---
title: "Look something up on the web in an automation"
description: "Add a Web Search step so an automation searches the live web with Tavily or Perplexity, then writes a cited answer into one of your custom fields."
---

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

# Look something up on the web in an automation

A **Web Search** step searches the live web while an automation runs and writes what it finds into
one of your [custom fields](/help/insights/fields/), with the sources listed underneath. Use it to
check a claim a caller made, pull a company's current headcount before a sales review, or confirm a
regulation a recording refers to.

The search reads the web. It does not read your recordings. To ask questions of your own media, use
a [Speak AI Chat step](/help/automations/actions/) instead.

## How do you add a Web Search step?

1. Open the automation on its canvas.
2. Select the **+** where the step belongs, then pick **Speak** as the app and **Web Search** as the
   action.
3. Fill in the step on its **Configure** tab, then save the automation.

## What do you set on the step?

- **Search provider**: **Tavily** or **Perplexity**. Pick this first, because it decides which
  **Search effort** choices appear below.
- **Search query**: what to search for. Required. Insert a token here to build the query from the
  recording, so one automation serves every file.
- **Save answer to field**: the custom field the answer is written into. Required, and it takes one
  field, not several.
- **Country**: narrows the search to one country. Optional, and it accepts a token, so a single
  automation can follow a field that already holds the right country. On Tavily it is ignored
  rather than rejected in two cases: when **Search effort** is **Fast** or **Ultra fast**,
  which do not accept a country at all, and when the country is outside Tavily's supported
  list. The search still runs, it just is not narrowed, so pair Country with **Basic** or
  **Advanced** when the narrowing matters.
- **Search effort**: how hard the provider looks. Tavily offers **Ultra fast**, **Fast**, **Basic**
  and **Advanced**. Perplexity offers **Fast**, **Low**, **Medium**, **High** and **Very high**.
  More effort means a slower, more thorough search.
- **How the AI should answer (optional)**: replaces the instructions Speak gives the model when it
  writes the answer up. Leave it empty and Speak uses its own, which tells the model to judge only
  what the sources support and to say so when the evidence does not settle the question.
- **AI Model**: which model writes the answer up from the sources. Optional; leave it empty and the
  step uses your workspace default. It does not affect the search itself, only the write-up.

## What lands in the field?

The answer, then a **Sources:** list of the pages it came from, each with its title and URL:

```text
The company reports 480 employees as of its most recent filing.

Sources:
1. Example Corp 2026 annual report
   https://example.com/investors/annual-report
2. Example Corp company profile
   https://example.com/about
```

Writing that field behaves like any other field an AI step fills in, so a **Field updated**
[trigger](/help/automations/triggers/) watching it fires as normal.

## Does a web search cost credits?

Writing up the result runs through the same engine a Speak AI Chat step uses, so it draws on your
[credits](/help/account/credits/) the same way. Price a representative run before you turn the step
on across a busy folder.

## Can a Web Search step run on a schedule?

No. The step writes its answer to one file's field, so it needs a run that is about a single file.
A scheduled automation processes a batch at once and the step stops with:

"Web search needs a single file to write its answer to, so it cannot run on a batch or scheduled run."

Trigger the automation per file instead, for example on **Media analyzed in folder**.

## Why did my web search step fail?

A Web Search step stops the run only when something is genuinely broken, not when it has nothing
to search for or nothing to report. The reason appears on the step in the automation's run
history:

| What you see | What it means |
|---|---|
| "The field this web search writes to no longer exists. Pick another destination field." | The destination field was deleted after the step was saved. |
| "Web search needs a single file to write its answer to, so it cannot run on a batch or scheduled run." | The automation runs on a batch or a schedule. |
| "No ... key available. Connect ... in Integrations to use web search." | Neither your workspace nor Speak has a key for that provider. Connect one in Integrations, or switch the step to the other provider. |
| "Web search could not summarise its results" | The search returned sources, but the model writing them up failed. |

A provider that rate limits or fails temporarily is retried once before the step gives up, so a
single blip does not stop the run.

## What happens when a search finds nothing?

A search that finds nothing completes the step and writes a note into the field, so the run
reaches the end and any step after it reads an explicit finding rather than a gap:

```text
No sources were found for this search.

Query: current headcount at Example Corp
```

Two cases produce that note. The provider returned no results at all, or Perplexity answered
from its own knowledge instead of searching, which leaves nothing to cite. Perplexity's own
explanation sits above the **Query:** line when that is what happened.

A search that ran and found evidence that does not settle the question is different again. It
succeeds as normal, and the model writes up what the sources do and do not support.

## What happens when there is nothing to search for?

Speak skips the step, leaves the field exactly as it was, and carries on with the rest of the
run. Two cases reach it:

- A token in **Search query** resolved to nothing on this file, so there is no query to send.
- The query resolved to `NO_QUERY`, which an earlier step can write into the field the query
  reads when a file has nothing worth looking up. It has to be the whole query, so a search that
  merely mentions the word still runs.

A skipped step is recorded as completed in run history. An untouched field and a field reading
"No sources were found for this search." mean different things: the first means nothing was
asked, the second means Speak looked and found nothing.

## Can a later step use the answer?

Yes. A Web Search step publishes its answer as a token, the same way a Speak AI Chat step does, so a
step after it can quote the answer in its own prompt. Insert the token from the picker rather than
typing it. You can also branch on what it found, by writing to a field and then adding a
[branch](/help/automations/branches/) that tests that field.

The fastest way to reach us is the live chat in the app (the chat bubble in the bottom corner). You can also email [success@speakai.co](mailto:success@speakai.co).

Want a hand setting this up? [Book a free consult](https://calendly.com/speak-ai/consult?utm_source=docs&utm_campaign=consult) and we'll do it together on your account.

---
Related: [Automations](/help/automations/) · [Actions](/help/automations/actions/) ·
[Branches](/help/automations/branches/) · [Variables](/help/automations/variables/)

Source: https://docs.speakai.co/help/automations/web-search/index.mdx