Overview
The page you land on, panel by panel: the filter bar, the visibility trend, rankings, the priced market, top cited domains and the latest chats.
The Overview is one screen that answers “where do we stand”, scoped by the filter bar at the top. Everything on it reads from the same scope, which is why the filters live once at the top rather than being repeated per panel.
The filter bar
Four scopes, and each one changes the denominator of every figure below it:
- Period — how far back the chats are drawn from.
- Model — one model, or all of them. Models disagree; the all-models figure is an average over that disagreement, so it is worth checking a metric per model before concluding anything about it.
- Locations — one market’s answers, or all of them. The same question asked from two countries is two different answers, not one answer measured twice.
- Topic — one topic’s prompts, or all of them.
A filter appears only when there is more than one thing to filter by. A project measured in one market draws no locations filter, because a menu offering “All locations” and “US” is two choices that select the same rows.
Visibility
The headline is visibility over the whole window, with its sample size and confidence interval beside it — those two are what make the percentage a reading rather than a number.
The chart carries two series. The filled one is you; the plain line is the
category leader on that day, which is not necessarily the same brand each
day. D / W / M change the granularity.
Below the chart, the panel says how many chats it is showing and how many failed. A failed chat is a model that did not reply — usually a timeout, and retried on the next pass. It is shown rather than hidden because a figure measured over most of the chats that were asked is a slightly different claim from one where everything came back.
It also says what was left out. Answers belonging to a paused prompt are not counted, and the caption names how many — so a figure that moved the morning somebody paused three prompts has its explanation on the same line rather than looking like a change in the market.
Rankings
Every tracked brand, yours marked, with all four metrics side by side. This is the panel to read when a single figure looks good — the interesting information is usually in the disagreement between columns. See Metrics overview for the patterns worth recognising.
Market
What the questions above are drawn from, priced — and then narrowed twice. Three rows, each one a part of the row above it:
- Category value — what advertisers pay per month for these intents, built from the keywords this brand and its rivals are contested on. The full width of the panel, because it is the whole.
- Tracked by your prompts — how much of that value a topic you ask questions about reaches, in dollars and as a share of the category. What is left is named beneath it — untracked demand, and the part Keywords gives a button for.
- Named in answers — the captured share, drawn inside the tracked part rather than against the whole, because a share of money nobody asks about would not mean anything.
The last row is drawn only when it is a reading. Before any answers have landed against a priced topic, or when the topics carrying it are too thin, the row says so in words instead of printing a figure — “51 chats — too few to read” rather than a number somebody would quote.
The panel does not render at all until a pool has been built, which happens on the schedule.
The money and the percentage are never multiplied
The panel says what share of a priced market the answers reach. It never prints a figure for “value captured”, because being named in an answer is not a click. See how a keyword is priced.
Top domains
The sites the chats actually leaned on, ranked. Each row carries:
- Retrieved — the share of chats that reached for it at least once, cited or not.
- Citation rate — citations ÷ the chats that retrieved it: how heavily a chat that opened it ends up quoting it.
- Type — what kind of site it is.
The panel is ranked by Retrieved, and it is the same query the
Sources page runs — rows here and rows there are the same
domains with the same numbers, grouped the same way.
The Type column is the one to scan. A category whose chats rest mostly on
Reference and Editorial sources behaves differently from one resting on
UGC: the first is won with documentation and coverage, the second in
communities.
Rows open into the source’s own page. Full detail is on Sources.
Domain types
The same citations, aggregated by kind — what proportion of the evidence behind your category’s chats is corporate, editorial, reference, UGC, owned or competitor.
This is a shape-of-the-category panel rather than a scoreboard. It tells you where effort pays before you have spent any.
Latest chats
The last panel, spanning the page: the answers the panels above are summaries of. Every figure on this screen is an aggregate, and an aggregate is worth whatever a reader believes about the evidence under it — so the evidence is on the same screen rather than a click away.
It has its own filters, because it is a list rather than a figure: search the prompt text, and narrow by brand, by source domain or by model. Each of those is built from what has actually been sampled, so a filter with one real choice behind it is not drawn.
Open a row for the whole answer, with the brands marked and the sources it used beside it. The full archive, paged and exportable, is on Chats.