Why ChatGPT gives different people different answers about your brand
·7 min read
ChatGPT does not give the same answer to everyone who asks about your brand, and it rarely gives one person the same answer twice. OpenAI's developer docs say its chat models are non-deterministic by default, and each answer also depends on whether ChatGPT searched the web, what it remembers about the person asking, which model replied, where that person is and how they worded the question. A screenshot of one answer is a single draw from that range. You need several draws, taken under the same conditions, before you know what your customers are likely to read.
Does ChatGPT give the same answer to everyone?
No, and questions about a brand tend to vary more than questions about settled facts. Ask for the boiling point of water a hundred times and the wording shifts while the number holds. Ask "What is Acme Analytics and what does it cost?" and the facts themselves can change between runs, because a small brand has far fewer sources behind it.
| Cause | What changes | Can you control it? |
|---|---|---|
| Sampling | Wording, order, which details make the cut | No |
| Web search or training data | Your site today or an older training snapshot | Partly |
| Memory and custom instructions | What the model assumes about the asker | No |
| Model | Depth, length, emphasis | No |
| Location | Which local sources a search returns | No |
| Prompt wording | The question actually answered | No, but you can test buyers' wordings |
| Timing | Which version of the web or training data | Partly |
Why does ChatGPT give different answers?
Seven inputs change between answers, and OpenAI's own docs back each one.
It picks words by sampling
ChatGPT writes one token at a time and picks each from a set of likely candidates instead of always taking the top one. OpenAI's cookbook on reproducible outputs says its chat APIs are "non-deterministic by default," so outputs may differ from request to request. ChatGPT gives users no setting to change that. Small differences compound. If one run opens by calling you a project management tool and another calls you a team chat app, the rest of each answer follows its opening.
Some answers search the web and some come from training data
ChatGPT decides per question whether to search. OpenAI's article on searching the web with ChatGPT says it may search automatically when a question would benefit from current information, and users can also pick Search from the tools menu. Without a search, the answer comes from training data, which OpenAI's article on whether ChatGPT tells the truth says covers events only up to a cutoff.
For a brand, those are two different answers. The unsearched one may predate your rebrand. The searched one reflects your current pages, plus whatever stale directory listing came back. Searched answers show their sources, which is how you tell them apart. The retrieval side is covered in how ChatGPT search works.
Memory and custom instructions shape the answer
Logged-in users bring context. OpenAI's Memory article describes saved memories, a setting that lets ChatGPT draw on past chats, and custom instructions, which are standing guidance on how it should respond. Someone who told ChatGPT they run a five-person agency and prefer open-source software can get a different CRM shortlist than a procurement lead at a bank.
Paid and unpaid plans get different models
OpenAI's FAQ for the unpaid tier names a default model that differs from paid plans and says the default can change over time. The ChatGPT release notes show the defaults and the model picker changing several times in 2026, so an answer saved last quarter may have come from a model that is no longer the default.
Location changes what a search returns
The same search article says ChatGPT may estimate a user's general location from their IP address to improve local results, and that sharing precise device location is off by default. For a local business, a user in Leeds and a user in Denver can get different sources.
The exact wording is a different question
OpenAI's prompting advice for ChatGPT tells users to adjust wording and add context to get better results, which amounts to saying that wording changes the result. "Is Acme any good?" invites reviews. "Acme vs Basecamp" invites a comparison. "Best project tool for agencies" invites a shortlist you may not be on.
Timing decides which data the answer reflects
OpenAI's search article warns that search results and citations can be incomplete, outdated or incorrect. A review site that updates your pricing this week can change searched answers the same week. A model trained before your rebrand keeps the old name in unsearched answers until a newer model replaces it.
Why one screenshot of a ChatGPT answer proves little
A screenshot is one sample from one account, with that person's memory, model, location and wording baked in. None of that shows in the image.
When a customer sends one saying ChatGPT called your product discontinued, ask the boring questions. Did the answer cite sources, and which? Was the person logged in? What exactly did they type? If it cited a 2022 forum thread, you have a source problem you can fix. If it cited nothing, you may be looking at a stale training-data answer or a one-off sampling miss. The same applies to the glowing answer your CEO got in an account that has spent months asking about the company.
How to test what ChatGPT says about your brand
Treat it like a survey, with fixed questions, clean conditions and written records.
- Write down the questions buyers ask, such as "What is Acme?", "Is Acme worth it for a small agency?" and "Acme alternatives". Reuse the exact text every time.
- Start from a clean session. Log out, or open a Temporary Chat. OpenAI's temporary chat article says these are unpersonalized by default and then use neither memory nor custom instructions.
- Run each prompt five or more times, each in a fresh chat, so you can see which claims repeat.
- Run the set once as plain questions and once with Search chosen from the tools menu. Note whether each answer shows sources.
- Log the date, the model shown, your location and every cited URL. Without them, next month's comparison means nothing.
- If you sell in several languages or countries, repeat the set in each.
A claim that shows up in most runs is what ChatGPT reliably says about you. A claim that shows up once is noise, unless it is harmful, in which case go find its source. For shortlist questions like "best CRM for agencies", AI Answer Visibility asks ChatGPT, Perplexity and Gemini that question and shows whether each answer names or cites you, and who it names instead.
What can you change when ChatGPT answers vary?
You can't change the sampling, anyone's memory, their model or their location. You can change the two things searched answers are built from, which are your own pages and the third-party pages a search returns.
Start with your site, the one source about you that you fully control. Say what you sell, who it is for and what it costs in plain text on the page, not only in images or a pricing calculator. If you renamed the company, say so in a sentence, such as "Acme was called Widgetly until 2025." Brand Entity Checker checks that your name, logo, Organization markup and sameAs profiles consistently tie the brand to your site.
Then follow the citations you logged. If the same outdated review or directory listing keeps appearing across runs, that page is shaping your description, and fixing it there is usually worth more than another edit to your own site. Update your review-site profile, ask the publisher for a correction, or post current facts in the forum thread. Earning citations for your own pages is a bigger job, covered in how to get cited by ChatGPT.
Unsearched answers are out of reach. Nobody can edit a model that has already shipped, and what you publish now reaches those answers only if a later model trains on it. Be wary of anyone promising to change what ChatGPT says about you by next week.
Check what ChatGPT and other AI engines say about your brand
Brand Narrative Check runs a controlled version of the test above. It reads your home page, then asks ChatGPT, Perplexity and Gemini the same question with web search on. The question asks what your brand is, what it offers, who it is for, what it costs, which strengths and weaknesses reviewers mention, and its main alternatives.
The report shows whether each engine recognizes your brand, how each describes it and in what tone, the strengths, weaknesses and alternatives each names, the sources each cites, and potential discrepancies with your home page, quoted so you can review them. It compares answers with your home page only and does not track answers over time.
A run costs 150 credits and takes one answer from each engine, so treat it as three samples, not a verdict. A claim all three engines make is the one to act on first. A claim one engine makes once deserves a second run before you chase it.