What Is an AI Answer, and Where Does It Come From?

The short answer

An AI answer is text the system writes at the moment you ask — it is not a page copied from somewhere, and it is not a record pulled from a database. It is assembled on the spot from two ingredients: what the model learned during its training, and, for many questions, sources it retrieves from the web while answering.

That is why two people can ask the same question and get different answers, and why you can ask the same question twice and watch the wording — and sometimes the facts — change.

What actually happens when you ask

When you type a question into ChatGPT, Google’s AI features, Perplexity, Copilot, or a similar system, here is roughly what happens behind the screen:

First, the model reads your question. It works with patterns of language, not with a filing cabinet. Nothing is being “looked up” in the traditional sense yet.

Sometimes, it goes and reads. For questions about current events, specific businesses, prices, hours, or anything it might not know reliably, many systems search the web at that moment, pull a set of pages, and use what those pages say as raw material. This is called retrieval. Not every question triggers it, and each platform decides differently — and none of them publish exactly when or why.

Then it writes. The model generates an answer word by word, drawing on its training, the retrieved pages if there are any, or both. The result is new text. Even when it closely follows a source, it is a paraphrase composed in that moment, not a quotation lifted from a file.

Citations, when they appear, are pointers — not a full accounting. A citation shows a source the system consulted or wants to credit. It does not prove that everything in the answer came from that page, and the absence of a citation does not mean no source was involved. Think of citations as a partial reading list, not a receipt.

Three things an AI answer is not

It is not a search result. A search engine returns a list of pages that exist. An AI answer is a new piece of writing that may draw on several pages and its own training at once.

It is not a database record. There is no stored entry somewhere that says what your business is, which an editor could open and fix. There are only sources — your website, your listings, reviews, articles — that answers are assembled from, again and again, each time someone asks.

It is not guaranteed to be the same twice. Generation involves a degree of variation by design. Ask the same question next week and the system may retrieve different sources, weigh them differently, or phrase the answer in a way that changes what a reader takes from it.

A small-business example

Say you run a bakery, and a customer asks an AI assistant: “Where can I get a gluten-free birthday cake in town?”

If your website has a page that says, in plain text, that you bake gluten-free cakes to order, and your listings and reviews say the same thing, the system has material to work with. It may name your bakery, describe the cakes, and cite your site.

If your gluten-free cakes are only mentioned in a photo of a handwritten sign, or nowhere at all, the system has nothing to retrieve. It will answer using the bakeries it can document — your competitors — or give a generic answer. Nothing went wrong mechanically. The answer was assembled from the record that exists, and your part of that record was missing.

What this means for your business

The practical consequence is the most important thing on this page: you cannot edit an AI answer, because there is no answer stored anywhere to edit. Answers are regenerated each time. What you can change is the record the answers are drawn from — the pages, listings, and sources about your business that a system can find and read.

That is also why one strange answer is not, by itself, proof of a problem, and one good answer is not proof you are safe. What matters is the pattern across many questions and many days — which is why Reputation.blog’s work starts with documenting what AI systems actually say, on the record, before changing anything.