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

# The analytics overview

> Analytics opens on four questions: are we getting better, where are we losing and to whom, what did we do and did it work, and what changed out there. Every answer carries its n.

**Analytics → Overview** is where the analytics open. It answers four questions, in the order they are usually asked:

1. [Are we getting better?](#are-we-getting-better)
2. [Where are we losing, and to whom?](#where-are-we-losing-and-to-whom)
3. [What did we do, and did it work?](#what-did-we-do-and-did-it-work)
4. [What changed out there?](#what-changed-out-there)

Each question is one block. Every block has **Ask about this**, which opens a chat with the agent on that block's own numbers, and **See all**, which goes to the tab the block summarises (on question 4, each of its three lists has its own **See all**).

The window control above the blocks sets 7, 30 or 90 days, and the window is part of the URL, so a reload or a pasted link lands on the same window. Arriving from a row of the **Markets** table narrows the overview to that market, and the header offers the way back to every market.

## Are we getting better?

One line: how often assistants name you (your [share of voice](/measure/share-of-voice)) per day across the window, with the number of answers read and one sentence under it.

### The sentence

The sentence is written in one place, on the server, so its wording cannot drift between surfaces. It always states **n**, and it says "a real change" only when the movement is one we can prove:

| What you see | What it means |
| - | - |
| *Named in 31% of answers, up from 26% (n=1,840), a real change.* | The change against the previous window of the same length cleared the [significance test](/measure/reading-the-numbers#why-does-a-change-say-real), and the window was read on at least 70% of its days. |
| *Named in 31% of answers (n=1,840), nothing we can prove yet.* | Any difference from the previous window is inside the noise of the sample. |
| *Named in 31% of answers (n=212, read on 9 of 30 days), nothing we can prove yet.* | The difference passed the test, but fewer than 70% of the window's days were read (or, when it says *coverage unknown*, how many days were read is not known). Two thin samples weeks apart are not a trend. |
| *No answers were read in this window yet (n=0), so there is nothing we can prove yet.* | Nothing was read. This is not a zero. |

If the markets read changed between the two windows, the sentence compares only the markets read in both, by country code, and names the rest: *Over the markets read in both windows (US, GB): named in 31% of answers (n=1,640), nothing we can prove yet. DE added, so left out of the change.* A market read before and not now is named as *no longer read*. When no market was read in both windows, the sentence says there is nothing to compare it with yet. Under the chart, a line says the same in full: *Germany added in this window. The change is measured over the markets read in both windows, so a new market does not show up as a move.*

The sentence never says why the number moved. Whether a change of ours moved it is question 3, and even there the answer is movement observed after the change, not proof of cause.

### Markers

A marker sits on each day one of our changes went live. Several changes on one day are one marker that names them. Delivered work that is not on a live page yet, such as a draft nobody has published, gets no marker, because no engine could have read it.

Hovering a day shows its rate and the n behind it.

### Goals

Under the line, each active goal is a chip with its current value, its target and whether it is on pace. A goal whose value was not measured shows a dash, not a zero. The chip opens the goal on the **Goals** tab.

## Where are we losing, and to whom?

The prompts in state **lost**: across the engines read in the window, you are named on none and a rival is named on at least one. See [Prompts](/tracking/prompts) for how a prompt's state is set.

The block shows the five worst, and says how many there are in total. Worst means:

1. lost on the most engines;
2. then the most searched (a prompt with no known search volume ranks after every prompt with one);
3. then the order of the prompts board.

Each loss shows:

* **The prompt**, which opens its drawer, with how many engines it is lost on, its topic, and how often you were named, with n.
* **Who is named instead**, on the engine where the loss is worst.
* **What decided it**: the [deciding source](/measure/deciding-sources) that carried the most weight on that engine, or across every engine if nothing was cited on that one. Your own domain is never named here. If no answer cited anything, the row says so.
* **What we are doing about it**: the latest work on that prompt and where it stands (Proposed, Approved, In progress, Declined, Delivered, or Paused at the spending cap), linked to the work item. With no work on the prompt, the row offers **Start**. Every row also links to the prompt's [win plan](/measure/win-this-answer).

