A changing model mix can move the score without moving the brand
Two valid August views pointed in opposite directions because new model surfaces entered after 8 August.
- Published
- 2026-09-02
- Revised
- 2026-09-02
- Author
- Cleotic Research
- Method
- comparison-v1.1
brand breadth in the fixed panel
Denominator: four model surfaces observed across both months
brand breadth in the changing-roster view
Denominator: all other eligible observed model surfaces
August roster change
Denominator: GPT-5.5 ended after 8 August; Luna and Muse entered
Changing the model mix can reverse the headline
Response length and tracked-brand breadth fell in the fixed four-model panel but rose in the changing-roster view. Both views apply the same exclusions. Only the fixed panel supports a like-for-like month comparison.
United Kingdom / July to August 2026 / Denominator: Fixed panel: four continuous model surfaces; changing roster: all other eligible observed surfaces / Method: comparison-v1.1
View chart data as a table
| Measure | Fixed four-model change | Changing-roster view |
|---|---|---|
| Response length (characters) | -1.3% | 5.7% |
| Tracked-brand breadth (brands per response) | -0.8% | 4.4% |
| Responses mentioning a tracked brand (share) | -0.1% | 0.3% |
| Latency (ms) | -4.3% | -8.3% |
| Output tokens | 1.3% | 6.7% |
The denominator changed the direction of the story
In the fixed four-model panel, tracked-brand breadth fell 0.8% and response length fell 1.3%. In the changing-roster view, brand breadth rose 4.4% and response length rose 5.7%. Both views apply the same exclusions, so the reversal comes from the model mix.
GPT-5.5 appeared for the first eight days of August. GPT-5.6 Luna and Muse Spark appeared for the final 23. Their profiles can describe those windows, but they cannot turn the roster transition into a controlled month comparison.
One score cannot serve trend reporting and market coverage equally well
A fixed panel is the more defensible view of change because it holds the requested model surfaces constant. A changing-roster view is still useful for describing the current set of available systems. The mistake is to label the second one as pure brand movement.
Every executive chart should name its panel, show additions and removals, and preserve a bridge back to the last compatible period. Without that information, a portfolio or model update can be mistaken for the result of marketing work.
What marketers should do
- 01Maintain a fixed-model trend series and a separate current-market series.
- 02Annotate every model addition, removal or version transition.
- 03Do not splice incompatible panels into one unlabelled line.
- 04Recalculate historical comparisons when a known data-quality exclusion changes.