The mechanism has three parts, and understanding them is what makes the fixes make sense instead of feeling arbitrary.
1. Trusted sources, not a live crawl
When an AI assistant answers "best AC repair company near me," it isn't running a fresh web search the instant you ask (though some, like Perplexity and Google AI Overviews, do blend in live retrieval). It's drawing on a mix of training data and, increasingly, real-time lookups against sources it already trusts for local business facts — the same kind of sources a human researcher would check: business profiles, review aggregators, and directories.
2. Synthesis, not a ranked list
Google returns ten blue links and lets you decide. An AI assistant makes the decision for the reader and states it with authority — "call Acme HVAC" — which means it has to resolve conflicting or incomplete information about you into one answer. Thin, inconsistent, or contradictory source data makes that resolution harder, and an engine that can't resolve confidently either skips you or hedges with a shorter, less complete answer.
3. Per-engine differences
ChatGPT, Gemini, Claude, Perplexity, Copilot, and Grok are built by different companies with different training data, different retrieval partnerships, and different update cadences. That's exactly why a real audit checks every engine your customers actually use — a good result on one engine tells you nothing about the other five.