**Start** does not create anything. It opens a chat with the loss written in: the prompt, how many engines it is lost on, the rival on the worst engine and how often you were named there, the deciding source and its citations, and a request to read the win plan, propose the first piece of work, and ask you before starting anything that spends. Nothing is sent until you press send.

If the work log for a prompt could not be read, the row says so and does not offer **Start**. Offering to start work that may already exist is how the same work gets paid for twice.

## What did we do, and did it work?

The last five changes that shipped in the past 90 days, newest first by the day they went live. Each change shows its kind, the day it went live (or was delivered, for work that is not on the web), the named rate before and after with n on each side, a verdict, and the reason for the verdict in one line.

The 90 days do not follow the window control. A change's last re-read lands about 34 days after it goes live, so a shorter look-back could only ever say **Too early**.

### The verdict words

The verdict comes from the change's [re-reads](/measure/before-and-after), read across every tracked prompt and engine the change was meant to move. These are the four words, with the text you see when you hover each one:

| Verdict | On hover |
| - | - |
| **Moved up** | After this went live, the prompts it was meant to move were named more often, by more than the noise in the sample. That is movement after the change, not proof the change caused it. |
| **Moved down** | After this went live, the prompts it was meant to move were named less often, by more than the noise in the sample. That is movement after the change, not proof the change caused it. |
| **Nothing we can prove yet** | We re-read the prompts it was meant to move. Any difference was within the noise in the sample, or there was too little measurement before it went live to compare against. |
| **Too early** | We re-read the prompts a change was meant to move 7, 14 and 30 days after it goes live. None of those re-reads has a verdict yet. |

How a change gets its word:

* **The latest re-read with a verdict decides.** The +30-day re-read replaces the +14, which replaces the +7, because a later window has more evidence. They are not averaged.
* **Moved up** when more prompt-engine pairs were named more often than less often, past the significance test. **Moved down** is the reverse. When some pairs went the other way, the reason line says how many.
* **Nothing we can prove yet** when no pair moved past the test, when pairs moved up and down in equal numbers, when there was too little measurement before the change to compare against, when the prompts it was meant to move are no longer tracked, when the change was never found on the live page, when its re-reads were stopped, when no tracked prompt is tied to it, or when no re-read is on record for it.
* **Too early** while no re-read has a verdict and one is still booked or being read (the reason gives the date of the next booked one), and for delivered work that is not live yet, whose re-reads are booked once it is.

The same four words, with the same rules, appear on a prompt's changes timeline in its drawer. There the verdict is about that prompt's engines only.

### Committed vs delivered

Under the changes, a **Committed vs delivered** line compares the work committed for the current period with the work delivered, counted in work units (articles, fixes and pitches), never in money. Committed and delivered work is not reported yet, so the line reads *Not available yet. Committed and delivered work is not reported yet.* in place of the counts.

## What changed out there?

Three lists of things that happened outside your site in the window: **Competitor moves**, **New deciding sources** and **Wrong claims about you**.

None of the three is reported yet. Each list shows its heading and *Not available yet. We are not watching for this yet.* instead of an empty list, because an empty list would say we looked and found nothing.

## When a block cannot be read

Every block is read on its own, with its own time limit, so one slow or failed read never takes the others down. A block is in one of three states:

* **Read**: the numbers, or a sentence saying the window had none ("No prompt is lost to a rival in this window").
* **Not available**: the data behind the block is not reported for this workspace yet. The block says *Not available yet.* and what is missing.
* **Failed**: the read went wrong or took too long. The block says *We couldn't load …* with the reason and a **Retry**. Nothing is lost by retrying; it is a read.

On every block, a dash and a zero mean different things:

> A dash means not measured: no answer was read for it. A 0 means we read the answers and the count was zero.

## Export

**Export CSV** in the title row writes the whole overview to one file. Every rate travels with its n, an unmeasured value is an empty cell rather than 0, and a block that was not read is written as a comment line with its state and reason, so a file opened months later cannot be mistaken for "nothing happened".

## More detail

Below the four questions, a closed **More detail** section keeps the engine split, the competitors and the markets tables.

## Related

* [Reading the numbers](/measure/reading-the-numbers): n, the dash, and when a change is real.
* [Before and after](/measure/before-and-after): the re-reads behind every verdict.
* [Who decides an answer](/measure/deciding-sources): the sources behind each loss, and the next move for each.


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