# Proof

> How Prefera checks whether a published fix moved AI answers, using targeted and control questions, a re-check after 7 days and a follow-up after 21.

Source: https://prefera.si/docs/concepts/proof
Docs index for agents: https://prefera.si/llms.txt

Proof answers one question: after you published a fix, did AI engines start recommending your brand more than they would have anyway? It compares the questions the fix aimed at with similar questions you left alone. It does not promise a result, and a fix that did not work stays on the page.

## Targeted and control questions

- **Targeted questions** are the tracked questions a fix was written for.
- **Control questions** are similar questions Prefera picks before you publish and leaves alone. Every fix gets up to four of them.

AI answers move from week to week even when you change nothing. The control group shows how much that is. Prefera credits the fix only with the change in the targeted questions beyond what the control questions did on their own.

A **flip** is a sample that moved to recommending your brand, not just naming it.

## When the re-check happens

| Check | When | Compared with |
| --- | --- | --- |
| Baseline | The last regular run before the fix goes live | Starting point |
| Re-check | The first run at least 7 days after it goes live | The baseline |
| Follow-up | The first run after 21 days | Confirms the change holds |

Prefera reads these from your regular tracking runs, with the same country and the same engines, so no extra answers are used. If your regular runs do not cover a question, Prefera asks the targeted and control questions itself, 3 samples per engine.

While a fix is being measured, its targeted questions stay on your scheduled runs until the follow-up is done. They do not use your question slots. They show under **Measuring for Proof** on the **Question map**.

## Verdict labels

Each fix with a result gets one of these labels.

| Label | Meaning |
| --- | --- |
| **Flipped N** | The fix moved N samples beyond the control group, and the gap is larger than the noise band. |
| **Within noise** | The gap is no bigger than the noise band, so Prefera does not call it a result. |
| **Lost ground N** | The targeted questions fell behind the control group by more than the noise band. |
| **Not enough data** | There were too few valid answers to compare fairly. |

The noise band is at least 2 samples. It gets wider when there are few samples, or when an engine gives the same answer to a question every time. With fewer than 6 valid answers in any group, Prefera says there is not enough data. A fix with no control questions available cannot be given a result. Prefera shows "No control questions were available for this fix, so no result can be claimed."

A fix that is waiting for its re-check has no label yet and is listed as **Waiting**.

## Not measured

Some pieces are **Not measured: no questions targeted.** This happens when a piece is not tied to any tracked question. For example:

- you create it from a topic in your own words;
- you create it from a page, and that page does not target tracked questions.

There is no baseline and no re-check for these. When you mark one as published, the message says "Not measured: no questions targeted." The piece is simply live.

To have a piece measured, create it from tracked questions. See [Questions, answers and engines](/docs/concepts/tracking).

## Where to see it

1. Open **Results** in the sidebar.
2. Press **Proof** in the bar at the top.

The page opens with a headline written from your results. Filter the table with **All**, **Flipped**, **Within noise**, **Lost ground** (shown when there is one) and **Waiting**. Choose a period: **Published in the last 30 days**, **Last 90 days** (the default) or **All time**. Open a fix to see each targeted question and its control questions before and after.

The page also has **How we count a flip**, which explains the method, and **Coming up**, which lists the next check for each live fix. You can export the table as CSV. Each fix on **Results** also shows its Proof result.