How often do AI answers change which brands they name?

How different are two answers to the same prompt?

Different enough that a matching list is the exception. SparkToro and Gumshoe found the same brand list twice less than 1 percent of the time, and the same order less than 0.1 percent.

The test, reported by Search Engine Journal in January 2026, ran 2,961 prompts through ChatGPT, Claude, and Google's AI Overviews, 60 to 100 times per prompt on each platform. The prompts asked for brand recommendations in categories from chef's knives and science fiction novels to cancer care hospitals and digital marketing consultants. Gumshoe.ai sells AI tracking tools, and Rand Fishkin disclosed that.

Headphones show what that looks like. Bose, Sony, Sennheiser, and Apple appeared in 55 to 77 percent of 994 responses to varied headphone prompts.

SEJ also reports that a brand's visibility percentage across many queries held steadier than its ranking position. Our guide to AI search covers how brands get named at all.

How much do cited sources change week to week?

Most of them turn over. Sistrix found Google's AI Mode replaced 56 percent of its cited sources each week, and ChatGPT Search about 74 percent.

Sistrix tracked 82,619 prompts across six countries from 17 December 2025 to 8 April 2026, 1,548,213 snapshots in all. Google's AI Overviews behaved differently in the same data, staying completely stable for just over half of queries.

Which sites each tool cites in the first place also differs, and our comparison of ChatGPT and AI Overviews goes through where they overlap.

Why does this happen?

Part of it is the models themselves. The same model asked the same prompt minutes apart cites different sites, before the web has had time to change.

cloro, a company that sells AI visibility tracking, ran the same prompt twice a few minutes apart on the same model and found about 40 percent of cited domains repeated. Per prompt, the overlap ranged from 0.06 to 1.00.

The odd part is the gap between Google's two surfaces, one mostly still and one churning weekly. Sistrix calls them different citation architectures and treats the drift as a structural feature of the platforms, which names the gap without explaining it.

How many runs make a number trustworthy?

Around 300 answers per period gets you to roughly plus or minus 5 points, by one vendor's maths. Twenty prompts leave a margin of about 17.5 points, which is too wide to show a trend.

Both vendors sell tracking tools, so we read these as arithmetic, not independent research. cloro puts about 72 answers at roughly plus or minus 10 points, and says spotting a move from 20 to 30 percent takes about 294 answers in each period.

The Hoth's guide to score accuracy, by Rachel Hernandez, shows what a small test hides. At 20 prompts, a score of 20 percent has a 95 percent interval of roughly 2.5 to 37.5 percent. At 100 prompts the margin narrows to about 7.8 points, and at 400 to about 3.9.

The Hoth also recommends running each prompt at least three times per period. A tool that runs your prompt list once a month and prints a position is reporting one draw.

What should you report instead of rank?

A mention rate, meaning the share of many answers that name you, collected on a schedule and compared quarter over quarter. Put your competitors in the same report as a control.

We build it from a fixed prompt set, with each prompt run several times a period and every answer checked for your name and each competitor's. The Hoth suggests collecting monthly and comparing quarters, since a monthly swing of a few points can sit inside the margin of error. Work on your own pages takes one to three months to show, so the quarter suits that too.

In our AI search work we capture a baseline across AI tools on day one and report your share of AI answers against it every month.

The competitor rows do the most work. If every brand in your category drops by a similar amount in the same month, the tools changed and your pages probably did not; if you drop alone, look at your pages.

A screenshot of ChatGPT naming you first proves very little. In SparkToro's test, the same order came back less than once in a thousand runs.