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

# Task Observation

> Let the AI moderator watch a participant do a task on their screen or camera, then ask reflective questions about what happened.

**Task observation** (previously called event detection) lets the AI moderator watch a participant perform a task, then ask a few reflective questions about specific moments once the task is done.

Think of it as a researcher sitting behind the participant while they shop online or handle a product. The researcher does not interrupt and does not count anything. When the participant says they are finished, the researcher picks a handful of moments worth talking about: "You searched twice, then went straight to a comparison page. Walk me through that."

It is a qualitative interviewing feature. It is not a UX analytics tool, and it does not track clicks.

<Note>
  Task observation is in **private beta**. It is enabled by invitation for a small number of organizations while we co-develop it with pilot customers. If **Task observation** does not appear under **Live vision**, contact your Conveo representative. The feature carries a price uplift, shown in the topic guide when you enable it.
</Note>

***

## How it works

Task observation runs on one open-ended question at a time. That question becomes a task.

1. **The moderator gives the task.** For example: "Find a pair of wireless headphones you would buy." Screen sharing or camera access was already granted when the participant joined, so nothing extra is asked of them.
2. **The participant does the task while the AI watches silently.** The moderator never speaks, prompts, or steers during the task.
3. **The participant says they are done.** There is no button. The participant tells the moderator they have finished, like answering any other question. If nothing usable was seen yet, the moderator gently checks in at most twice, then moves on.
4. **The moderator asks about specific moments.** "I noticed you compared two models on Amazon. What made you hesitate?" How many moments, and how deep the follow-ups go, is set by you.

The moderator only asks about things that were actually visible and that match your guidance. If a task produced nothing worth asking about, it asks less or moves on.

***

<a className="compatibility" id="when-to-use-it" />

## What it can and cannot do

Task observation sees exactly what a person would see looking at a screenshot once per second. Nothing more.

| It can                                                                                                                                                                                         | It cannot                                                                                                                                                                              |
| ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Notice which sites, pages, apps, or tools were on screen, read from the visible address bar, page title, or logo.                                                                              | Track clicks, taps, or click-through rates. There is no browser extension and nothing runs inside the websites the participant visits.                                                 |
| Recognize seven kinds of moment: visiting a page, searching, comparing options, using an AI assistant, dwelling on one thing, checking out, and visible friction such as an error or dead end. | Produce heat maps, scroll depth, mouse movement, or time-on-page.                                                                                                                      |
| Ask reflective questions about those moments after the task, phrased from your guidance.                                                                                                       | Calculate conversion rates, funnel percentages, or "what share of participants used the search bar". Moments are qualitative notes on one participant's session, not analytics events. |
| Recognize AI usage inside a page, such as an open chat thread or an AI-generated overview in search results.                                                                                   | Detect frustration or emotion from the screen. "Friction" means a visible error or dead end, not a feeling.                                                                            |
| Work on any website or app without setup, because it needs nothing from the site.                                                                                                              | Probe during the task. The moderator stays silent until the participant says they are done.                                                                                            |
|                                                                                                                                                                                                | Catch anything faster than a second, smaller than readable text, on a second monitor, or on a phone in the participant's hand.                                                         |

**Rule of thumb:** if the research question starts with "what percentage of people…" or "how many times did they…", task observation does not answer it. If it starts with "why did they…" or "what happened when they…", it does.

***

## Turning it on for a question

1. Open the topic guide and select an open-ended question.
2. Open **Live vision** and choose **Task observation**.
3. Choose a **Source**: **Screen sharing** or **Camera**. Each question uses one source.
4. Write your **Guidance**.
5. Open **Observation settings** to set the number of moments and the follow-up depth.
6. Run a full [test interview](/docs/setup-to-launch/testing-your-study) with the intended source before launch.

<a className="compatibility" id="picking-a-source" />

### Choosing a source

| Source             | What Conveo can observe                                                                                                                |
| ------------------ | -------------------------------------------------------------------------------------------------------------------------------------- |
| **Screen sharing** | Whatever the participant shares: a browser, an app, a document. Best for shopping journeys, prototype walkthroughs, and AI tool usage. |
| **Camera**         | The participant's webcam view. Best for physical tasks: handling a product, showing a store shelf, using a device.                     |

Choosing **Screen sharing** turns on the study's screen-sharing requirement if it was off, so the participant shares their screen for the whole interview, not just this question. If you had already chosen **Single window**, that setting is kept. Removing the last screen-sharing task observation question turns the requirement back off, so re-check [device and screen-sharing settings](/docs/setup-to-launch/detailed-topic-guide-settings) after editing.

Screen sharing works in desktop browsers and in the Conveo app on iPhone and iPad. Mobile browsers cannot share a screen, and Android app support is still being verified. Plan screen-sharing tasks for desktop participants.

<a className="compatibility" id="listing-events" />

### Writing guidance

**Guidance** is the only steer you give the AI. It shapes what gets noticed during the task, which moments get asked about afterwards, and how those questions are phrased. Write it as a few sentences, not a checklist, and cover three things:

* **What to watch for.** "Watch which AI tools they open and whether they compare prices across shops."
* **What to dig into.** "Ask why they trusted or ignored the AI answer, and what made them switch shops."
* **What to leave alone.** "Do not ask about delivery times."

Guidance sharpens the moderator's attention. It cannot add new kinds of detection: asking it to "count how many times they click the search bar" or "check whether they notice the sponsored label" will not work, because neither is visible in a screenshot.

Changes to guidance apply to interviews that start after you save.

<a className="compatibility" id="probing-depth-per-event" />

### Observation settings

| Setting              | Options                                                                  | What it does                                                                                               |
| -------------------- | ------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------- |
| **Minimum moments**  | 1, 2, or 3                                                               | The moderator asks about at least this many moments, as long as the task produced usable evidence.         |
| **Maximum moments**  | 1, 3, 5, 7, or 10                                                        | The moderator never asks about more than this. Also adjustable from the moments pill next to the question. |
| **Depth per moment** | No follow-up questions, 0–1 follow-ups, 1–2 follow-ups, or 2+ follow-ups | How far the moderator digs into each moment.                                                               |

Keep the maximum low. Three moments with one follow-up each is already six questions on top of the task itself. Ten moments with 2+ follow-ups makes for a very long interview.

***

## What the participant sees

The participant hears the task question and performs it on their shared screen or camera. They do not see your guidance, and they are not told which moments were noticed.

While they work, a small **Analyzing…** indicator shows that Conveo is watching. When they say they are finished, the moderator asks its questions, then continues with the interview.

Write the task so the participant knows what "finished" means: "find one you would buy and tell me when you have it" works better than "browse around". Do not tell participants to "start sharing your screen": sharing was already set up when they joined, and the instruction makes people think it is not working.

***

## Tips

* **One task, one natural end.** "Find and add to cart" or "pick the one you would buy" gives the participant a clear moment to say they are done.
* **Match the source to the task.** Camera observation cannot inspect an app screen that is not visible to the camera, and screen sharing cannot see a phone in the participant's hand.
* **Describe visible behaviour in guidance, not inner states.** "Compares prices across shops" is observable. "Feels confused" is something to ask the participant about.
* **Test it on yourself first**, and tell your client it is a beta. If the moderator asks too many questions, lower **Maximum moments** or the depth.