How to write an AI visibility report a client can act on

A client-ready AI visibility report answers five questions in a fixed order: what did we measure, what changed, what did the AI answers actually say, what did we do, and what happens next. The order matters more than the metrics. A client who can see the population behind the numbers, the evidence from real answers, and a dated list of next actions can decide something. A client who only sees a percentage cannot, especially in a week when the percentage did not move.

In short

• Open every report with the population measured (which prompts, which engines, how many answers), so a number can never be read without its denominator.
• Report changes next to the noise they might be, and say plainly when a change is too small to read.
• Put two or three real answer excerpts or citations in the report; a percentage without evidence gives the client nothing to check.
• End with dated next actions that each name an owner and what you will look at to judge them.
• A flat week is still a full report: the value is in the stable baseline, the work logged, and the next test.

What should a weekly AI visibility report contain?

Five sections: the population measured, the changes, the answer evidence, the work completed, and the next actions. Each stops a specific misreading.

• Population measured stops a client comparing two numbers built from different prompts or engines.
• Changes stops a small wobble being sold as a trend.
• Answer evidence stops the report being a black box; the client can read what an engine said.
• Work completed connects the number to effort, without claiming the effort caused it.
• Next actions turns the report into a decision instead of a status update.

Keep it to about one page. If a section has nothing to say, write one sentence saying so.

Which numbers belong in the report?

Report a small, stable set of readings, and define each one once. The weekly report in DiscoveredBy carries four: brand visibility, citation rate, citation share and average citation position (Weekly report). The headline figure on the report screen shows how it moved from the week before; for the other readings, compare against last week's report yourself. You do not need more headline numbers than that.

Define them the way the metrics reference does, because the definitions carry the population:

• Brand visibility is the share of analysed answers that name your brand.
• Citation rate is the share of collected answers that cite your domain.
• Citation share is the percentage of AI-cited sources that come from your domain across a set of tracked prompts (glossary). The docs do not spell out the weekly report's exact counting rule, so quote the report's figure and do not recompute it.
• Average citation position is where your cited pages sit among an answer's sources (the report screen labels its tile "average best position"). It is a citation position, not a brand ranking, and it reads blank for a week in which nothing of yours was cited.

Two details will save you an awkward call. Brand visibility and citation rate divide by different populations (analysed answers versus collected answers, and analysis lags collection), so print both denominators. And the platform treats fewer than 30 observations as provisional, so label any slice under that (one engine, one country) rather than reading a change into it.

For the mention versus citation distinction, see AI visibility, brand mentions, and citations: what each number tells you.

The report template

Copy this outline for each client and fill it in weekly. The bracketed prompts say what belongs in each slot.

AI VISIBILITY REPORT
Client: [name]            Week: [Monday date] to [Sunday date], UTC
Prepared by: [name]       Prior report: [date]

HEADLINE (two sentences, written last)
[What changed or did not, and the one thing we want you to decide.]

1. POPULATION MEASURED
- Prompts tracked: [n]. Added this week: [n]. Removed or paused: [n].
- Engines measured this week and last week: [list]. Any engine missing
  or newly added: [say so, and say the comparison uses shared engines only].
- Answers collected: [n]. Answers analysed for brand mentions: [n].
- Sample warnings: [any slice under 30 observations is provisional].
- Anything that makes this week not comparable to last week: [list, or "nothing"].

2. CHANGES
| Reading                  | This week | Last week | Change | Read it as        |
| ------------------------ | --------- | --------- | ------ | ----------------- |
| Brand visibility         |           |           |        | [flat/moved/unknown] |
| Citation rate            |           |           |        |                   |
| Citation share           |           |           |        |                   |
| Avg citation position    |           |           |        |                   |
- Prompts that gained or slipped: [names, with the direction]
- Competitors that moved: [names, only if the comparison is like for like]
- Unavailable readings: [name any reading with no data, and why]

3. ANSWER EVIDENCE
- Two or three saved answers or citations, each with: prompt, engine, date,
  what the answer said, and which sources it cited.
- What the evidence shows: [one sentence per example].
- What it does not show: [one sentence: examples, not a sample].

4. WORK COMPLETED
- Published: [articles, with URLs]
- Changed: [pages updated, what changed, the date it went live]
- Prompt set changes: [added, paused, and why]
- Last week's next actions: [done / not done / dropped, with reasons]

5. NEXT ACTIONS
| # | Action | Owner | Due | What we will look at to judge it |
| - | ------ | ----- | --- | -------------------------------- |

DECISIONS NEEDED FROM YOU
- [Only items that need the client. Otherwise write "none".]

