How Does ChatGPT Decide Which Businesses to Recommend?

When someone asks ChatGPT "who should replace my water heater?", the answer comes from one of two places. Sometimes the model answers from its trained memory: a compressed impression of the web as it existed during training. Sometimes it decides the question needs current information, searches the live web, reads a handful of pages, and composes an answer with citations (OpenAI, "Introducing ChatGPT search").

You can usually tell the modes apart: live-search answers show source links, memory answers don't. The distinction matters practically. Live-search answers respond to changes you make within days or weeks, because they re-read the web every time. Memory answers only change when the model itself is updated, on a schedule you don't control. This is why the realistic strategy targets the live-search layer first: it's the one that listens.

For local business questions, live search is increasingly the default, because the model knows its memory of any given plumber in any given suburb is unreliable. That's good for you: it means the contest is decided by what's on the web this week, not by what a model memorized last year.

Watch the citations on local recommendation answers and the reading list is consistent: the business's own website, Google Business Profile data and reviews, platforms like Yelp, local directories and chamber listings, and editorial roundups from local publications (Local Falcon). In other words, the same places a diligent human researcher would check, read at machine speed with machine impatience.

Impatience is the operative word. The system reads a handful of pages for a few seconds each. It doesn't hunt through your PDF brochure, interpret your image-only price list, or execute the JavaScript your site needs to display its content. Whatever isn't sitting in plain, structured, immediately readable text on those few pages effectively doesn't exist for that answer.

This is also why your own website still matters enormously even in a reviews-and-directories world: it's the one source where you control every word, and it's almost always in the reading list.

Across our weekly scans and the public research, the businesses that get named share observable traits. Their pages answer questions directly, with concrete services, areas, and prices a model can quote. Their facts agree everywhere: website, Google profile, Yelp, directories. Their review evidence is visible and current: engines regularly justify picks by citing ratings and review counts. Independent pages, roundups, directories, news, confirm they exist and matter. And their pages carry structured data that removes the guesswork about what kind of business they are.

Notice what's not on the list: domain authority scores, keyword density, backlink counts, the traditional SEO scoreboard. Those still matter for ranking in search results, and search results feed AI reading lists, so the worlds connect. But the AI's final selection leans on extractable, verifiable, corroborated facts more than on any ranking metric. The original research on generative engine optimization measured exactly this: adding concrete statistics, quotes, and citations to pages raised how often AI systems used them, by up to about 40 percent in some tests (arXiv 2311.09735).

Seer Interactive's follow-up work adds the volume dimension: brands mentioned consistently alongside their service across many pages got named more, a pattern they summarized as repetition beating thought leadership (Seer Interactive). The machine believes what it sees confirmed everywhere.

Run the same recommendation question ten times and you will not get the same list ten times. SparkToro and Gumshoe.ai tested this at scale in 2026, thousands of prompts across ChatGPT, Claude, and Google's AI, and found the lists vary so much that identical asks almost never produce identical results (SparkToro).

The variance comes from everywhere: which pages the live search happened to fetch, how the question was phrased, the model's own sampling randomness, and ongoing updates to models and sources. None of it is personal.

The practical consequence: stop thinking in terms of "ranking #1 in ChatGPT" and start thinking in terms of share. Your business is either in the pool of candidates the engine draws from often, occasionally, or never. Everything in this article moves you up that ladder, and measuring it honestly means asking repeatedly over time. A single spot check, good or bad, tells you almost nothing. That's why we run the same panel of questions every Monday rather than declaring victory or defeat off one answer.

If the mechanics above are right, and everything we can measure says they are, then improving your odds is ordinary checklist work. Make sure AI crawlers can fetch your site at all. Write pages that answer real questions in liftable sentences. Make every listing agree. Keep reviews flowing. Earn a few independent mentions. Add structured data so nothing about your category is a guess.

If ChatGPT currently doesn't name you and you want to know which of those is the bottleneck, our diagnostic guide, Why doesn't ChatGPT recommend my business?, walks the six causes in order with a check for each. And the free checker reads your site the way the machines do and tells you what they can't find, which is the fastest way to stop guessing.