Where ChatGPT and Gemini actually find their local recommendations

When someone asks ChatGPT "who can deep-clean my apartment in San Francisco?", the engine doesn't consult a secret index of good businesses. It runs a web search, reads a handful of pages, and recommends whoever those pages support. Which pages get read decides who gets recommended.

To find out which pages those are, we asked ChatGPT and Gemini the same 217 unbranded local buyer questions — 10 categories, 22 businesses, 4 cities — and kept every answer that exposed its citation URLs (389 of 434 completed answers did).

The three source families that matter

• Business websites (62.7% of paired answers). The single most-cited source type is the business's own site. Engines fetch service pages, FAQs, and pricing pages directly and quote them. A business with no crawlable, locally specific pages is invisible at the exact moment it matters.

• Yelp (46.1% overall — 92.2% on ChatGPT). ChatGPT leans on Yelp so heavily for local service queries that an unclaimed or thin Yelp profile is close to disqualifying in that engine.

• Category directories (41.0%). Zocdoc for dentists, The Knot for wedding vendors, Houzz for contractors — every category has two or three directories the engines treat as authorities.

What this is not

This is a directional benchmark, not a universal ranking. The mix shifts by city and category — which is why every SmartCited scan recalculates it from live answers for your business, instead of assuming the average applies to you.

What to do with it

The practical reading: you control the #1 source family outright (your website), you can strengthen #2 within Yelp's rules (a claimed, complete profile and owner responses — never solicited reviews), and #3 is a finite claiming-and-outreach checklist. That's the plan a SmartCited scan writes for you, grounded in the citations your own category actually produces. Run a free scan or browse an example report to see it on a real business.