By Carson Cobb. Data collected November 2026.
“ChatGPT recommended us yesterday and not today.” I hear some version of this all the time. AI answers are written fresh every time, so some change is guaranteed. The useful question is how much, and where in the list.
So I asked ChatGPT the same 12 buyer questions 4 to 5 times each, in fresh temporary chats, across two days. That’s 55 answers, and I tracked every brand named in each one.
The short version
- The first pick almost never moves. The brand named first came back in 98% of the other runs of the same question.
- Position decides stability. Second place came back 93% of the time, third 74%, fourth 65%, and fifth or lower 56%.
- The same top pick in every run for 10 of 12 questions. In the other 2, two brands traded first place.
- Only 40% of brands were named in every run. 19% appeared only once.
- Wording matters more than randomness. Rewording the question changed the list far more than asking it again did. See how question wording changes AI recommendations.
Stability by position
For every brand in every answer, I checked whether it was named again in each other run of the same question.
| Position in the answer | Named again in another run |
|---|---|
| 1st | 98% |
| 2nd | 93% |
| 3rd | 74% |
| 4th | 65% |
| 5th or lower | 56% |
| Any position | 77% |
The drop after second place is the important part. Positions 1 and 2 are close to locked. From position 3 down, a brand’s presence is closer to a coin flip than a guarantee.
How the lists changed
Across the 12 questions, ChatGPT named 83 different brands at least once.
| Brands | Share | |
|---|---|---|
| Named in every run | 33 | 40% |
| Named in some runs | 34 | 41% |
| Named in only one run | 16 | 19% |
Two runs of the same question shared, on average, 66% of the brands either of them named. That sounds low, but it’s much higher than the overlap between different AI tools for the same question, which was 39% to 45%. See ChatGPT vs Gemini vs Google AI Mode.
List length moved too. One question about protein powder got 3 brands in one run and 9 in another. A longer list doesn’t push the top brands out. It adds more brands at the bottom, and those are the ones that come and go.
Top picks
| Questions | Same top pick in every run | Top pick changed |
|---|---|---|
| 12 | 10 | 2 |
The brand named first most often held first place in 52 of 55 answers. In both questions where first place changed hands, the two brands trading places were each named in at least 4 of 5 runs. The top of the list rotates between a small set of strong brands. It doesn’t jump to brands that were barely there.
Growing brands held their spots about as well
24 of the 83 brands have their own Shopify store and appear in my dataset of growing US brands. 11 of them (46%) were named in every run they could have been, against 22 of 59 (37%) for every other brand. Being a smaller brand did not make a spot less stable. Position did.
What this means for checking your own visibility
- One answer is not a result. Ask the same question at least three times, in fresh temporary chats, before you decide whether you’re recommended.
- Record the position, not just yes or no. Position 1 or 2 is a strong signal. Position 4 means you’re in the part of the list that changes between conversations.
- Count how often, not whether. “Named in 3 of 5 runs” is a far more honest number than a screenshot.
- Moving up beats getting in. Going from fifth to second matters more than going from absent to fifth, because fifth comes and goes.
- Compare like with like over time. Same questions, same number of runs, fresh sessions each time. That’s the only way a month to month change means anything.
This is how I track products for clients: the same buyer questions, repeated runs, positions recorded, on ChatGPT, Gemini and Google AI Mode. See the AI Search and Conversion Audit, or start with the free AI Visibility Check.
How this was done
- Questions: 12 of the 48 buyer questions from Who ChatGPT recommends, across apparel, beauty and skincare, food and beverage, home, pet and supplements.
- Runs: 4 or 5 runs per question, 55 in total. The first run of each question and a repeat of some questions on November 4, 2026, and three more runs on November 5, 2026.
- How they were asked: ChatGPT on a paid plan, default model and default thinking setting, each run in a new temporary chat.
- Brands: every brand named in an answer as an option to buy, in order of first mention, matched against a list of known brands and aliases for each question and checked by hand.
- Named again: for each brand in each run, the share of the other runs of the same question that also named it.
- Limits: two days of data on one AI. Stability over months will be lower, and Gemini and AI Mode may behave differently.
No store or brand is named in this report.