How Agencies Report AI Visibility to Clients (With a Real Report)

Not more data — an answer. "Is this AI thing working, is it worth what we pay you, and what are you doing next month?" A client report earns its keep when those three questions are answerable from its pages alone. That framing rules out the two default failure modes. The first is the dashboard screenshot dump — forty charts, no argument. The second is the vanity single number, an "AI score" that went up. The first is unreadable; the second is indefensible the moment a client asks what the number means. The report that works sits between them: few numbers, each one defined, each one carrying its sample size, arranged to support a narrative a human wrote.

After a year of sending these, our set is: visibility (share of scan runs naming the brand, with its ± margin), share of voice against named competitors, average position when listed, mentions by engine, and — the one that changes the meeting — the tie to outcomes: sessions, conversions and revenue that AI assistants actually referred, from the client's own GA4. Two disciplines make these survive scrutiny. Every rate carries its n: "35.3% visibility on 2,940 runs" is a claim, "35.3%" alone is a decoration. And null is null: an engine with too few decisive runs shows a dash, not a zero — a rule that has saved us more client-meeting credibility than any chart. If a competitor's sentiment score is built on three mentions, the report says so or doesn't show it. What to leave out: raw mention counts without denominators, blended cross-engine scores, and anything measured once. We wrote up the full metric definitions — and the five metrics that lie — in the metrics guide; the client report is that guide's short, opinionated edit.

The report in these images is our own agency's August report, exactly as generated. The cover carries the agency's branding, the client's name, and the period. The KPI row leads with visibility at 35.34% ± 3.0 — and its month-over-month change annotated "within margin", the report literally refusing to sell noise as progress. That one label does more for client trust than any chart: it teaches the client that when the report does claim a move, the move is real. Then the trend: visibility and share of voice over the month, with a timeline annotation on 18 August — "Website relaunch" — pinned to the chart, so the story explains itself at the glance ("the dip is the relaunch, and it recovered"). Annotations are the difference between a chart and an explanation; every meaningful account event should be on the chart that felt it. Then mentions by engine — a donut across all ten engines, which quietly answers "where does this number even come from" — and the competitor table: visibility, position, sentiment and share of voice for every tracked competitor, same period, same methodology. Clients rarely remember their own number; they always remember being ahead of the rival.

The section that converts skeptical clients is never the visibility chart — it's the one where AI visibility sits next to GA4: the sessions, conversions and revenue whose source was a known AI assistant. Stated honestly, that number is a lower bound (analytics attributes some AI referrals to direct), and the report should say so in exactly those words. Understatement is a feature here: when the lower bound already justifies the retainer line, the argument is over. We've written about why this pairing matters beyond reporting — a brand can gain AI visibility while its site loses sessions, and only seeing both trends on one page catches it. For the client report, the practical point is simpler: the AI section belongs in the same document as Google Ads and GA4, not in a second tool with a second login. One report, one period, one answer to "is it working".

AI referral traffic is the visits that arrive through a link inside an AI answer. GA4 records them in two ways. When the assistant passes a referrer, the session source is its host: chatgpt.com, perplexity.ai, gemini.google.com, claude.ai. GA4 also files some of these visits under its own AI Assistant channel. When the assistant tags its links, the tag wins, and with no medium on the tag the visit usually lands in Unassigned. Links ChatGPT cites often carry utm_source=openai, and Gemini has been seen adding utm_source=gemini. A filter that only knows host names misses those visits. So the count needs one definition that covers both. A session counts as AI-referred when its source is a known assistant host or tag, or when GA4 put it in the AI Assistant channel. A session that meets both counts once. In the report that becomes three numbers a month: AI-referred sessions, the conversions they produced, and revenue where the client tracks it. Split them by assistant. "Most of it came from ChatGPT" is a sentence a client remembers. Then say what the number cannot see. Some clicks from the apps arrive with no referrer and no tag, and GA4 files them under Direct. Clicks from Google's AI Overviews and AI Mode arrive as ordinary Google organic. Search Console's Generative AI report shows impressions there, not visits. So call the figure a lower bound, in those words. If the lower bound already covers the fee, nobody asks for the upper one.

The monthly report is written for the marketing lead. The board sees something else: one page, often a slide someone built from yours. Plan for it, or someone will crop your chart and the margin goes with it. The board page answers the same three questions in fewer words. Is it working: one visibility figure with its range, its run count and the month it is compared with. Is it worth the money: AI-referred conversions or revenue, labelled as a lower bound. What happens next: one line. Boards add a fourth question, whether the company is ahead. So put the client next to the two or three rivals the board knows by name. Keep the method off the page, one click away. Nobody on a board wants a margin of error explained. They do want to know the number wasn't picked to flatter. A footnote says how many answers it rests on and which engines. A change inside the margin is still written as unchanged. That rule matters more upstairs, because a board remembers last quarter's claim. In CrunchJunkie this is a second report on the same client: executive summary, KPI row, competitor table. It goes out on the same schedule as the monthly one, as a PDF or a share link.

The mechanics should be boring. In CrunchJunkie the monthly report is drafted by describing it: "Monthly client report: AI visibility plus Google Ads plus GA4 — in German", and the draft assembles from the client's connected sources, fully editable. Scheduling sends it monthly without anyone remembering to; the client gets a live share link — same report, live data, no attachment archaeology — or a PDF when procurement insists. Report language matters more than agencies expect: a report a client reads in their own language (ours generate in English, German, Spanish and French) beats a prettier one in yours. And the AI-written executive summary is drafted from the saved numbers — never invented client-side — then edited by a human, because the narrative is the part the client actually reads.

A monthly report tells the client what happened. It shouldn't be the first time the agency hears about it. If visibility fell in week one, the client shouldn't read about it in week five. The useful triggers are few. Visibility falling against its own recent past. A competitor moving ahead in the answers. A page the engines used to cite going down, or dropping out of their answers. AI-referred conversions falling in the same weeks as visibility, which is the one worth a phone call. Each trigger needs a bar for when it counts. Ours compares the last 14 days with the 28 before. It raises a signal only at 95% confidence or more, and only for a fall big enough to matter. Without a bar like that, alerts fire on noise every week and get muted by month two. What happens next is the playbook. In CrunchJunkie a signal can send an email, post to Slack or call a webhook. It can also draft an action plan, a report or a content brief for a person to approve. Nothing the AI drafts goes out without that approval. Then put the event and what you did about it on the next report, as an annotation on the chart. That page is the one that justifies the retainer.

The last mile is ownership. White-label — your logo, not your vendor's — is available from our Starter plan; from Pro the report lives on your own domain, so the client's bookmark says your brand. That's not vanity: an agency's AI-visibility service is worth a monthly line item precisely because the client experiences it as the agency's capability, not a reseller arrangement. We wrote separately about productizing white-label AI-visibility reporting — pricing structures, per-brand economics, the pitch. The report in this article is the deliverable that makes that line item renew itself.