AI visibility for agencies: reporting AI answers to clients

AI search visibility for agencies is best reported as evidence: the buyer questions tested, the answers returned, the pages cited, and the work that follows. It is not a permanent rank.

This guide shows how to structure the report, separate software cost from labour, work through one client example, and explain the limits of sampled AI search results. It also includes a short process you can reuse each month.

How do you turn AI visibility into a client report?

Start with buyer questions, not a generic visibility score. A client needs to see the questions that matter to its market, the engines tested, whether the brand appeared, and which sources the answer cited. The report should connect each finding to a page, so the account team can assign work instead of admiring a percentage.

A simple report lists the question set and date of the run, shows mentions and citations by engine, names the cited pages and domains, and turns gaps into actions. Those actions might include improving a comparison page, clarifying a product fact, or fixing access to a page. Keep the raw answer available as evidence, because a client will reasonably ask what the assistant actually said.

The engine differences matter. In CiteGraph's sample, Perplexity cited an average of 5.5 sources per answer and named the scanned product in 10.5% of answers. Gemini cited 4.2 sources and named the product in 9.4%, while ChatGPT cited 3.5 sources and named it in 6.1%. Claude cited 4.9 sources and named the product in 10.7%. A single blended score hides that spread.

Report the engine view beside the combined view. The combined view is useful for a weekly direction of travel, but the engine view tells you where a page change may have an effect. You can compare this with the weekly research on who owns AI answers, which gives broader context without treating a category average as a client forecast.

What should you put in an agency AI visibility report?

Give each client a short front page and a detailed evidence section. The front page should say how many questions were tested, how many runs were made, which engines were included, and the observed mention and citation results. It should also include a plain recommendation, such as “rewrite the comparison page for these buyer questions”, rather than “improve AI visibility”.

The evidence section should show the answer, the cited URL, the source domain, and the reason that page matters. Separate a brand mention from a citation. A product can be named without its own page being cited, or have a cited page without being named. Those are different jobs and should create different tasks.

Use the free AI visibility checker for a quick client conversation, then use a fuller scan when the client needs repeatable evidence across questions and engines. The free AI readiness check is useful before reporting visibility because it runs fifteen checks and shows the fix for each. Readiness is a diagnosis, not proof that an assistant will cite the site.

Keep the client language concrete. “Your site was named in this sample” is defensible. “Your site ranks fourth in AI” is not, because the data contains no fixed ranking position and the answers do not create one stable results page.

How much does reporting cost you per client?

Price the work from two separate costs: the software allocation and the human time. CiteGraph Starter costs $29 a month, Growth costs $69, and Scale costs $149, according to the CiteGraph pricing page. The plans cover different project counts, question runs, tracked rivals, and written pages, so the monthly price alone does not show the delivery workload.

See the worked example below for the per-client software calculation. The important point is that this allocation is only one part of delivery. Add the time to choose questions, review answers, write the explanation, attend the client call, and turn findings into tickets.

This distinction stops a common pricing mistake. A report can have a low software allocation and still lose money if review and explanation are included in a low fixed fee. Put software, labour, and reporting frequency in your internal margin model. Do not treat the subscription price as the cost of delivery.

CiteGraph plans start with a 7-day trial opened with a $1 card confirmation and include 2 scans. A single scan without a subscription is $12. Those options let you test the workflow before committing, but they do not tell you the labour cost, which remains agency specific.

How do you work through one client report with real numbers?

Take a three-project agency using Growth. The plan costs $69 a month, includes 3 projects, 10 questions each, 5 runs per question on 4 engines, 10 tracked rivals, and 60 written pages a month. Allocate one project to each client. The software allocation is $69 divided by 3, or $23 per client per month before labour.

Now build the report for one of those clients. Select up to the plan's 10 questions for that project. Run each question 5 times on each of the 4 engines, then record whether the client was named and whether one of its pages was cited. Keep those outcomes in separate columns. The scan gives you repeated observations rather than one answer copied into a slide.

