How to build an AI visibility report for your boss or client

What a report must contain

Seven things, in this order: the prompt set you ran, Visibility with its trend and sample size, Average position, Share of voice against named competitors, what changed and the likely reason, the top sources behind the answers, and one recommended action. Everything else is optional.

• The prompt set. The actual questions, how many there are, and whether they changed since last month. Without this, the trend is measuring your edits to the list.
• Visibility, with the trend and the sample size. The share of answers that mention your brand, shown as a line over the period with “based on N answers” underneath.
• Average position. How high you appear in the lists AI gives, on average, where 1 is the top. It explains a Visibility number that looks fine but feels wrong.
• Share of voice against named competitors. Your mentions as a share of all tracked brands in the same answers, with the competitor names visible rather than an anonymous “others” line.
• What changed, and the likely reason. Two or three prompts that moved, a quoted answer for each, and your best explanation with its confidence stated.
• Top sources. The domains feeding the answers, and which of them feed answers about you rather than only about your competitors.
• One recommended action. A single page to write or fix, with the reason it was chosen. A list of ten actions is a way of shipping none.

A copy-paste monthly structure

Seven sections in a fixed order, with placeholders you refill at the end of each month. Keeping the order and the scope identical is what makes this month comparable with last month.

1. Scope. [N] prompts × [M] engines, daily runs, [date range], one market, English (US). List any prompt added or removed, and why.
2. Headline. Visibility [X]% (based on [N] answers), change since last month. Average position [X] (based on [N] answers). One sentence on direction.
3. Competitive. Share of voice for you and each named competitor on the same prompts, same date range, same sample size printed once for the chart.
4. What changed. The prompts that moved, one quoted answer each with the date it was collected, and the likely reason. Mark anything you are inferring as an inference.
5. Sources. Top cited domains overall, then the domains that appear in answers about you. One line on the gap.
6. Caveats. Scores hidden for small samples, engines that publish no source list this period, figures marked est., prompts whose wording changed.
7. Action. The one page to write or fix, who owns it, and what number we expect to move.

Put the detail behind the numbers in an appendix rather than the body: the exported rows, the per-engine split, the answers you quoted. The people who want it will ask, and the people who do not will not scroll.

The five questions you will actually get

They ask about certainty, the direction, the drop, the competitor, and what we are doing. Getting the first one right is what keeps the other four from turning into an argument.

1. Is this real or is it noise? Answer with the sample size and the range: the number is based on [N] answers, and this is the range the true value most likely falls in. If the sample is too small, say the score stays hidden until there are enough answers to trust it.
2. Why did Visibility drop? Name the prompts that moved and read the answers behind them. A fall concentrated in two prompts is a prompt-level event, not the channel collapsing.
3. What is [competitor] doing that we are not? Compare Share of voice on the same prompts, then the domains cited in answers where they appear and you do not. That difference usually points at specific pages rather than a strategy.
4. Which of these numbers can go in the board deck? Any number with its sample size and date range attached. Percentages without their N and single-day screenshots are the ones that get repeated out of context six months later.
5. What are we doing about it? One action, chosen because it sits in the sources that already feed answers about you, with the metric we expect to move and when we will check.

What never to ship

Bare percentages without their sample size, unlabelled estimates, and single-day screenshots presented as trends. Each of these survives the meeting and embarrasses you two months later.

• A bare percentage. “Visibility 42%” is not a finding. The count and the date range are part of the number.
• An unlabelled estimate. Volume, authority and impact figures are relative, so they carry the est. mark and the word estimate next to them. Presenting them as counts invites a correction you cannot defend.
• A single screenshot as a trend. One day of answers is a mood, and engines move day to day. Trends need the period on the axis.
• A cross-engine average with no split. Engines behave differently, including in whether they publish source lists. Show the split or report one engine at a time.

Where each section comes from

The overview carries Visibility, Average position, Sentiment and Share of voice, each printed with “based on N answers”, which fills the headline and competitive sections in one screenshot. The Sources view and its CSV export fill the sources section, and the answer archive holds the quotes for what changed.

Practically: run the same prompt set all month, take the overview the morning the report is due, export the sources to CSV so the domain table goes in from a file instead of being retyped, and pull each quoted answer from the archive with its date. If runs were paused during the month, Run now collects a fresh set before you write, and the scope section should say so.

For the reasoning behind the sample-size rule see how many answers before you trust a number, and for the sources section see how to read AI citation and source data. While billing is off the workspace is free with unlimited team members, so the report can be a view people open themselves rather than a deck you rebuild every month.