How we collect data
AI models
Which AI assistants we ask, how, and what data each one gives — sources, searches, location.
Survelium asks the same AI assistants for all brands. In the model filter on each page they appear with their full name (for example "ChatGPT Live Search"), so you know exactly which variant was measured.
All answers use web search
An assistant can answer from memory (what it learned in training) or by searching the web before it answers. Today people use the search version almost exclusively, so that's the one we measure: the answer reflects what the assistant finds on the web now, not what it knew a year ago. This is also where the cited sources come from.
Some assistants we read from the public interface, exactly what a person sees in the browser; the others we ask directly, with web search turned on.
The list
| Model | How we ask it | Cited sources | AI searches | Brand country |
|---|---|---|---|---|
| ChatGPT | public interface | yes | yes | yes |
| Copilot | public interface | usually | no | yes |
| Google AI Mode | public interface | yes | no | yes |
| Google AI Overview | Google result | yes, when it appears | no | yes |
| Gemini | directly, with search | yes | yes | no |
| Claude | directly, with search | yes | yes | no |
| Perplexity | directly, with search | yes | yes | no |
- Copilot sometimes answers in conversation mode, with no sources. The answer counts; it just doesn't cite anything.
- Google AI Overview is the AI box in Google results. For many searches Google doesn't show it at all; then the answer is empty and means "you don't appear" (see How we collect data).
- Google AI Mode always shows a fixed number of sources, so the number of sources there doesn't tell you much.
Old models in the filter
The list of models changes over time: we add new assistants and remove variants that no longer reflect what people use. The model filter also shows removed models, so their history stays visible.
All models together
"All models" puts all answers together. A model that answered more times weighs more; we don't average the models. See Metrics.