# AI Web Search and Sources

> Why a model stops learning at a cutoff date. How web search lets it read today's pages, and how to check the links it gives you.

**Course:** [AI Assistants and Chatbots](https://learn.benibix.com/learn/ai-use-case)  
**Time:** 10 min

**After this lesson you can:**

- Explain what a knowledge cutoff is
- Say what web search adds to an answer
- Check an answer against the links it shows

## Why does an AI assistant need web search?

A model only knows what was in its training data, the text it learned from. That text stops at a date called the knowledge cutoff. The newest Claude models, for example, have a cutoff of June 2026. Web search lets an assistant look things up online, read the pages and answer from what is there today.

Training a model takes months, so its cutoff is always behind the day you use it. Ask a model with no search who won last night's game, and it cannot know. Worse, it may still answer. It writes a smooth reply that sounds right but is made up. That is called a hallucination.

With web search on, the assistant writes its own search words, much as you would type them. It reads the pages that come back and picks the parts that answer you. Then it writes a reply from those parts, not just from memory. Tying an answer to real sources like this is called grounding.

A good answer shows the links it used, often as small numbers after a sentence. Those links are the point. Click one and check that the page says what the answer says. An answer can still mix up two pages, or lean on a weak one, even when the links look neat.

Some assistants search on their own when a question needs fresh facts. Others wait for you to switch search on. If an answer is about news, prices, laws or anything else that changes, check whether it searched. If no links appear, it is answering from memory, and its memory stops at the cutoff.

Web search is not free of risk. The pages it reads are written by strangers, and some are wrong. A few even hide text meant to trick an AI. Search makes an assistant more up to date, not more careful. You still decide which sources to trust.

![A timeline. Training text covers the time up to the cutoff date. After it there is a gap, and events in the gap are unknown to the model.](https://learn.benibix.com/diagrams/llm-knowledge-cutoff.svg)

### Example: Emma checks a train fare

Emma asks her assistant what a train ticket to the coast costs. With no search, it says $45, a price from before its cutoff.

She turns web search on and asks again. Now it reads the rail firm's page and says $52, with a link. She clicks it, and the page says $52.

**Words in this lesson:** [Web Search (AI)](https://learn.benibix.com/term/ai/web-search), [Training Data](https://learn.benibix.com/term/ai/training-data), [Knowledge Cutoff](https://learn.benibix.com/term/ai/knowledge-cutoff), [Hallucination (AI)](https://learn.benibix.com/term/ai/hallucination), [Grounding (AI)](https://learn.benibix.com/term/ai/grounding)

**Question:** An answer about this week's news shows no links. What is most likely?

**Answer:** It answered from memory, which stops at its cutoff. An answer built from web search shows the pages it read. With no links, the model answered from its training data, which ends at the knowledge cutoff, so it may not know this week's news at all.

---

Source: https://learn.benibix.com/learn/ai-use-case/ai-web-search-and-sources
Author: Behnia Zarei (a developer in Vienna, Austria, who builds and teaches data, AI and business intelligence at BENIBIX)
Publisher: BENIBIX Learn (https://learn.benibix.com/)
Last updated: 2026-11-18
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote freely with attribution.
