How AI searches the web: LLMs with live search explained
Short answer An AI chatbot answers from training memory, which ends months before today, unless it decides the question needs current facts. Then it turns the question into short searches, runs them on a search index, reads some of the pages and writes an answer citing them. Your business appears only if those searches find your pages and can read them.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callHow does an LLM search the internet?
It doesn't browse like you do. A large language model (LLM), the engine inside ChatGPT, Claude, Gemini or Copilot, answers first from what it learned in training. When a question needs fresher facts, it runs a lookup loop:
| Step | What the AI does |
|---|---|
| 1. Decide | Judges whether a search would improve the answer |
| 2. Rewrite | Turns your question into a few short search queries, adding your rough location for local questions |
| 3. Search | Sends them to a search index: a stored copy of the web, not the live web |
| 4. Read | Reads the results, and sometimes opens a page at that moment |
| 5. Repeat | Searches again, more narrowly, after reading the first results |
| 6. Answer | Writes the reply and cites the pages it used |
Google's developer docs describe step 1 for Gemini:
"The model analyzes the prompt and determines if a Google Search can improve the answer." (Google AI Studio docs: Grounding with Google Search)
OpenAI describes steps 2 and 5 for ChatGPT: it "typically rewrites your query into one or more targeted queries", then "may send additional, more specific queries" after reviewing the first results. Microsoft says Copilot's query "consists of a few words informed by the user's prompt" (Microsoft Learn).
Researchers call this retrieval-augmented generation (RAG): look things up, then write. The 2020 paper that named it paired a language model with a searchable index of Wikipedia (Lewis et al.). Google calls the same idea grounding.
What it means for you: the AI can only cite what step 3 returns. A page those short searches don't find, or a crawler can't read, isn't in the answer.
When does ChatGPT search, and when does it answer from memory?
It searches when the answer may have changed since training. OpenAI: "ChatGPT may search the web automatically when your question would benefit from current information." You can also force it by choosing Search in the tools menu. Web search works on every ChatGPT plan, Free included, and for people who aren't signed in (OpenAI Help Center).
Anthropic's developer docs list when Claude searches and when it doesn't:
| Usually searches | Usually answers from memory |
|---|---|
| Recent news and announcements | Settled facts, math, science basics |
| Current prices, rates, scores | Writing and brainstorming |
| Organizations, people or products that might have changed | Text you pasted into the chat |
| Any request to look something up | Greetings and small talk |
The third row includes your business. Memory has an end date: OpenAI's newest models list knowledge cutoffs between April 30 and June 1, 2026, and Anthropic's current Claude models June 2026 (OpenAI API docs: Models; Claude docs: Models overview). A price, address or service you changed after that reaches an AI answer only through a search. If the chatbot skips the search, it may repeat what was true before.
Which search engine does each AI use?
Each searches a different index, so being found by one doesn't mean being found by all. Setup for each is in how to rank in Perplexity, Gemini and Claude.
| AI | Where its searches go | What reads your site |
|---|---|---|
| ChatGPT | Partner search providers, Microsoft among them, plus OpenAI's own search | OAI-SearchBot builds the index; ChatGPT-User opens pages when someone asks |
| Copilot | Bing | Bingbot |
| Gemini | Google's Search index | Googlebot |
| Claude | Anthropic's web search tool | Claude-SearchBot crawls for search; Claude-User opens pages when someone asks |
| Perplexity | Its own search | PerplexityBot builds the index; Perplexity-User opens pages when someone asks |
Google defines Gemini's grounding as "providing content from the Google Search index to the model at prompt time" (Google's common crawlers). Perplexity says it searches "in real-time" (Perplexity Help Center).
Does the AI read my website live?
Mostly it reads a saved copy. "Live search" usually means searching an index, which holds each page as the crawler last saw it. Some assistants also open a page when someone asks: OpenAI says ChatGPT "may visit a web page with a ChatGPT-User agent." Anthropic and Perplexity run similar fetchers.
Three things follow:
- Changes take time to show. The AI cites the copy it has. Microsoft says IndexNow, a way to tell Bing and other engines that a page changed, "helps ensure that AI systems reference the most current version of a page when generating answers" (Bing Webmaster Blog, February 2026).
