Why do AI assistants give different answers each time I ask?
Short answer
Because they write each answer fresh instead of looking it up in a fixed list. The text is generated with some built-in randomness, and when the assistant searches the web it may read different pages each time. Your wording, the place you name, your chat history and the product version also change the result. Variation is normal.
AI assistants give different answers each time because every answer is written fresh, with some randomness built in, and not fetched from a stored list. A search engine works from an index and an ordering that are fairly steady from one minute to the next. An assistant composes a new piece of text for each request. Ask twice and you get two compositions. For a business owner this is the single most useful thing to understand before reading anything into one reply.
The answer is generated, not looked up
The variation starts with how language models produce text. A model builds its reply a piece at a time, each time choosing among several likely continuations. That choice is usually made with a degree of randomness, which is deliberate: it keeps the writing from being stiff and repetitive. The side effect is that the same question can lead down slightly different paths. When the question is “who should I call for a slab leak in Oceanside” and five companies are plausible, different runs can name different ones or put them in a different order.
Searching adds a second source of variation
When an assistant searches the web to answer, the pages it reads can differ from run to run. It may phrase its search differently each time and get different results back. Of those, it reads a different handful. The web itself also moves: a directory page is updated, a new review appears, a list is republished. So two runs an hour apart can be built from different material, even before the wording varies.
An assistant can also answer from training alone, without searching. Those answers vary too, because the text is still generated with some randomness, but they draw on older material and tend to change less from one week to the next.
Things on your side that change the answer
Some of the variation comes from the person asking. If you’re testing, these are the ones to hold steady:
| What varies | How to control it in a test |
|---|---|
| The wording of the question | Save the exact prompts and reuse them word for word |
| The place named, or your own location | Name the city in the prompt every time |
| Earlier messages in the same chat | Start a new chat for each question |
| What the product remembers about you | Use a temporary or signed-out mode where one is offered |
| The assistant and the version of it | Record which one you used and the date |
You can’t remove the randomness inside the model. You can remove most of the rest, and what remains is closer to what a stranger would see.
What it means for your business
You don’t have a ranking in an AI assistant. You have a rate. A hypothetical plumber in Oceanside might be named in seven of twenty runs for one question and in none for another. That’s a more honest description than “ChatGPT recommends us” or “ChatGPT ignores us”, and it’s the only kind of number that can show whether a change helped. If the rate moves from seven in twenty to fourteen in twenty across the same prompts, something changed. If it moves from seven to eight, nothing did.
That’s how I set up AI visibility tracking: fixed prompts run repeatedly, with a count of how often you’re named and which sources are cited. The practical follow-ups are how I measure AI visibility in more detail and how often AI visibility should be checked.
What it means for a screenshot
A single screenshot proves very little, in either direction. If a salesperson shows you one answer in which you’re missing, that may be real or it may be one unlucky run. If an agency shows you one answer with their client at the top, the same applies. The answer isn’t fixed, even for the company that built the assistant, which is also the reason nobody can guarantee an AI recommendation. So ask how many times the prompt was run and what the count was.