How AI search works: how ChatGPT, Perplexity and Google choose their sources

When someone asks an assistant a question, the reply feels like it comes from one clever mind. It does not. Behind that single paragraph there are usually several separate steps, and each one is a place where a business can be found or missed.

None of the companies publish every detail, and their systems change often, so treat this as a working model rather than a specification. It is accurate enough to make good decisions.

Two sources of knowledge

A large language model (LLM) is a program trained on a huge amount of text so that it can predict and write language well. From that training it holds a broad, blurry memory of the world, up to a cut-off date. It does not store web pages like a library. It stores patterns, which is why it can explain an idea fluently and still get a small business’s address wrong.

The second source is live retrieval. When a question needs current or specific information, modern assistants search for it and read what they find. This is often called retrieval-augmented generation, or RAG, which simply means “look things up, then write the answer using what you found.”

What happens when you ask a question

Take a question such as “Which company can automate appointment reminders for my clinic in Gujarat?” A typical assistant that can browse does roughly this.

• It decides whether it needs fresh information. Local, recent and specific questions almost always do.
• It rewrites your question into one or several search queries. This is sometimes called query fan-out: one question becomes many small searches, such as “appointment reminder automation clinic India”.
• It sends those queries to a search index and gets back a list of candidate pages. Depending on the product, that index may be its own, Google’s, Bing’s or a mix.
• It fetches and reads the most promising pages, then picks the passages that answer the question.
• It writes a response in its own words, often with citations pointing back to the pages it used.

The important consequence is that your page has to survive every step. If it is not in the index, it is never a candidate. If it is slow or unreadable, it is skipped. If it is readable but vague, another page’s clearer passage is used instead.

Why some businesses get named and others do not

Assistants are not choosing the “best” business in any absolute sense. They are choosing the passages that best answer the question and that look reliable enough to repeat. In practice that favours a few qualities.

• Directness. A page that states the answer in its first sentences is easier to use than one that circles for four paragraphs.
• Specificity. “Website projects usually take four to six weeks” is quotable. “We deliver on time” is not.
• Freshness. Dates and recent updates matter, especially for topics that change, such as prices, rules and tools.
• Consistency. The same facts appearing on your site and on independent profiles make a claim safer to repeat.
• Authority signals. A named person, a real location, reviews and mentions elsewhere all help the assistant treat you as a real business.

Training crawlers and search crawlers are different

This trips up a lot of website owners. Companies typically run more than one crawler, and each has a separate job. One collects material that may be used to train future models. Another fetches pages so the assistant can answer a live question. A third acts when a user explicitly asks the assistant to open a link.

Because they are separate, you can choose what to allow. Blocking a training crawler does not remove you from live answers, and blocking the search crawler does. If your goal is to be found, the search crawlers are the ones you must not block. You can check by opening yourwebsite.com/robots.txt and reading the rules.

Where llms.txt fits

Some sites publish a plain-text file called llms.txt that lists their key pages and summarises the business for AI tools. It is a proposal, not a guarantee: the major assistants have not confirmed that they rely on it. It is cheap to add and harmless, so it is reasonable to have, but it is no substitute for clear pages that are properly indexed.

Why assistants sometimes get small businesses wrong

If there is little clear information about a business, the model fills gaps with plausible guesses. Common mistakes are the wrong city, an outdated service list, or confusing two businesses with similar names.

The fix is not a trick. It is to publish the facts plainly and repeat them consistently: who you are, where you are, what you offer, who it is for and how to contact you. Do that on your site, on your professional profiles and in directories.

A short glossary

• LLM: a large language model, the engine that writes the answer.
• RAG: retrieval-augmented generation, meaning the assistant looks things up before answering.
• Index: the searchable copy of the web that a search system keeps.
• Citation: a link or mention showing which page an answer used.
• Schema markup: code that describes your page and business in a standard vocabulary machines can read.
• Crawler: a program that visits pages and reads them.

What to do with this

Make sure your pages are indexed and open to search crawlers. Put clear answers near the top of the pages that matter. Keep your business details identical everywhere. Then check monthly by asking the questions your customers ask. Most of the work is ordinary good writing, done with a machine reader in mind as well as a human one.