Logged-In vs Logged-Out AI Testing: Why Your ChatGPT Brand Results Differ

OpenAI's help page on ChatGPT search says that if memory is enabled, ChatGPT may use relevant saved memories when it rewrites a search query. The same page says product results consider your query and context, such as Memory or Custom instructions.

That matters for brand checks. If you've spent months asking ChatGPT about your company, your account already holds your brand name, your category and your interest in it. A buyer asking the same question starts with none of that context.

So the answer you see may be shaped by you. It can name your brand more readily, or frame your category the way you usually describe it. We cover the mechanics in more detail in whether ChatGPT's memory changes what it says about your brand.

The next move: treat any answer from your own logged-in account as personalised until you've rerun it clean.

Memory

Saved memories can feed the query ChatGPT sends to search. Two people typing the same words can trigger different searches.

Custom instructions

OpenAI names custom instructions as context that shapes product results. If yours say you work in B2B software, answers lean that way.

Plan tier

Secondary reports from September 2026 put ChatGPT Free and Go on gpt-5.6-luna and Plus on gpt-5.6-sol, with higher tiers on a different model. OpenAI hasn't confirmed that mapping on a page we could read, and model lineups change often. Treat the specifics as unconfirmed. The point holds either way: your plan may not match your buyer's plan.

Model routing and retrieval

OpenAI's model docs list a February 16, 2026 knowledge cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path (chat-latest). An answer given without search reflects whatever that model learned before its cutoff. ChatGPT also decides per question whether to search at all, and OpenAI says it typically rewrites your question into one or more targeted queries. Each of these steps can differ between your session and a buyer's. For the wider picture, see how ChatGPT describes your brand, and why.

No column is the "real" answer. A logged-out session is the closest you can get to a stranger. An API test gives you the most control, but the API surface isn't the consumer app. Pick the column that matches the question you're asking.

Reproducibility means you can run the same check next month and compare like with like. As of October 2026, that takes a few disciplined steps.

1. Write a fixed list of questions in the words buyers use. Keep the wording identical between runs.
2. Run them logged out in a private window, or with memory and custom instructions turned off.
3. Record the date, the surface, and whether the answer showed sources. If no citations appear, you don't know whether it searched.
4. Run each question more than once. One answer is a sample, not a result.
5. Compare each answer against the facts stated on your own site, and log every wrong price, name or claim.
6. Repeat on a schedule, with the same list and setup.

Whether an answer searched the web matters more than most testers realise. Published estimates of how often ChatGPT searches disagree: Profound reports roughly 18% of conversations trigger a search, Semrush's US clickstream put it at 34.5% of queries in February 2026, and cloro measured 80.5% on its own 100-prompt set. They measure different things, so treat none of them as your number. Our free test of answers with search on vs off shows how far the two can drift.

Other engines need the same care. Claude.ai's web search is a user toggle with no published default. Perplexity's Agent API only searches if web search is switched on. Google says AI Overviews and AI Mode may use different models and techniques, so test them separately. The full method is in our reproducible protocol for testing AI prompts.

A quick check from your own account is still useful, if you read it for what it is.

It can tell you:

• That ChatGPT is capable of giving a particular answer about you.
• That a specific wrong fact exists in at least one answer, which is worth tracing.
• Which sources it cited, and whether they're your pages or other sites.

It cannot tell you:

• What any other user sees, on any plan.
• How often an answer appears across buyers.
• Your "rank" or position. No engine publishes position data for AI answers that such a number could be checked against.
• Whether a fix worked. One before and one after are two samples, and answers vary between runs anyway.

If the spot check turns up a wrong fact, the next step is verification, not panic. Our guide on how to tell whether ChatGPT is accurate about your company walks through it.

• Use your logged-in account for a quick look, or to see what a returning user with history might get. Never report it as what buyers see.
• Use a logged-out session with memory off when you want an answer closer to a first-time buyer's. This is the baseline for any check you plan to repeat.
• Use an API test when you need control: a pinned model, search forced on or off, and logged sources. Label it as a proxy for the app.
• Use a repeatable schedule when you need a trend. The same questions, setup and surfaces every run, compared over time. Our guide to monitoring what ChatGPT says about your brand covers cadence.

If you want a clean starting point without setting anything up, the free checker on this site runs up to three buyer questions and scores the answers. It's a snapshot, not continuous tracking.

Whatever you test, remember what's in your control. You can't control what a model says. You can control whether AI crawlers reach your pages and whether your key facts are stated plainly in the HTML.