# Web Research (/docs/concepts/web-research)



Most Crowdee verification pipelines reason from an AI model's training knowledge — fast, but limited to what the model already "knows." Web research stages are different: they perform a **live search or crawl of the current web** before handing the result to an LLM for analysis. This is what powers [Source Research & Credibility Assessment](/docs/pipelines/verify-source-credibility) and any future pipeline that needs to trace a claim back to its actual current source, rather than reason about it in the abstract.

## How a Web Research Stage Runs [#how-a-web-research-stage-runs]

<Steps>
  <Step>
    **Resolve the query**

    The stage inspects the run's context data and determines either a specific URL to fetch directly, or a search query to run when no URL is already known.
  </Step>

  <Step>
    **Search or fetch**

    If a URL is known, the stage fetches that page directly. Otherwise it searches the web and retrieves the full content of the top matching results.
  </Step>

  <Step>
    **Analyze**

    The retrieved page content — title, URL, and cleaned text — is passed to an LLM alongside the stage's task-specific instructions. The model's output is validated against the stage's structured output schema, same as any AI stage.
  </Step>
</Steps>

If a specific result fails to load (dead link, blocked, timeout), that result falls back to its search snippet rather than failing the whole stage — the stage only fails outright if no sources could be gathered at all.

## Failure Modes [#failure-modes]

| Condition                                | Behavior                                                           |
| ---------------------------------------- | ------------------------------------------------------------------ |
| A fetched page returns no usable content | That source is dropped; the stage continues with remaining sources |
| No sources could be gathered at all      | Stage fails; the pipeline run transitions to `failed`              |

<Callout type="info">
  Because a web research stage reflects the live state of the web rather than a fixed corpus, re-running the same pipeline on the same subject at a later date can produce a different result if the underlying web content has changed. This is expected — the stage is tracing *current* sources, not returning a cached judgment.
</Callout>