On October 19, 2026 we asked ChatGPT for the best mattress for side sleepers three times. Casper came first once and Helix twice. Same question, same day. That's the thing we wanted to measure, so we ran two small tests on how a large language model handles brands.
Test one was tokens. A model reads chunks of text, not whole words. We gave 15 brand names to o200k_base, the open-source tokenizer GPT-4o uses. Every one split into two or three tokens: ClickUp into Click and Up, Lemonade into Lemon and ade, Mailchimp into Mail and chimp. Plumber and Kubernetes stayed whole.
Test two was repeat questions. Five buyer questions, like "best crm for small business", each asked three times. ChatGPT named 28 products in total, and 22 of them showed up every time. Where it slipped was the top pick, which changed on 2 of the 5 questions. Freelancer accounting software went Wave, QuickBooks, Wave. CRM, project management and email marketing never changed their first pick.
Mattresses were the shakiest list, 7 products named and just 3 in every run. Plenty of similar brands to choose from probably explains it. Across the whole test, 6 products came and went, 2 in CRM and 4 in mattresses. We call what stays the stable core and what moves the rotating tail.
We counted citations too. ChatGPT cited a median of nine sources per answer, and for 13 of the 15 answers it ran just one search.
One day, one run, so it's a snapshot. Try it yourself: ask ChatGPT one buying question from your category three times and note who comes first.
RankControl, an AI SEO and AEO platform, ran both tests. RankControl's AI visibility tracker tracks citations across 7 AI engines. The full write-up is on Medium.