category-education · general · GEO Basics
How ChatGPT, Perplexity, and Google AI Overviews Actually Choose Sources
AI engines answer a query by first assembling a short list of candidate sources, then writing from the ones that pass a trust and extractability check. Ranking is not part of that process the way it is in classic search. Understanding the shortlist is the entire basis of GEO.
Type a question into ChatGPT, Perplexity, or a Google search that triggers an AI Overview, and the model does not return ten blue links for you to evaluate. It has already done the evaluating, in a way that has little to do with the ranking algorithm most marketers spent a decade optimizing for.
The shortlist happens before the answer is written
Every generative engine runs some version of the same two-step process: retrieve a set of candidate sources for the query, then generate an answer grounded in whichever of those sources clears a trust bar. Retrieval draws on a mix of the engine's own index, real-time search results, and, for some models, licensed data partnerships. None of that step resembles a ranked list. A source either makes the candidate set or it does not, and once it does, the model still has to decide whether to cite it.
Three signals that decide who gets cited
- Extractability. A claim stated in one clear, self-contained sentence is easy for a model to lift and attribute correctly. A claim buried in a long paragraph, or split across sentences that each depend on the last, gets paraphrased loosely or skipped.
- Corroboration. Models weigh how many independent sources say the same thing. A single confident claim from one site is worth less than the same claim appearing, worded differently, across several sources the model already treats as credible for that topic.
- Freshness and consistency. Dated content, regularly updated content, and content that does not contradict itself across pages all read as more trustworthy than a page that has gone stale or that disagrees with the company's own other pages.
What this means for your content strategy
None of this is about writing for the model at the expense of the reader. It is about writing plainly and specifically enough that a model extracting a claim cannot get it wrong, then making sure the same claim appears, in the company's own words, across the other places the brand already has a presence: press coverage, review platforms, partner sites, community threads. A brand that states something once, on one page, is easy for a model to skip past. A brand that repeats the same true thing consistently, in several credible places, becomes hard to leave out.