Grounding model

Every grounded AI engine has a grounding model: the architecture that decides what gets retrieved, ranked, and quoted. The differences across engines are substantial and matter for optimization.

Always-on retrieval. Perplexity, Bing Copilot. Every non-trivial query triggers retrieval. The citation surface is dense (5-10 sources per answer) and the answer style is highly extractive: the engine stitches quoted sentences from sources rather than freely generating.

Model-discretionary retrieval. ChatGPT Search, Claude, Gemini. The model decides whether each query benefits from retrieval. Recency-sensitive or factually-anchored queries trigger grounding; abstract or creative queries don't. The citation density is variable.

SERP-anchored grounding. Google AI Overviews. Retrieval is heavily biased toward pages already ranking in classical organic results. Citation share correlates strongly with top-10 organic ranking. The grounding model rewards what Google's classical ranker already rewards, with extractability as the additional dimension.

Hybrid corpora. Most engines pull from the open web plus vertical-specific sources. Perplexity has academic and finance modes; some vendors add proprietary partner content; Bing has a Microsoft Graph integration for enterprise. The corpus shape determines who gets cited.

For optimization, the practical implication: an engine that grounds based on classical organic ranking (AI Overviews) rewards classical SEO investment. An engine that grounds based on its own retrieval index (Perplexity) rewards content the retriever can match precisely. An engine that grounds discretionarily (ChatGPT, Claude) rewards content that survives both retrieval and the model's selection layer. Optimize for the grounding model of the engines you care about most.