2 food categories, each asked the questions real UK consumers ask, every month. Do they behave like one market or several? On the measures below, Food behaves like 2 different markets.
Each strip is one measure. The grey band is the programme's full range; the gold span is where Food's sub-categories sit inside it, one dot each. A ratio under 1 means they cluster tighter than the programme — a category with a shared logic — and over 1 means they are further apart than categories chosen at random.
| Measure | Sub-category | Value | Within-category spread | Programme spread | Ratio |
|---|---|---|---|---|---|
| Concentration | Food Delivery | 61% | 13.9 | 13.8 | 1.01 |
| Chocolate | 33% | ||||
| Leader visibility | Food Delivery | 78% | 22.5 | 22.4 | 1.00 |
| Chocolate | 33% | ||||
| Read → cited | Chocolate | 38% | 6.7 | 11.6 | 0.58 |
| Food Delivery | 24% | ||||
| Model agreement | Food Delivery | 100% | 50.0 | 35.7 | 1.40 |
| Chocolate | 0% |
The most-named brand in each sub-category's answers. No brand leads more than one sub-category here — each is its own contest. 1 of the 2 leaders are upsets: not the market leader the programme seeded.
The independent sources AI cites in the answers — counted only where they appear in the top sources of at least two of Food's sub-categories. Each sub-category has its own set of referees.
When Claude or ChatGPT opens a brand's own page while answering, how often does it make the cited sources? Food's sub-categories convert alike — a shared content problem, or a shared solution.
Method: every sub-category runs 20 unbranded consumer prompts monthly against ChatGPT, Claude and Gemini. Alikeness compares each measure's spread within Food to its spread across every category in the programme (population standard deviations); model agreement is the share of models naming the same leader; read→cited uses Claude and ChatGPT only. Correlational observations of model behaviour at the time of each run. Data updated Wed, 02 Sep 2026 04:25:01 GMT.