Next, show the client what the wider sample can and cannot support. CiteGraph's research records 9,791 AI answers and 64,667 citations across 77 scans of 44 sites in 48 categories. Those scans asked 507 distinct buyer questions, and the engines cited 9,192 distinct domains and 21,465 distinct pages. That is useful context for patterns, but it is not a forecast of this client's next answer.

A line in the report might read: “In this client's tested sample, the brand was named in the answers recorded for these questions, and these pages supplied the citations.” Follow it with the question list, engine split, answer excerpts, and page URLs. Then attach one action to each gap. The report has done its job when the client can choose the next page or technical fix without asking what the percentage means.

The engine comparison is worth showing because naming rates differ in the research sample. Perplexity and Claude sit above ChatGPT in the recorded naming rates, while Gemini sits between them, but those are sample results, not positions a client can buy. The methodology page explains how the measurements are made. Use the spread to decide which engine's cited pages deserve the first review, rather than treating all engine results as interchangeable.

How do you explain uncertainty when the client wants a rank?

Say this plainly: “We measured a sample of questions and runs, not every answer an assistant could produce.” Then show the size and shape of that sample. A result based on a small question set can move when the wording changes, when the engine changes, or when the cited page changes.

We don't have a margin of error to quote here, so we won't invent one. State the sample size and the engines tested next to the result, and leave the rest out rather than estimate a margin of error you don't have. We measure samples; we do not know what we have not measured.

Avoid translating an observed mention rate into a rank. The research records whether a product was named and which sources were cited. It does not provide a universal position such as first, fourth, or tenth. If a client asks for a rank, explain that the more honest unit is inclusion in a defined answer sample, supported by the page and engine evidence.

You could label findings “strong enough to act” when the same page or issue appears across repeated runs and more than one engine. “Needs more observation” could describe a result based on a narrow question set or one that changes between runs. Define the label once, in writing, so two account managers don't apply it differently on different calls.

Which AI visibility tools should you use for repeatable work?

Choose a plan by the number of live client projects you run each month, not by the headline price. Then check whether the plan gives you enough repeated runs, tracked rivals, and written output for the reports you actually deliver.

The important comparison is whether a tool lets you preserve the question set, repeat the measurement, inspect citations, and turn findings into client work. We'd avoid any report format that hides the answer and source page behind a single score, because that's what creates re-explanation work later.

If implementation matters, CiteGraph runs as an MCP server. Claude, Cursor, or VS Code can read a scan and apply readiness fixes in a repository, according to the MCP documentation. That is useful when the agency owns the technical work, but it does not replace editorial judgement about which buyer question deserves a page.

Keep the tool choice tied to client volume and reporting depth. Your delivery process determines whether the plan becomes useful evidence or another unused allowance.

How can you build the report in a short working session?

Use this procedure for a first pass:

1. Pick the client's buyer questions and assign them to a project.
2. Run the questions across the available engines and save the answers.
3. Mark brand mentions and citations separately.
4. Group cited pages by the action they suggest, such as comparison content, product facts, or access fixes.
5. Write one finding and one next action for each important gap.
6. Add the sample description, engine split, and limitation statement before sending the report.

The procedure is deliberately plain. Do not spend the first session polishing a scorecard while the answer excerpts and source pages remain hidden. A client can act on a page, a question, and a clear next step. A client cannot act on “AI presence” without the evidence underneath it.

For a recurring service, save the question set and keep the report structure stable. Change the content when the evidence changes, not because the template needs a new colour. The useful trend is whether the client appears more often in the defined sample, whether its pages earn citations, and whether the same gaps keep returning.

The next question to ask is: which buyer questions and source pages should this client fix first? The answer lives in the cited answer evidence, supported by the weekly research and the measurement details in the methodology guide, not in a universal AI rank.