Why does ChatGPT give different answers?

Four sources, stacked. Sampling: a language model picks each next word from a probability distribution, and consumer assistants deliberately keep some randomness in that pick, so the same prompt takes different paths on different runs. Retrieval: when the assistant searches before answering, the pages it happens to fetch differ between runs, and the answer follows its sources. Context: anything earlier in your conversation, plus personalization, tilts the output. Updates: the model behind the product changes without notice, and yesterday's behaviour is not a promise.

If answers vary, one answer is one draw from a distribution, and a screenshot of it proves whatever the person who took it wanted. In one real scan of ours, a brand was named in 18 of 63 answers on one engine and 0 of 63 on the other two. Check once and you would call that brand visible or invisible depending on which draw you happened to catch, and either conclusion would be wrong. The honest number is a rate with its denominator: named in 18 of 63, not "appears in ChatGPT".

This is why Aureo asks every tracked question nine times a week, three times on each of three engines, and prints the n beside every metric. Not because nine is magic, but because a rate needs a denominator, and a trend needs the same denominator every week. The sample report shows what that looks like in practice.

If you are evaluating your own visibility, ask the question more than once before concluding anything, and keep the wording identical between runs. If you are evaluating a tool, ask how many answers each number rests on; a vendor that cannot answer is selling you single draws. And if an answer got your brand wrong, remember you are correcting a distribution, not a page: the fix is making the right information retrievable and quotable, which is where LLM SEO starts.

• Does ChatGPT give the same answer to everyone?
  No. Two people asking the identical question can get different answers, and the same person asking twice often does. Model updates, sampling randomness, retrieved sources and conversation context all move the output. This is by design, not a malfunction.
• Can I make ChatGPT answer consistently?
  Developers can reduce variation through API settings, but the ChatGPT, Perplexity and Gemini your buyers use ship with variation on. For anyone measuring brand visibility, the consumer behaviour is the one that matters, because that is what a buyer actually sees.
• Does this mean AI answers cannot be measured?
  They can, the same way any noisy signal is measured: repeatedly. Ask the same question several times per engine, count how often a brand appears, and report the count with its sample size. What cannot be trusted is a single screenshot, in either direction.
• Why did a tool show my brand in an answer that my own test does not show?
  Both runs can be real. One draw found you, another did not, and neither one alone says how often you appear. The number you want is a rate: named in how many answers out of how many asked. If a report does not state that denominator, it is showing you one coin flip.