There was no universal AI visibility trend
Grok recorded the largest fall in brand breadth, while the other three model surfaces changed little.
- Published
- 2026-09-02
- Revised
- 2026-09-02
- Author
- Cleotic Research
- Method
- comparison-v1.1
Grok brand-count change
Denominator: equal-weighted prompt and project comparison
Grok tracked-brand presence change
Denominator: paired July to August prompt comparison
Grok response-length change
Denominator: characters per successful response
The aggregate hid four different movements
Grok recorded the largest fall in tracked-brand breadth. Gemini was slightly positive, while Claude and Seed changed little.
United Kingdom / July to August 2026 / Denominator: 136 prompts paired across four continuously observed model surfaces and six projects / Method: comparison-v1.1
View chart data as a table
| Model surface | Brand breadth change | Tracked-brand presence change | Response length change |
|---|---|---|---|
| Claude Sonnet 5 | -0.015 | +0.20 pp | +4 chars |
| Seed 2.0 Lite | -0.023 | -0.68 pp | -11 chars |
| Gemini 3.5 Flash | +0.005 | +0.16 pp | +12 chars |
| Grok 4.3 | -0.064 | -0.09 pp | -97 chars |
One aggregate concealed four different movements
Grok lost 0.06 tracked brands per answer, while its tracked-brand presence rate was almost unchanged. Its answers were 97 characters shorter on average. Claude and Gemini moved slightly upward on response length, while Seed declined by 11 characters.
The pooled result is mathematically correct for the fixed panel, but it is not a diagnosis. A marketer who sees only the average could spend time changing a cross-channel programme when the observed shift belongs mainly to one model surface.
Report the market and the mechanism separately
A useful dashboard needs two levels. The fixed-model aggregate answers whether the measured market moved overall. The model, prompt and category cuts show where the movement occurred. Collection success belongs beside both because missing answers can otherwise masquerade as lost visibility.
This does not mean every model needs its own strategy. It means the evidence should identify whether a problem is broad enough to justify shared work or narrow enough to investigate on one surface first.
What marketers should do
- 01Keep a fixed-model aggregate for controlled trend reporting.
- 02Show model-level changes before assigning a cause or an owner.
- 03Drill from model to prompt cluster and then to the answers that changed.
- 04Keep a separate current-market view for new models, labelled as composition-sensitive.