Logged-out ChatGPT is its own surface

A ChatGPT visibility number is a measurement of one ChatGPT surface: the API, the logged-out web app, or a signed-in account with its own memory, model and region. Those surfaces have always differed. This week OpenAI made the gap harder to ignore by testing a separate application for people who are not signed in.

#What changed

Tibor Blaho reported that OpenAI appears to be testing a separate lightweight ChatGPT web app for logged-out users, internally named "unauth-mweb". According to his post, it runs on its own experiment router, asset bundle, page structure and endpoint family instead of loading the full ChatGPT client, and responses stream as ready-to-display HTML. The "mweb" in the name suggests a mobile web design.

Search Engine Roundtable's daily recap for October 2, 2026 picked it up through a post from Jay at Radarkit, an AI visibility tracking vendor, which described ChatGPT as quietly moving logged-out users to a different app that hides most of what visibility tracking relies on. OpenAI has not announced the change, so treat the details as observed behavior in a test, not documented product.

The practical point does not depend on the details. Many trackers that avoid the API do so by opening a clean, logged-out browser session and typing the prompt. That session is now, at least for some users, a different codebase from the one a signed-in customer sees.

#Logged-out was never the default customer

Logged-out ChatGPT has existed since April 2024, when OpenAI started letting people use it without an account. Even then it came with limits: chats could not be saved or shared, and voice conversations and custom instructions required an account. OpenAI also said it added extra safeguards for accountless users, blocking prompts in more categories than for signed-in users. Web search followed later. ChatGPT search became available to everyone in supported regions with no sign-up required on February 5, 2025, after reaching all logged-in users in December 2024.

Signed-in sessions pick up a variable logged-out sessions do not have. In April 2025 OpenAI added Memory with Search, where ChatGPT may rewrite a prompt into a search query using details it remembers about the user. TechCrunch's example: a user known to be vegan and based in San Francisco asks for restaurants nearby, and the search becomes "good vegan restaurants, San Francisco." A different query retrieves different pages, which means different citations.

So a logged-out session was already a proxy. It approximates a new user with no history, on whatever model and feature set OpenAI gives anonymous traffic. A separate logged-out app widens the distance between that proxy and the signed-in buyer.

#The API is a third surface

Some tools skip the interface entirely and call OpenAI's API with web search enabled. That is a cleaner, repeatable setup, and it is also a different product.

Promptwatch, another tracking vendor, ran identical commercial prompts through the ChatGPT web interface and the API on the same day in the same locale. In its write-up the two methods overlapped on just 25.6% of their sources. The interface returned sources on 84% of prompts against 26% for the API, with 392 citations across 149 domains versus 170 citations across 67 domains. The post does not state the number of prompts, and a vendor that sells interface-based tracking has a stake in the result, so read the figures as one data point rather than a settled rate. The direction is still consistent with how the products are built: the interface decides when to search and how many sources to show, while an API call does what the developer configured.

#Why a separate app matters for measurement

A tracker that reads the logged-out page is parsing markup to find the answer text, the cited links and the source cards. If logged-out traffic moves to a different app with a different page structure and a different response format, three things can happen without any warning in your dashboard:

• The parser breaks and returns empty citations, which looks like a visibility drop.
• The parser still works but the new app shows fewer source elements, which looks like a citation drop that only exists on the logged-out surface.
• Sessions get split between the old and new app as the experiment runs, which looks like noise.

None of these reflects what signed-in buyers see. A chart that falls in the first week of October could be a real change in ChatGPT's sourcing, a scraper problem, or a sampling change on OpenAI's side. Without knowing which surface produced each data point, you cannot tell them apart.

#What to do with your ChatGPT numbers

Write the surface on every number. For each ChatGPT metric you report, record whether it came from the API, a logged-out session or a signed-in account, plus model, region and whether search ran. The Rankry team's guide to ChatGPT visibility tracking lists the same configuration variables (browsing state, memory, region, model version and the shopping surface). Ask your vendor for this in writing if the dashboard does not show it.

Annotate October 2, 2026. Put a marker on any ChatGPT trend line at the date the logged-out test was reported. If a series built on logged-out sessions moves after that date, check the raw answers before reading it as a change in sourcing.

Check for empty captures. A sudden rise in answers with zero citations, or answers with no text, is more likely a capture failure than a ranking event. Count them separately from answers that ran and cited other sites.

Do not merge surfaces into one score. An API series and an interface series measure different things. Report them side by side. If they diverge, that is information about the product, not an error to average away.

Keep a small signed-in panel. Logged-out sessions remain useful as a baseline for a new visitor. For high-value prompts, add a few runs from signed-in accounts with memory turned off and note the plan tier. That is closer to a returning customer, and it shows how far the anonymous number sits from it.

Repeat runs before reacting. Single answers vary run to run on every surface. Look at rates across repeated runs before treating a change as real.

#The takeaway

There is no single ChatGPT to rank in. There is an API, a logged-out app that OpenAI now appears to be rebuilding, and signed-in accounts shaped by memory and plan. Each can be measured, and each tells you something. The mistake is reporting one of them as "ChatGPT visibility" without saying which.