Why do ChatGPT's answers change, and when is a change real?
The Tasadr team · ·12 min read
ChatGPT and Gemini write every answer afresh, from search results that move, sometimes shaped by where the asker is and by their account. One answer is a snapshot, not a trend, and a change only counts when it repeats on different days. This guide covers why answers change, what moves and what holds, and a three-day test.
On this page (9)
Why does ChatGPT give a different answer to the same question?
Because the engine keeps no ready-made answer. It writes one afresh each time, from search results that may have changed, and it can be shaped by your location, your account's memory and model updates.
A customer in Riyadh might type «ليش إجابة ChatGPT تتغير كل مرة؟» ("why does ChatGPT's answer change every time?"). Here are the causes as the companies that run the engines describe them in their own help pages, not as we guess them.
This guide is part of how to measure the success of optimising for AI answers, which covers what to measure and how to read it.
The answer is written afresh
ChatGPT's own help page lets you regenerate a response with Try again, that is, ask for another answer to the same question. Google's Gemini help says its apps are continuously evolving and may produce inaccurate information.
Search results move
When ChatGPT needs current information it searches the web. OpenAI says it rewrites your question into one or more search queries, and may send more specific queries after reading the first results. Those results change from one day to the next.
The question branches into sub-questions
Google AI Overviews and AI Mode split a question into subtopics and search them at once, and Google says the two may use different models and techniques. A change in one sub-question is enough to change the sources, as explained in what is Google AI Mode?
Your location and your account
OpenAI says ChatGPT may use an approximate location from your IP address for local results, and, if memory is on, may draw on what it remembers about you when it rewrites a search query. So your answer in Dubai may not be your customer's answer in Riyadh.
The model itself changes
OpenAI publishes a log of its model updates, and in 2026 alone it updated ChatGPT's models and retired older ones more than once. An answer an older model wrote may not be the one its replacement writes.
What usually changes, and what holds?
The wording changes almost every time because it is written afresh. What deserves your attention is three other layers: the cited sources, the businesses the engine names, and whether there is an answer at all.
So read every answer in these four layers, and do not mistake movement in one for movement in another:
- Wording: sentence order, length and word choice. A change here says nothing about your site, so do not track it.
- Sources: the links the engine put in or under the answer. They tell you whether you are cited, so track them question by question.
- Names: the businesses the engine recommends in the text, with or without a link. They are what your customer reads first.
- Whether there is an answer: Google says some Gemini responses come without links, and that AI Mode may show links instead of an answer when it is not confident. An answer with no source is a result in its own right.
And how much do the sources change from one day to the next?
There is no single figure that holds for every question: how much they move can differ from one question to the next and from one engine to the next. So do not carry what you saw on one question over to another; find the amount on your own questions, one question and one engine at a time, with the three-day test below.
When is a change real rather than chance?
When it repeats. A change you see on different days, from two different accounts, on the same engine, is closer to the truth than one a single answer brought, however welcome or worrying it is.
Four signs tell you a change is worth acting on:
- It repeats on different days, not in two questions asked within the same hour.
- It repeats in two accounts, one of which does not know your business, so your account's memory is not the cause.
- It touches the sources or the names, not just the wording.
- It has a cause you know of: you edited your page, a competitor published a new page, or the company announced a model update.
And the lucky answer?
A citation that happened did happen: some customer read that answer. But it does not say it will happen again, and your absence from one answer does not say you have lost your place. Read a single answer as news about its day, not a verdict on your site.
How do you test this yourself in three days?
Pick one buying question, ask ChatGPT and Gemini for three days from two accounts, one that does not know your business, record each answer's links and names, then read what repeated.
Pick the question
A buying question with the service, the city and a condition, as your customer types it. In an example for a fictional Riyadh developer, mudun-homes.sa, the question is «وش أفضل مطوّر لشقق التمليك في شمال الرياض؟» ("what's the best developer for freehold flats in north Riyadh?"). Choosing the question is covered in how to choose the questions you measure on.
Set up two accounts
The first is your usual account; the second does not know your business. In ChatGPT, web search is available without signing in, or you can open a temporary chat and choose Unpersonalized before it starts: OpenAI says a temporary chat does not create or update memories, but it can use your existing ones unless you make that choice. In Gemini, use an account you have never asked about your business from.
