Hiding key facts in media only
A requirement mentioned only in a video or image may not be confirmed by a text-based check.
Questions
For older or smaller models it can help. Reasoning models already do it, and vendor guidance for them often says explicit step-by-step instructions are unnecessary.
Chain of thought is the technique. A reasoning model is one trained to apply it routinely before answering.
