Skip to main content
Your agent decides for itself how to present an answer. Most of the time that decision is good, but it can only work with what your question implies. If you know what you want to see, saying so in the prompt is the single most reliable way to get it. This page lists every visualisation an agent can produce, when each one is the right choice, and the exact phrasing that steers the agent towards it.

​
How a visualisation gets made

1

The agent queries your data

It writes SQL shaped for the visualisation — one column for the axis, one column per series.
2

It builds the chart, table or KPI row

The full result set is used, not just the rows you see summarised in chat.
3

It renders in chat, or gets embedded in a report

Charts appear inline in chat. Tables and reports open in the right-side panel.
An agent can build up to 5 charts in one step, so asking for several views of the same data in one prompt is efficient — you don’t need a separate message per chart.

​
Choosing a visualisation

The single most common mistake is asking for a chart of categories over a date-style axis. If your X values are names — campaigns, channels, products, markets — ask for a ranked bar. Names stay readable next to each bar instead of being squeezed and rotated underneath.

​
Chart types

​
Time series

The default. A metric plotted over dates, or over any ordered category axis.
Time series chart with revenue as bars on a left euro axis and orders as a line on a right count axis

Bars and a line on two axes — revenue in euros on the left, order count on the right.

What you can steer:
Use a second axis when the two metrics have genuinely different units or magnitudes (revenue in euros vs. a count of transactions). Without it, the smaller series flattens into the baseline and tells you nothing.Say which metric belongs on which side, and in which unit. The agent labels and formats each axis separately, so the right axis won’t inherit a euro sign it shouldn’t have.
Both are time-series only.
  • Reference lines mark a threshold on either axis — a budget, a target CPA, a launch date.
  • Highlighted periods shade a date range with an optional label — a campaign flight, a site outage, a promotional week.
Just describe them: “mark our €2.50 target CPA”, “shade the two weeks the site was down”.
Line chart of weekly revenue with a dashed target reference line and a shaded period labelled Summer sale

A single line with a gradient fill, a dashed target line, and a shaded, labelled period.

Keep reference line labels short — they’re drawn to the right of the line and a long label runs out of room.

​
Stacked and 100% stacked bars

Both stack the series on top of each other; they answer different questions. The same data, both ways — note how the growth in total sessions is visible on the left and deliberately absent on the right:
Stacked bar chart of monthly sessions by channel, showing total volume growing

Stacked: total volume plus composition.

100% stacked bar chart of the same sessions data, each bar normalised to 100% with percentage labels in each segment

100% stacked: mix only, with auto data labels.

Use 100% stacked when the question is genuinely about mix — a channel split, a device split, new vs. returning. Volume changes disappear on that axis, which is the point, but it also means it’s the wrong chart if you care how big the total got. Either way, hovering a segment shows its share of that bar’s total.
A 100% stacked chart needs at least two stacked bar series and can’t be combined with a right-hand axis or an absolute value reference line — a fixed number has no meaning on a 0–100% axis. If you ask for a percentage view of a single metric, expect the agent to give you an absolute chart instead and say why.

​
Ranked bar

Horizontal bars, ranked by value. The right choice for any “top N” or “who’s biggest” question.
  • Category names sit next to each bar, so long campaign names stay readable.
  • Sorted largest-first by default; ask for “ascending” or “smallest first” to flip it.
  • Ask for a specific cut-off: “top 10”, “top 15 markets”. Around 10 reads best; 50 is the maximum.
  • Bars can be stacked too — “top 10 campaigns by revenue, split by device”.
“Top 10 campaigns by revenue last month, with the revenue shown on each bar”
Horizontal ranked bar chart of the top 10 campaigns by revenue, sorted largest first with currency labels on each bar

Long campaign names stay horizontal and readable; values are labelled at the end of each bar.

​
Scatter

Two numeric metrics against each other — the chart for “is there a relationship here?”
“Scatter plot of spend vs. ROAS per campaign, coloured by channel, sized by conversions, with a trendline”
Scatter chart of campaign spend against ROAS, with bubbles sized by conversions, coloured per channel, and a dashed linear trendline labelled R squared 0.12

Bubble size is conversions, colour is channel, and the dashed trendline carries its R² — here 0.12, i.e. spend explains very little of the variation in ROAS.