Ask for three days
Ask the same question in the same words, in a new chat, on each of the three days and at about the same time. Do not edit your page during those days, or you will not know whether the answer changed because of you or on its own.
Record it in a table
One row per answer: date, engine, account, whether your site was cited, the cited links, and the businesses the engine named. Here are rows from the same illustrative example:
- Day 1, ChatGPT, usual account: mudun-homes.sa cited, alongside two other developers' sites.
- Day 1, ChatGPT, signed out: mudun-homes.sa absent; the two other sites cited.
- Day 2, ChatGPT, signed out: mudun-homes.sa absent; one of the same two sites cited.
- Day 3, ChatGPT, signed out: mudun-homes.sa absent; the same site cited for the third time.
Read what repeated
In this illustrative example the developer was cited only in its usual account, so what it saw first was most likely its own account's influence, not its place with its customers. One competitor repeated on all three days, so that is its competitor for this question.
If one competitor keeps repeating, start with their page, as in which competitor ChatGPT recommends instead of you. The full manual test is in how to test your site yourself today.
Which five questions should you ask anyone who measures for you?
Ask about the result you receive, not about the inner workings. Anyone measuring AI answers for you should answer five questions clearly before you pay them anything.
- Do you ask the question afresh at every measurement, or show me an old saved answer?
- Do you tell a change one answer brought from a change that held, and when do you alert me?
- Do I see the answer's excerpt and the links the engine gave as evidence under every result?
- Do you ask in my customer's language and dialect, and read Google's answers as they see them in their country?
- Do you show me the range when the result swings, or a single figure that hides the swing?
And why these five?
Because each one shows whether a provider reads a single answer as a trend, the mistake that costs you decisions built on movement that never held. Tasadr's answers follow in the next section, and the full selection criteria are in test before you pay.
How does Tasadr handle changing answers?
Tasadr (تصدَّر) asks each question afresh at every measurement, keeps the answer's excerpt and its links with their date, and does not call movement a change until it holds. We call that a "confirmed change".
- A new competitor: we do not call it new on the strength of one answer, only once it holds.
- A lost citation: we do not announce it lost on one answer without it, only once its absence holds.
- A new win: it counts at once, because a citation that happened did happen.
And what reaches you?
The rule is written in advance and is the same for every account. You get an alert only for what passes it, and what does not stays in the answer log for you to read whenever you like. The detail is in when do we say something has changed? and the term in what is a confirmed change?
The Tasadr score also shows its range when a measurement swings between two figures, so you see the swing yourself instead of having it hidden behind a single number.
What can nobody promise you?
Nobody can promise you a stable answer. The engines write every answer afresh and update their models and sources when they choose, and nobody outside them knows everything that shapes their answers.
- Nobody reads every customer's answer: location, account and memory can change what each one reads.
- Three days show you the swing; they do not show the trend over a month.
- We know of no fixed time after which an edit starts to show in answers; what sets the pace is in how long GEO takes to change results.
Sources
- Searching the web with ChatGPT · OpenAI Help Center
- Memory in ChatGPT · OpenAI Help Center
- Model Release Notes · OpenAI Help Center ·
- Learn about responses from Gemini Apps · Gemini Apps Help
- View related sources from Gemini Apps · Gemini Apps Help
- Get AI-powered responses with AI Mode in Google Search · Google Search Help
- AI features and your website · Google Search Central ·
Frequently asked questions
How do I track changes in AI answers about my site?
Why does ChatGPT's answer change every day?
Does my customer see the same answer I see?
Do ChatGPT's memory and my history affect the answer?
Does a changed answer mean my site has slipped?
Do AI Overviews answers change too?
Related guides
- How to know if your site shows up in ChatGPT and GeminiA check by hand that tells you today whether your site leads ChatGPT and Gemini answers, how to read it, when it is not enough on its own, and an action plan for what comes next.Read the guide
- How do you measure the success of optimising for AI answers?The indicators that measure GEO success, how long it takes for results to change and what decides it, how to work out the return, and what a report on your site's place in AI answers should hold.Read the guide
- What is the best tool to analyse your place in AI answers?Eight questions to ask any tool that measures AI answers before you pay, and when you need a tool rather than an agency.Read the guide
Try Tasadr free for 7 days
Add your site, write up to 10 questions your customers ask, and see who leads the answer today. No credit card needed.
Start your free trial