The seven answer engines
There is no official ranking or citation API for AI answers, and every engine grounds its response on a different index with a different retrieval strategy. Lokrix models seven distinct answer engines — ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Gemini, Grok and Claude — because being surfaced in one tells you little about the others.
Why they are separate engines
Each engine draws citations from a different source of truth. ChatGPT retrieves through the Bing index — roughly 87% of its citations match a Bing top-10 result. Copilot also grounds on Bing. Gemini grounds on Google Search. Perplexity runs its own live search across Google and Bing and heavily over-weights Reddit (about 47% of its top citations) and freshness. Claude and Grok both use live web search.
Google is the clearest case for keeping engines apart: AI Overviews and AI Mode both synthesise from Google's own index, yet they cite the same URL only about 14% of the time. Optimising for one does not carry over to the other, so Lokrix scores them independently.
How each engine grounds
- ChatGPT — retrieves via the Bing index; Bing rank and freshness dominate.
- Google AI Overviews — synthesised from Google's own index, shown above the classic SERP.
- Google AI Mode — Google's conversational surface; cites different URLs than AI Overviews.
- Perplexity — live Google+Bing search, over-weighting Reddit and freshness.
- Gemini — grounds on Google Search.
- Grok (xAI) — live search, with a real-time signal from X.
- Claude — answers with web search when the question needs current information.
What runs live by default
Five engines run live out of the box: ChatGPT, Claude, Perplexity, Grok and Gemini. Google AI Overviews and AI Mode have no direct API — they are observed through a SERP provider (Serper or SerpAPI), so they only run once a search-provider key is configured for the workspace.
Reading per-engine results
Because answers are non-deterministic, every engine result is a Monte-Carlo estimate: each prompt is run K times across temperature, persona and locale to produce a presence probability with a 95% confidence interval — never a single-run yes/no. Per engine you get a component score, presence, share-of-voice and citations, which roll up into the overall Lokrix AI Score. See sampling & confidence intervals for the method.