GEO guide · Chapter 2 of 9

Where do ChatGPT, Gemini, Claude, Copilot and Perplexity get their facts?

For anyone who has asked an assistant about their own business and been surprised by the answer. After reading, you will know which sources each assistant draws on and where to fix wrong or missing information.

Updated November 6, 20264 min readBy Pop & Soda

In one paragraph

AI assistants get facts from two places: what the model learned in training, and live web searches run at answer time. ChatGPT search draws on its own crawler and partners including Bing, Copilot uses Bing, Gemini and Google's AI Overviews use Google Search, Perplexity runs its own index, and Claude can search the web when the feature is on. Say a pet food brand launched 3 new recipes in March: a model trained earlier won't know them, but a search-backed answer can.

Training data is the model's memory, and it has a cutoff date

Everything a model knows without searching comes from its training data, which stops at a cutoff date. Training data is a very large collection of text, drawn from sources such as public web pages, licensed content and public archives like Common Crawl. If your business was written about widely and consistently before that cutoff, the model has a default picture of you. If not, it may know nothing, or it may confuse you with something else.

You cannot edit a model's memory directly. You can only change what is published about you, which future training rounds may pick up. Vendors release new models every several months, and each one has its own cutoff, so memory improves in steps rather than overnight.

Live search fills the gaps, and it is where fast wins happen

When an assistant searches the web before answering, it can use information published yesterday. This step is called retrieval or grounding. The assistant turns your question into one or more search queries, reads a handful of top results, and writes an answer from them, often with numbered links to its sources.

Most assistants decide on their own when to search. Questions about prices, opening hours, recent events, local businesses or 'best of' recommendations usually trigger a search. General knowledge questions often do not. That means your pages need to rank well enough to be among the few that get read.

Worked example

a DTC skincare brand

Say a skincare brand cut the price of its best-selling serum from $38 to $32 and changed the formula last spring. A customer tells the founder that ChatGPT still quotes $38 and the old ingredient list. The founder asks ChatGPT twice: once in a plain chat, and once with search switched on. The plain answer repeats the old facts, which come from model memory. The search answer cites two retailer listings that were never updated.

The fix splits in two. For live search, the brand updates the retailer listings, adds the current price and ingredients as plain text on its product page, and confirms the page is indexed in both Google and Bing. Within a few weeks, search-backed answers show $32. For model memory, the brand can only keep its facts consistent everywhere and wait for future model updates to catch up.

Each assistant leans on a different search index

Being visible in Google alone is not enough, because several assistants rely on other indexes. A search index is the catalog of pages a search engine has collected and can return. Here is the broad picture as of late 2026; vendors change details often, so check their own documentation.

ChatGPT: model memory plus ChatGPT search, which uses OpenAI's own crawler (OAI-SearchBot) and third-party search providers, with Bing widely reported as one.
Microsoft Copilot: Bing's index.
Gemini and Google AI Overviews and AI Mode: Google Search's index.
Perplexity: its own crawler (PerplexityBot) and index, combined with other search sources.
Claude: model memory plus a web search feature that the user or app can switch on.
Meta AI and Grok: model memory blended with web results, and Grok also draws on posts from X.

Third-party sources often outweigh your own website

Assistants frequently cite review sites, directories, forums, news articles and roundup posts ahead of a brand's own pages. They do this because independent sources look more trustworthy for questions like 'which is best'. Reddit threads, Google reviews, Yelp, G2, Wikipedia and industry publications appear in answers again and again.

This is good news and bad news. If an old directory listing shows your former address, the assistant may repeat it. If a respected roundup praises you, the assistant may repeat that too. Chapter 6 covers how to earn those mentions.

Do this next

Verify your site in both Google Search Console and Bing Webmaster Tools, and submit your sitemap to each.
Search your brand name on Bing as well as Google and fix anything missing or wrong.
Ask each major assistant 'What do you know about [your business]?' with and without web search, and note the differences.
Record which URLs the assistants cite when they talk about you or your category.
Update or claim outdated third-party listings, starting with the ones assistants cite.
Check that your robots.txt allows Bingbot, Googlebot, OAI-SearchBot and PerplexityBot.

Quick answers

What is GEO (generative engine optimization)?

Questions people ask

Can I submit my website to ChatGPT?

There is no submission form. Make sure your site is indexed in Bing and Google, allow OAI-SearchBot in robots.txt, and keep key facts in plain text so ChatGPT search can read them.

How often do assistants update what they know?

Live search results update as often as the underlying index, which can be days. Model memory updates only when the vendor releases a new model, which typically happens every several months to a year.

Do Gemini and Google AI Overviews use the same sources?

Both draw on Google's systems and Google Search's index, though they are separate products and can give different answers. If you rank well in Google Search, you have a better chance in both.

Start with a free scan.