A scatter needs at least two rows where both metrics are numeric and non-empty. Rows with a NULL on either axis are dropped, so a sparse metric can produce a much emptier chart than you’d expect from the row count.

​
Making the numbers readable

Number formatting applies to axis ticks, tooltips and on-chart labels. It’s worth being explicit about it, because a raw axis of 1234567.891 undoes an otherwise good chart. Numbers are formatted in Dutch notation (nl-NL) and default to euros. Ask for another currency by name if you need one. You can also set precision — “two decimals”, “no decimals”.

​
Data labels

By default charts have no labels on the plot; the tooltip carries the detail. Ask for labels when the chart will be read on paper or in a slide where nobody can hover:
“Show the values on the bars”
Labels are placed only where they measurably fit — labels inside stacked segments too thin to read are dropped automatically, and dense line charts stay clean. This means it’s safe to ask for labels on a stacked chart. If you need every label regardless of overlap, say so, and keep the number of data points small.

​
Tables

Ask for a table when the exact numbers matter, or when you want more columns than a chart can carry. Tables open in the right-side panel and can be sorted. Columns can be styled by type, which you can request directly:
“Table of the top 20 products: product name as the title, revenue as currency, growth as a change column, and country as a badge”

​
KPI tiles in reports

Reports can open with a row of KPI tiles — a scorecard of the headline numbers, each with an optional change vs. a comparison period. Each tile carries:
  • A label — the metric name
  • A value — the formatted headline number
  • An optional delta — the change, its comparison period (“WoW”, “MTD”, “vs target”), and whether that movement is good or bad
That last part matters and is worth stating in your prompt. Direction and sentiment are separate: cost per conversion going up is bad, conversion rate going up is good. Tell the agent which of your metrics are inverted and the colours will match:
“Open the report with a KPI row for revenue, sessions, conversion rate and CPA, each with the week-over-week change. Remember CPA is a cost metric — rising CPA is bad.”
Report KPI row with six stat tiles showing revenue, orders, conversion rate, average order value, cost per acquisition and ROAS, each with a coloured week-over-week delta

Six tiles in a 3-column grid. Cost per acquisition is up 11,2% and renders red with an up arrow, while revenue up 5,1% renders green — same direction, opposite sentiment.

A KPI row holds up to 9 tiles, laid out three per row. Ask for more than that and the agent will split them across separate rows or sections — which usually reads better anyway. Group them deliberately: one row for outcome metrics, one for efficiency metrics.
KPI tiles replace the old habit of stuffing metrics into a highlight box or a pipe-joined sentence. Use highlights for what they’re good at — a single narrative callout, one key finding, a data quality flag.

​
Structuring a report

Reports are built from blocks: text, lists, highlights, KPI rows, charts and tables. The agent will lay one out for you, but a prompt that specifies the structure gets a far more consistent result — especially for a report that runs on a schedule and needs to look the same every week.
“Create a weekly performance report:
  1. Executive summary — a short paragraph, then a KPI row with revenue, orders, AOV and ROAS, each with the week-over-week change.
  2. Trend — revenue and orders per day, revenue on the left axis in euros, orders on the right as a count.
  3. Channel mix — sessions per week as a 100% stacked bar by channel, with the percentages shown on the chart.
  4. Top campaigns — ranked bar of the top 10 campaigns by revenue, plus a table with spend, CPA and ROAS.
  5. Recommendations — a numbered list of at most five actions.”
Reports built this way are the ideal thing to attach to a scheduled job — write the layout once and a consistent, PDF-exportable report lands every week.

​
Limits worth knowing

​
When you don’t get the chart you asked for

You got a time series with a category axis. Ask for a ranked bar: “show that as a ranked bar chart instead”.
The two metrics are on incompatible scales. Ask for a second axis, or split them into two charts.
Say what the unit is: “revenue is in euros — format the axis as currency and abbreviate the large numbers”.
They should. For a 100% stacked chart, the agent reads the rendered percentages directly rather than recalculating them, so the narrative and the chart agree. If you see a genuine mismatch, flag the answer as incorrect so an admin can review it.
Rows shown inline in chat are a preview — charts and tables are always built from the complete result set, whatever the row count. If a limitation genuinely applies to your answer, the agent will state it explicitly.
When a correction is about your data rather than your question — “the revenue column is in cents”, “CPA is a cost metric” — add it to the agent’s context so you don’t have to repeat it in every chat.