3 · Documentation
Channels and systems
Model knowledge, live search and answer box: why we keep three channel types apart and what the contrast between them reveals.
An AI answer can come from two places: from what the model learned in training, or from what it finds on the web at the moment of the question. For a brand these are two entirely different situations. augenmerk measures both separately and calls the difference channel contrast.
Three channel types
| Channel type | What it answers | Sources |
|---|---|---|
| Model knowledge | Does the model know the brand at all? Answer without web search, from training alone. | none |
| Live search | Does the AI find the brand when it searches? Answer with a current web search. | yes, with source list |
| Answer box | Is the brand in Google's AI answer above the search results? | yes, from Google's search |
Model knowledge names no sources, because it does not search. Every citation rate in augenmerk therefore refers only to the channels that search; otherwise a zero that means nothing would look like a finding.
The ten systems
| Provider | Model knowledge | Live search | Answer box |
|---|---|---|---|
| OpenAI / ChatGPT | OpenAI · model knowledge | ChatGPT · live search | |
| Gemini · model knowledge | Gemini · live search | Google · answer box (AI Overviews) | |
| Anthropic | Claude · model knowledge | Claude · live search | |
| xAI | Grok · model knowledge | Grok · live search | |
| Perplexity | Perplexity · live search |
Gemini takes part twice with the same model, once without and once with Google search. Only here does channel contrast arise within one provider: whoever compares OpenAI with Perplexity compares two different models; with Gemini the only difference is access to the web.
The contrast is the diagnosis
A single value across all channels would mix the two situations. "The AI does not know you" and "the AI knows you but does not find you in its sources" are two different findings with two different measures:
- Model knowledge high, live search low: the brand is known but missing from the pages the search channels currently read. A source problem. The measure lies with test and comparison sites, not with awareness.
- Live search high, model knowledge low: the sources know the brand, the models do not yet. Typical for young brands; model knowledge catches up with new model versions, in months, not runs.
- Answer box deviates: Google draws its answer from its own search, the chat systems from their sources. Movement shows where your own pages rise or fall on Google.
The findings catalogue lists these situations as entries of their own, see Channel findings.
What the search channels reveal
Some systems disclose the search queries they ran before answering. "Which frozen pizza is the best?" becomes, say, the search "frozen pizza test 2026". augenmerk stores these queries and shows them in the answers view. They show most precisely what an AI searches for when a person asks something: whoever ranks up front on Google for that search gets read as a source.
For other systems we know the source list but not the search. We therefore never claim that a search led to a particular source.
Retrieved and cited
The search channels present the model with more pages than it uses. Perplexity, for instance, retrieves around twenty addresses per answer and cites seven on average. augenmerk keeps the two apart: retrieved means the search considered the page relevant; cited means the model used it in the text. Not every system discloses the distinction; where the number of citations always equals the number of retrievals, the sources view marks that with an asterisk. More under Sources and website.
Updated: 2026-09-16