Skip to content
Public · Capture fidelity

How each engine is actually measured

"10 engines tracked" is only honest with this page attached: capture methods differ per engine, and the differences matter more than the count.

Why fidelity beats a platform count

A bare API model call, a search-grounded API call, and a real consumer session can return different answers to the same prompt. Tools that advertise one big platform number without saying which of these each “platform” really is are hiding the most important variable in the measurement. Below is our registry, generated from the same code that runs the captures.

API model answer (no web access)

A bare call to the provider’s model API without retrieval. It measures the model’s parametric answer — NOT what a consumer-app user sees, which may include search grounding and personalization.

Search-grounded API

The provider’s API with its search/grounding capability active, so answers draw on live retrieval. Closer to consumer behavior than a bare model call, but still an API capture, not the consumer UI.

Manually seeded / browser

Evidence entered from a real consumer surface (currently operator-seeded). Scheduled browser capture is policy- and spend-gated and is not running.

The matrix

10 primary engines + a separate grounded lane

EngineModelCapture modeDisclosure
ChatGPTgpt-4o-miniAPI model answer (no web access)
Claudeclaude-sonnet-4-6API model answer (no web access)
Geminigemini-3.5-flashSearch-grounded API
PerplexitysonarSearch-grounded APIThe sonar model runs a live web search on every call, like the Perplexity consumer product.
Grokgrok-4.3Search-grounded API
Copilotcopilot-bing-proxySearch-grounded APIConsumer Copilot has no public API, so we measure Bing Web Search results — the index Copilot is grounded in — as a disclosed proxy. An optional, off-by-default OpenAI fallback is labelled and excluded from engine counts.
DeepSeekdeepseek-v4-flashAPI model answer (no web access)
Metameta-ai-manualManually seeded / browserMeta AI has no public API — answers are copied from the real Meta AI app by an operator (manual seed), never fetched by an API or crawler.
You.comyou-smartSearch-grounded API
Phindphind-searchSearch-grounded API
ChatGPT (grounded)weekly grounded lanegpt-4o-miniSearch-grounded APIThe same base model as the ChatGPT lane, called through the OpenAI Responses API with the provider web-search tool enabled — the answer reflects live web retrieval. It is still an API capture, not the consumer ChatGPT app UI. Captured weekly on grounded-flagged prompts (budget-capped), while the bare-model lane runs on the normal cadence.
Not in this registry

Google AI Overviews and AI Mode are tracked on the search-results lane (SERP snapshots with inline-citation extraction), not through an answer-engine adapter — counting them as an “engine” would double-dip the search lane. Populations are never pooled across capture modes: the grounded ChatGPT lane is a separate series from the bare-model ChatGPT lane, so a methodology change can never silently rewrite a trend.