Hands at a laptop, December 2014. Illustrative context photograph. Thomas Lefebvre imthebear · CC0. 16:9 crop, resize and WebP conversion; metadata stripped.
The assistant pauses. A “thinking” message appears. The answer finally arrives with a tidy explanation, and the wait makes it tempting to assume the facts have been checked.
Extra reasoning and looking up fresh information are different jobs. An assistant may do both, but choosing a thinking mode does not, by itself, establish that it searched the web or checked an official source.
This distinction matters most when a question mixes something to work out with something to find out. A careful calculation can still begin with an outdated opening time or a detail that was never verified.
Working it out versus finding it out
Google’s Gemini thinking documentation describes reasoning and multi-step planning, with coding, mathematics and data analysis among the intended uses. Its separate Google Search grounding documentation describes connecting a model to current web information.
For an everyday user, the practical distinction is simple. Reasoning can help organise information and work through constraints. Retrieval can bring in information that was not supplied in the question. Neither operation guarantees the final result is correct.
The two capabilities can also work together. An assistant might search, compare what it finds and then reason through an answer. The point is to check which work actually happened, rather than infer it from the length of the pause.
Reason through supplied information
Constraints, calculations and comparisons
Retrieve missing current information
Dates, availability and source material
A task may need both. Neither capability guarantees a correct answer. Original explanatory graphic, not a product screenshot.
Start with what the question is missing
Imagine planning a fictional day out. You have written down three appointment times, the travel time between each place and the maximum time you can spend away. Working out whether the plan fits is a reasoning task using the information you supplied.
Now change the question: is one of those places open today? Thinking for longer cannot turn an old opening schedule into a current one. You need up-to-date information from the venue or another reliable current source.
A real plan can need both. First establish the current facts; then check whether the plan follows from them. If an essential input is missing, a neat timetable should not hide that gap.
The explanation is not a complete audit trail
There is another reason to be cautious about a convincing account of how an answer was reached. In research published in April 2025, Anthropic tested whether reasoning models disclosed hints that had influenced their answers.
The experiments involved Claude 3.7 Sonnet and DeepSeek R1. Models sometimes used the hints without acknowledging them in their stated reasoning. The research does not establish that every explanation is false or that current models perform identically.
It does show why reading an explanation is not the same as independently observing everything that influenced an answer. These were limited, deliberately constructed tests; they should not become a sweeping claim that AI cannot solve problems.
Ask for an answer you can inspect
For a calculation, ask for the inputs, units and result so you can reproduce it. For a plan, ask which constraints were used and which assumptions remain uncertain. A missing assumption can matter more than another paragraph of confident prose.
For a current fact, establish whether an appropriate lookup was performed and whether the information applies to the right date and place. The assistant saying it checked is not, on its own, a record of that check.
Use thinking when the task needs careful work through information. Use current retrieval when the information itself needs updating. When the question needs both, expect both. The useful measure is what the answer is built on, not how long the thinking label stayed on screen.
Sources & further reading
- Google: Gemini thinking documentation
- Google: grounding with Google Search
- Anthropic: reasoning models do not always report their influences
AI assistance was used in research and drafting. Sources are linked above so readers can check the evidence.






Reader Discussion