Log probabilities · Token probabilities on a log scale
Logprobs (log probabilities) are the probabilities, on a log scale, that a model gives to each possible token at a step of its answer. Some APIs expose them, which shows whether a brand was an obvious choice or a close call.
Standalone definition · 41 words
How to read the logprobs of an answer
At each step, the model computes a distribution over the next tokens. The OpenAI API can return, for each token, its logprob and the most likely alternatives. A logprob close to 0 means the model was almost certain.
Why logprobs matter for GEO
Metehan Yeşilyurt suggested in August 2025 using logprobs to measure AI visibility: instead of counting whether a brand appears, you check how likely it was in the slot where a brand is expected. He notes it is experimental and shares a free script.
The limits of logprobs for brand measurement
Not every API returns them, and consumer apps never do. Values change with the prompt, the model and the temperature. It is a trend indicator, not a position.
Frequently asked questions about Logprobs
Does ChatGPT show logprobs?
No, only some developer APIs do.
Is a logprob of -0.1 good?
Yes, it means about 90% probability for that token.
Terms related to Logprobs
Further reading on Logprobs
- Best GEO tools 2026: top 10
- Alternatives to GetMint for AI tracking
- ← Back to the GEO glossary (150 terms)