Write the headline last. Many clients will read only that, so make it a plain sentence, not a metric.

How do you report a week when nothing moved?

Report the stability as a finding, then show where the useful information sits underneath it. A flat headline is not an empty week: the prompt-level movement, the evidence, the work logged and the next test all still exist.

Here is a checklist for a flat week.

• Say "flat" and say why you believe it. State the change and the sample size, and say the difference is too small to act on. Show the last few weeks for context if you have them. Describe the number; do not editorialise.
• Look under the average. A steady brand visibility figure can hide some prompts gaining and others slipping. The weekly report lists both, and prompts with no change at all are left out of those lists (Weekly report).
• Check the population before you check the result. If the prompt set, engines or analysed count differ from last week, the flat line might be an artifact. The weekly report compares only engines measured in both weeks, and a summary that reports a brand visibility change says so.
• Show the work that is too new to have shown up. Pages published this week have had almost no time to appear in answers. Log them under Work Completed with their go-live date and a date to look again.
• Give the client a decision, not a shrug. A flat week is a good time to propose a test or a page to fix.

Do not write "no change means the work isn't working". A flat week is one observation. Whether a specific edit helped needs its own method: Did your content update help? A practical before-and-after review.

Worked example: a flat week for Quillstone

Enlarge image

Quillstone sells document-review software to mid-sized legal and compliance teams. Its agency reports weekly. This is the report for the week of Monday 7 September to Sunday 13 September 2026.

AI VISIBILITY REPORT
Client: Quillstone         Week: 7 Sep to 13 Sep 2026 (UTC)
Prior report: 7 days ago

HEADLINE
Quillstone's visibility and citation numbers moved by 0.1 point or less this
week, too little to read as a change. Two pages went live on Thursday and are too
new to read; we will look again on 28 September. We need one decision from
you: approve a fix to the pricing page.

1. POPULATION MEASURED
- Prompts tracked: 40. Added: 0. Paused: 0.
- Engines measured both weeks: ChatGPT (app), Gemini (app), Perplexity,
  Google AI Overviews. Nothing missing or new.
- Answers collected: 300 this week, 296 last week.
- Answers analysed for brand mentions: 288 this week, 296 last week.
  The 12 not yet analysed were collected on Sunday and are waiting on
  mention extraction, so brand visibility uses 288.
- Sample warnings: none for the totals. The Google AI Overviews slice is
  provisional (fewer than 30 completed answers, because Google showed no AI
  answer on many runs).
- Comparability: same prompts and engines as last week.

2. CHANGES (table follows, then prompt-level detail)

The arithmetic: 92 / 288 = 31.9%, 94 / 296 = 31.8%, 45 / 300 = 15.0%, 44 / 296 = 14.9%. Citation share and average citation position are shown as the report's saved readings, since the counts behind them are not shown here.

Prompt-level detail for section 2
- Gained: 3 prompts, including "redlining tool for contract review".
- Slipped: 3 prompts, including "document review software pricing".
- The three gains and three slips roughly cancel, which is why the headline
  is flat. The pricing prompt is the one we are acting on (see 5).
- Competitors: Brieflane was named in more analysed answers than Quillstone
  this week and last; the gap did not change.

3. ANSWER EVIDENCE (saved answers from this week; examples, not a sample)
a) Prompt "document review software pricing", Perplexity. The answer
   describes Quillstone's entry pricing using a figure from a two-year-old
   third-party listing. Sources cited: two comparison sites, no Quillstone page.
b) Prompt "redlining tool for contract review", ChatGPT (app). Names
   Quillstone third of four, cites Quillstone's redlining guide.
c) Prompt "best document review software for compliance teams", Gemini (app).
   Names Brieflane and Clausewise, not Quillstone. Cites a review roundup.
What this shows: (a) an answer relying on outdated third-party pricing;
(b) a page of ours cited; (c) a comparison page we are not on.
What it does not show: three examples, chosen by us; not a ranking of
sources and not a reason the engines wrote what they did.

4. WORK COMPLETED
- Published (Thursday): "Document review checklist for compliance teams".
- Changed (Thursday): the redlining guide (added a comparison table).
- Prompt set changes: none.
- Last week's next actions: 1 of 3 done. Done: redlining guide update.
  Not done: pricing page fix (waiting on client approval). Dropped: LinkedIn
  post (deprioritised; not a cited source in any saved answer this week).

