Why ChatGPT and Google AI name different brands every time you ask.

What the study found

In November and December 2025, about 600 volunteers put the same 12 requests for recommendations to ChatGPT, Claude and Google's AI. They asked about things like cancer care hospitals on the West Coast, cloud providers for software startups and new science fiction novels. Together they collected 2,961 answers. Rand Fishkin of SparkToro ran the study with Patrick O'Donnell of Gumshoe, a company that sells AI tracking (SparkToro).

The lists rarely matched. ChatGPT or Google's AI gave the same list of brands less than once in 100 tries. The same brands in the same order came back about once in 1,000. Some answers named 2 or 3 brands. Others named 10 or more.

Some brands still came up again and again. City of Hope, a hospital in Los Angeles, was in 69 of 71 ChatGPT answers about cancer care. It was named first in only 25 of them. In fields with few choices, such as cloud providers, the leading names turned up in most answers. In fields with many, such as novels, the names were scattered.

Why the answers change

The model picks each word with some chance. An AI model writes one word at a time, choosing from a list of likely next words. A setting called temperature decides how often it picks a less likely word. Above zero, two runs can part ways early and end up naming different brands. Watch it happen in our temperature explainer.

Even at zero, answers drift. In September 2025, researchers at Thinking Machines asked one open model to tell them about Richard Feynman 1,000 times, with temperature at zero. They got 80 different answers. All 1,000 began with the same words. Then 992 said Feynman was born in Queens, New York, and 8 said New York City. The researchers traced it to the servers: the result can shift with how many other requests a server is handling at that moment (Thinking Machines).

The search behind the answer changes. For many questions, an assistant searches the web first. Google says its AI Mode and AI Overviews may run several related searches for one question, across subtopics and data sources (Google). Different searches find different pages, and different pages name different brands.

The person asking differs. The volunteers used their own accounts and settings. The study did not control their location, device or chat history, and its authors list that as a limit.

What it means for your brand

The study's authors came round to one measure. Over dozens to hundreds of runs of the same questions, the share of answers that name a brand is, in their words, a reasonable metric. A brand's place in one list is not: they say any tool that gives a ranking position in AI is full of baloney.

• Treat one answer as one draw. A colleague who saw you named, or left out, saw one result out of many.
• Count how often you are named, across many tries of the questions your customers ask.
• Watch the trend over weeks, not one day. A drop on Tuesday may be noise.
• Read what the answers say about you, as well as whether you are in them.

Visibian asks your questions on a schedule, daily or weekly, and shows a range around each figure, so you can tell a real change from noise. How to measure AI visibility covers the rest.

What the study cannot tell you

Its authors say it was not peer reviewed and that they are not professional researchers. One of them works for a company that sells AI tracking, which they disclose. It covered three assistants and 12 questions over two months, and newer models may behave differently. It did not look at how the answers described each brand.

The figures above are the study's own. We read the published report on 30 October 2026.