Reading your first answer
What every part of an assistant answer means: the figure in the sentence, the rows in the panel, the basis badges, and the footer that says what was counted.
An answer has four parts and each one is making a different kind of claim. Once you can read them apart, you can tell in a second whether a number is safe to act on.
What you'll learn
- The four parts of an answer and what each is for
- What counted, projected and proxy mean on a row
- What "good" looks like at each stage of a workspace's life
The four parts
- The sentence. The finding, in prose, with the figures in it. Every figure here has rows behind it in the panel. If it could not be evidenced, it would not be stated.
- The panel. The evidence: the rows the sentence was computed from. Not an illustration of the answer.
- The trace. Which phases ran and what each one counted, for example "Checked 6 platforms, 1,203 answers".
- The footer. The window the figures cover, and when the underlying checks were made.
Read them in that order the first time, and in reverse order when a number surprises you.
Basis: counted, projected, proxy
Rows carry a basis, because not all evidence is the same kind of thing.
| Basis | What it means | How to treat it |
|---|---|---|
| Counted | Rows the platform actually stored | Quote it |
| Projected | A modelled figure, for example a lift estimate | Use the ordering, not the size |
| Proxy | Attributed indirectly, for example by time window | Indicative only |
A projected figure carries no measured error band today, which the panel says out loud. A proxy row says what the proxy is, for example "these were attributed by creation time rather than by a link".
The windows are not all the same
The footer exists because different figures are counted over different periods. Verbatim evidence runs over a rolling 90 day window. Your own current rate is a shorter window. The Industry Pulse benchmark carries a period of its own. A figure never wears another figure's window, so compare like with like by reading the footer.
Empty is a real answer
A row that says excluded, not measured, or not measurable is not a bug and not a zero. It means nobody measured the thing, and the platform will not invent a number to fill the space. See when a number is missing.
What good looks like
| Stage | What to expect |
|---|---|
| First week | Partial coverage, some surfaces excluded, a handful of sentences. The job is a baseline you believe, not a good number |
| First month | Every covered surface reporting, a rival list that has grown by itself, one change shipped through the ledger |
| Ongoing | The spread between surfaces narrowing, contested prompts converting, changes reaching Confirmed moved |
Do not judge week one on rate. Judge it on whether the prompts are right, because everything after that depends on it.
Try this in Outercite
Ask the assistant "What is my citation rate on each engine?", then read the footer before the number. Knowing the window changes how much weight the number deserves.
Related
How answers are grounded
The phases behind the sentence, and the share view that shows how a figure was made.
Data definitions
Exactly what each metric counts, over which window.
Panes and the ladder
Descending from the panel to the proof.
Last updated on