How AI models get their information about your business
Every AI answer about your business is fed by two pipelines. The first is training: a huge snapshot of text the model learned from, frozen months or years ago. The second is retrieval: a live lookup the assistant runs while the customer waits, reading search results, your site, and listings about you. Knowing which pipeline is in play tells you what you can change and how fast.
The two pipelines: what it learned and what it looks up
Think of it as what the AI learned versus what it looks up. What it learned about local businesses is thin and stale, because models train on a snapshot that ended months ago. What it looks up is fresh, and it comes from sources you can actually improve: your reviews, your listings, your own pages.
What it learned (and when learning stopped)
Models are trained on a large collection of text with an end date. Anything that happened after that date isn’t in the model’s memory. Your new phone number, your rebrand, the 40 reviews you earned this spring: none of it exists in the trained model until a future version is trained.
That’s why a raw model with no web access can sound confidently out of date. It isn’t lying; it’s reciting a snapshot. And for most local businesses that snapshot barely mentions them at all, which sounds like bad news and mostly isn’t.
What it looks up in real time
This is the pipeline that matters for you, because every major assistant now reads the live web before answering questions like “who should I call.”
What each platform says about its own lookup behavior, checked against their published pages on July 17, 2026:
| Platform | What its own documentation says |
|---|---|
| ChatGPT | Searches the web for current information and gives answers “with links to relevant web sources” (OpenAI announcement, Oct 2024)1 |
| Google AI Overviews and AI Mode | Runs searches over Google’s index, including a “query fan-out” that issues multiple related searches, and links out to the web (Google, Mar 2025)2 |
| Perplexity | ”Searches the internet in real-time” and attaches numbered citations to sources (Perplexity Help Center)3 |
| Claude | Can search the web and “provides direct citations” for what it pulls in (Anthropic, Mar 2025)4 |
Each platform decides for itself when to search. Simple factual questions sometimes get answered from memory alone. Local recommendation questions usually trigger a lookup, because the model knows its memory is weak there. We say “usually” because this is platform behavior we observe and the vendors don’t fully document; treat it as a strong pattern, not a rule. There are also readable signals for telling a live lookup from a memory answer, starting with whether the answer carries citations.
The sources that feed both
The same short list of sources keeps showing up on both sides of the pipeline, which is convenient: work on them once and you feed everything.
The list: your reviews and the actual words in them, the fields AI reads off your Google Business Profile, the short set of directories AI systems demonstrably cite, your own website if AI crawlers can actually reach and read it, and press, forums, and other third party mentions. How an assistant weighs all of that before naming anyone is its own subject, but the inputs are that short.
One short note on why the list is short: AI systems lean on sources that are big, structured, and hard to fake. That’s the filter. A random widget site never feeds an answer; a review platform with millions of entries does.
Why the same question gets different answers on different platforms
Because each assistant has its own training snapshot, its own search index, and its own reading list, the same question can get you named on one platform and skipped on another. That isn’t a bug you can file; it’s four separate systems disagreeing about the web, and the disagreements follow patterns you can actually work with.
Our take: chasing one platform is the wrong lesson to draw from this. The overlap in what the four read is much bigger than the differences, and the businesses that show up broadly are the ones that fixed the shared sources rather than tuning for a single assistant. Fix the inputs they all read, then watch where you appear.