5. NEXT ACTIONS

DECISIONS NEEDED FROM YOU
- Approve the pricing copy correction (action 1).

Notice what the flat headline still delivers: three prompts up and three down, one answer repeating an outdated price, two pages too new to judge, and one approval that unblocks the next step. A report that led with "visibility +0.1 pt" would show none of that.

How does a report in the product map to this template?

The product covers the measurement parts of the template. The narrative is yours.

The shared client digest is narrower than the report on screen. It carries the project name and accent colour, the report dates and summary, four readings, recorded activity and up to three recommendation titles. It excludes raw answers, prompt lists, analytics revenue, team details and recommendation details, and the docs say historical deltas and engine comparisons are not included in it, so the Changes table is yours to add (Client report links). So treat it as the attachment that gives a client a fixed, dated copy of the numbers, and write your own narrative around it.

Before you send anything, know these facts from the docs:

• An owner or editor creates the link. It works without a DiscoveredBy account, expires after 1, 7 or 30 days, and can be revoked. Anyone the client forwards it to can also open it, so treat it as a credential. Creating a link does not send an email.
• The content is fixed at creation; regenerate or edit anything afterwards and you need a new link. Print / Save PDF uses the browser's print dialog; there is no server-generated PDF.
• The digest keeps a "Prepared with DiscoveredBy" footer. It is a client-branded digest, not a white-labelled application.
• Older reports can show a reading as Unavailable, and the saved summary can still describe a reading the digest withholds, so read it before sharing.

For agencies with several clients, the Client reports tab shows the same completed week across your active projects, one card per project, with separate readings. The docs are explicit that these are not an agency-wide average or a ranking: prompts, countries and engines differ by client. There is also a combined agency report that puts several clients into one shared link. Each client's numbers stay separate, but anyone holding the link can read every included client, so only use it when every recipient should see all of them (Client reports).

Enlarge image

Common mistakes and what this cannot tell you

Enlarge image

Reporting a percentage without its denominator. "Visibility is 31.9%" is unreadable until the client knows it is 92 of 288 analysed answers. Put the counts in.

Changing the prompt set and reporting the result as a trend. Adding or pausing prompts changes what is measured. If you must change the set, report the week as a new baseline.

Presenting citations or mentions as proof of why an engine said something. They are observations of an answer, not of what the model weighed. Write "the answer cited X", not "the engine ranked us lower because".

Claiming your work caused the change. Published pages, edits and fixes are things you did. The report can log them next to the readings, but a change alongside an edit is not evidence the edit caused it. The before-and-after review method exists for that question.

Treating three excerpts as a sample. The saved citation examples in a digest are limited to five, ordered by date and position, and are not a representative sample. Say so in the report, as the example does.

Skipping the traffic question. This report covers what appears in AI answers, not whether it produces visits. How to measure AI referral traffic and conversions with GA4 covers that.

Frequently asked questions

How long should a weekly AI visibility report be?

One page is enough: headline, readings table, two or three answer examples, work log and next actions. Put prompt-by-prompt detail in an appendix.

Should I include the raw AI answers?

Include two or three chosen excerpts as evidence, and say they are examples. The docs describe the shared digest as excluding raw answers, so if you want a client to read one, you have to add it yourself in your own report. Check any excerpt for wording you would not want circulated before you send it.

What if the numbers drop?

Use the same template. The population section is your first check: did the prompts, engines or analysed count change? Then look at which prompts slipped and read the saved answers for them. A drop is a prompt to investigate, not a conclusion; Why AI recommends your competitor: how to investigate the answer gives a method.

How much variation is normal from week to week?

The docs give no threshold, so do not invent one. They do flag slices under 30 observations as provisional, and note that engines and analysed counts can change between weeks. Judge a change against your own recent weeks: if readings have been moving by similar amounts in both directions, a small move is probably ordinary variation.

Next step

Once the template is settled, the weekly work is mostly reading numbers you did not have to assemble. DiscoveredBy builds a dated weekly report for each project, lets owners and editors share a fixed client digest, and shows the same completed week across your projects in the Client reports tab. Depending on your plan, weekly-report access may be included; see plans and limits and alerts and reports for what applies.

Sign in or create an account and open the Client reports tab to see last week's numbers next to this template.