- The bot must get through. OpenAI asks site owners to confirm "the website host or content delivery network allows traffic from OpenAI's published searchbot IP addresses." A firewall or bot-blocking setting can stop these bots before they read anything. See should you block AI crawlers?
- It reads text, not scripts. In a December 2024 study of traffic on Vercel's network, ChatGPT's and Claude's crawlers downloaded JavaScript files but didn't run them, so anything a script draws on the page was invisible to them; Gemini, using Googlebot, did run them (Vercel). Google lists "Making sure that important content is available in textual form" among the basics for its AI features.
So put prices, service areas and hours in the page's text, not only in images, PDFs or widgets.
Why do AI answers with sources still get things wrong?
A citation shows where the AI looked, not that it read correctly. OpenAI's own help page says so:
"Search results and citations can be incomplete, outdated, or incorrect." (OpenAI Help Center: Searching the web with ChatGPT)
| Where it slips | What happens |
|---|---|
| The query is too short | A few words miss what made the question specific |
| The saved copy is old | You changed the page; the crawler hasn't been back |
| The source is wrong | An old directory listing still shows your previous address |
| No search ran | The AI answered from training data that predates your change |
| Personal context | ChatGPT may use saved memories when it rewrites a search, so two people can get different results (OpenAI Help Center) |
You can't change what a model learned in training. You can fix the page or listing it cites. To see what ChatGPT says about you now, run the 20-question test.
Can I add live web search to my own chatbot?
Yes, through the AI companies' developer tools. One common tip is out of date: Microsoft retired the Bing Search APIs on August 11, 2025 and points customers to "Grounding with Bing Search as part of Azure AI Agents" (Microsoft Lifecycle).
| Provider | Tool | Note |
|---|---|---|
| OpenAI | Web search tool in its API | A reasoning model "can refine its searches as it reads results" |
| Anthropic | Web search tool in the Claude API | US$10 per 1,000 searches plus usage; a single fact typically takes 1–3 searches |
| Grounding with Google Search in the Gemini API | Returns the answer with citations and the search queries used | |
| Microsoft | Grounding with Bing Search in Azure AI Agents | The replacement for the Bing Search APIs |
A chatbot on your own website usually needs the opposite: answers from your own pages, not the open web. Anthropic's tool, for one, can be limited to domains you list.
Example
Illustrative cases with invented numbers.
A dental clinic in Denver (US dollars)
It raised its cleaning price from US$120 to US$140 in September 2026, after the newest models' cutoffs. Asked without search, a chatbot may quote US$120; with search, it reads the price page. The fix: the new price as text on the site and on every listing, and IndexNow so Bing picks up the change sooner.
A roofer in Calgary (Canadian dollars)
The quote widget shows "replacements from C$9,000" only after a script runs. A crawler that doesn't run scripts sees an empty box. The fix: one plain sentence under the widget with the price range.
A law office that moved
Perplexity cites an old directory listing with the previous address. The firm's own site is right, but the directory came up first for the short search the AI ran. The fix: correct the directory, then ask again a week later.
FAQ
Does ChatGPT have access to the internet?
Yes, on every plan, Free included, even signed out. It searches when a question needs current information, or when you choose Search.
Is ChatGPT's information up to date?
Only when it searches. Otherwise it answers from training data; OpenAI's newest models have knowledge cutoffs between April and June 2026.
What is RAG?
Retrieval-augmented generation: the AI fetches documents first, then writes from them. Google's AI features retrieve "up-to-date web pages from our Search index."
Can I see what the AI searched for?
Sometimes. Copilot Chat shows "the exact web search queries" it sent to Bing; Bing Webmaster Tools (preview) shows a sample of AI grounding queries.
Different for you?
- You don't know whether your firewall or robots.txt blocks AI bots: should you block AI crawlers?
- You want each assistant's index and listings: how to rank in Perplexity, Gemini and Claude
- ChatGPT names other businesses for local questions: where ChatGPT finds local businesses
- You never set up Bing, which feeds Copilot: Bing Places and Copilot
- You want to see what ChatGPT says about you today: the 20-question test
- You'd rather have your pages kept current and readable for every engine: SEO & GEO Agent