How Agencies Report AI Visibility to Clients

A client AI visibility report should answer three questions: how often do assistants name and cite the client for the questions that matter, how does that compare with competitors, and what changed since last time and why. Everything else is decoration. The hard part is not the layout. It is reporting numbers that are honest about sampling, so that next month's dip does not turn into an argument about something that was noise.

Key Takeaways

• Report a rate over a fixed question set, always with the number of answers behind it.
• Split by assistant. A blended number hides which one leaves the client out.
• Keep mentions, citations and recommendations as separate lines.
• Show competitors on the same questions, since a share only means something relative to others.
• Annotate platform changes and your own work on the timeline.
• Do not promise rankings, citations or revenue. Report evidence and next steps.

Agree the Panel Before You Report

The report is only as stable as the question set. Agree it with the client once, in writing.

• Unbranded buyer questions. Questions that do not name the client, such as "what tools do X for Y". Branded questions inflate scores because the client has named itself.
• A small set of branded questions to catch misstatements, such as pricing, safety or "is it legit". See AI brand misrepresentation.
• A fixed competitor list, picked from the brands the client's buyers actually consider.
• Assistants. ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews cover most of what clients ask about.

Changing the panel mid-quarter breaks comparability. If you must, start a new baseline and say so.

The Metrics That Hold Up

Presence rate. The share of answers that name the client across the panel, per assistant. This is reportable with its run count and margin. A standard 95 percent binomial margin shows the idea: 30 queries answered once on one engine give about plus or minus 18 points near a 50 percent rate, while 30 queries run daily for a week and pooled give about plus or minus 7. Those margins assume every answer is independent. Repeated runs of the same question are not, so the pooled week is less certain than plus or minus 7 suggests, and adding questions narrows the margin faster than adding runs. Evertune's worked example shows margins shrinking with repetition on real prompts: at five repetitions one brand's visibility score carried a 27-point margin, falling to about 6 points at 100.

Share of voice. The client's count divided by the client plus competitors. A ratio, so it moves when competitors move. AI share of voice explains the method.

Citation share for top domains. Which sites assistants cite for the panel. Usable above the margin for its run count, and useful for deciding where to pitch or publish. One run finds only a small share of the URLs an engine cites for a query, so pool runs (see how often AI answers change for the figure).

Leader on comparison questions. Conductor found the first-named brand stayed the same across most runs on comparison prompts (the figures are in how often AI answers change), which makes it a stable item to track.

The Metrics to Avoid or Caveat

• Position in a single answer. It is noise.
• Citation of one specific page for one question. Too few repeats.
• Modelled "AI search volume". Some data vendors derive it from Google keyword volume or from panels with unpublished methods, so do not present it as measured demand.
• A single sample. How often AI answers change shows why.

A One-Page Structure

1. Headline. Presence rate and share of voice this period against the last, with answer counts. One sentence on whether the change is larger than sampling noise.
2. By assistant. A small table. Name, cite, neither.
3. Against competitors. Same questions, same period.
4. Sources. The five domains assistants cited most for the panel, and the client's own share.
5. What we did. Fixes shipped, pages published, pitches sent, directory listings updated, each dated.
6. What changed outside our control. Platform and model changes in the period. Do not read a shift starting on one as your effect.
7. Referral traffic. Sessions from AI assistants as a floor, not a total. See track AI referral traffic in GA4.
8. Next period. Three actions with the reason for each.

Managing Expectations

Clients ask whether this is working. An honest answer needs a control. The one field study with untreated pages (Watanabe and Nakayashiki) found that ChatGPT referrals to the treated pages rose sharply, but untreated pages on the same site rose too, so most of the raw growth was the platform, and the net effect was suggestive rather than proven (the figures are in how long does AEO take). Use the same logic: compare changed pages with similar unchanged ones, and say what the comparison shows.

Avoid promising specific citations or a number of days. Where evidence for a tactic is weak, say so. How long does AEO take gives the timeline to share with a client at kickoff.

Setting It Up Without Building It

Running the panel by hand is possible and slow. Lectern runs a fixed panel on a schedule, keeps the history, and reports mentions and citations by assistant, with each client as its own brand. Lead is priced per brand at $990 a month, so a retainer with several clients maps to one brand each; the details, and the free Hobby plan, are on the pricing page. What is and is not available for agencies today is on the agencies page. Lectern can also be used from Claude, ChatGPT, Cursor or any MCP client, as described on the MCP page.