> ## Documentation Index
> Fetch the complete documentation index at: https://docs.erdo.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Web Research

> How Erdo agents research facts on the web: dated sources, a sourcing rule on every web tool, and a claims table you can check.

# Web Research

When you ask an Erdo agent something that needs facts from the web, such as a
price sheet, a list of a company's projects or a funding round, it searches and
reads pages itself. Research that reaches someone else has to be right and show
where it came from, so the web tools are built around sources and dates.

## Every source carries its dates

* **Search results** show the date the search engine reports for each result,
  when it has one. Some are exact ("Oct 20, 2026") and some are relative
  ("1 year ago").
* **Pages the agent reads** show when the page was published, when it was last
  modified (when the page says) and when Erdo read it.

An empty date means unknown. The agent uses these dates to prefer the newest
source and to say what date a live figure, such as a current price or a sales
figure, is as of.

## The sourcing rule

Both web tools carry the same short rule, so every agent that can search the web
follows it:

* Prefer the owner's own page or release, and the newest dated source.
* Corroborate with an independent source.
* Keep the URL, the date and a short verbatim excerpt for each claim.
* State a fact only from a page the agent read itself, or from a source and
  excerpt it was handed. Otherwise label it unverified.
* Show conflicts rather than picking one value.
* Label a lone aggregator, broker or undated figure as unverified.

## The research skill

Erdo ships a default skill, **Researching external facts**, that agents find
when a task calls for research. It has the agent:

* write down the questions the answer must settle, what an authoritative
  source looks like and how fresh each fact has to be, before the first
  search;
* start from the owner's own list rather than search snippets;
* keep ranges and qualifiers exactly as the source states them;
* for larger research, keep a **claims table** as a dataset in the
  [conversation](https://docs.erdo.ai/conversations.md), with each claim's value, source URL, excerpt,
  dates and a status (`verified`, `single_source`, `unverified` or `conflict`).

Like any default skill, your organization can edit its copy. See
[Knowledge](https://docs.erdo.ai/knowledge.md).

## Checking the work

While the agent works, the claims table's card in the conversation shows how many
rows it holds. While rows are landing it says "Setting up" for a new table, or
"Refreshing" for one that already had rows. Open it to see
each fact with its source. A
fact marked `single_source` or `unverified` is the agent telling you it could
not settle it. Treat it as a lead to check, not a fact.

Before sending factual research outside the conversation, ask for an independent fact check as described in [Fact checks](https://docs.erdo.ai/docs-public/fact-check). The review researches the document text and reports its evidence without publishing or mutating the document.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.