Article
Why AI assistants give different answers each time
Ask ChatGPT the same question twice and you can get two different businesses. Why that happens, and what it means for measuring your visibility.
Published 2 min readBy Visibility
Short answer
AI assistants do not look answers up in a fixed table. Each reply is generated fresh, with a degree of chance built in, and may draw on the model's memory, on a web search, or both. A different model version, a different search result or simply a different roll of the dice can change which businesses are named.
On this page
Why is each answer different?
A language model writes its answer one word at a time, choosing each from a range of likely options. Assistants deliberately do not always take the single most likely word, because answers that do tend to be repetitive. The result is that the same question can produce a different answer, including a different list of businesses.
What else changes the answer?
- Web search or memory. An answer that searched the web reflects the pages it found that time; an answer from memory reflects the model's training. They often name different businesses.
- The model version. Assistants are updated regularly. A new version can know different things and answer differently.
- Wording. "Best restaurant in Gibraltar" and "where should we eat near the cruise terminal" are different questions to an assistant.
- Personal context. Your chat history, saved memories, settings and location can shape what you are told.
What does this mean for my business?
It means one answer is not a result. If you appear once, you may not appear the next time; if you are missing once, you may be there tomorrow. Visibility is a rate, not a yes or no: how often you are named across many askings.
That is why we ask every question to every assistant on a fixed schedule, record the model and date, and report how often each business appeared. See measure your AI visibility and our methodology.
Common questions
Is the assistant making mistakes when answers change?
Not necessarily. Several different answers can all be reasonable. Variation becomes a problem when it includes wrong facts, such as a closed business or an old address.
Can I make the answer more consistent?
For your own chats, turning on web search and asking specific questions helps. For how assistants describe you to others, the lasting route is making sure the web describes you clearly and consistently.
How many times should a question be asked to measure it?
Enough to see a pattern rather than a single draw, across several assistants, repeated over time. Our methodology page sets out exactly what we do.