AI visibility report template with the numbers that matter
A practical AI visibility report template built around lost answers, cited sources, actions and the limits of sampled data.
By CiteGraph · 21 Sept 2026 · every number below has a source
An AI visibility report should show the answers you lost, the pages AI assistants cited instead, the work that follows, and the margin around the result. The figures are sample evidence, so the report must show its questions, engines and denominators.
The rest of this post gives you the dossier structure. It covers what to copy, which numbers belong in the report, and where the data stops being certain.
Lost answers
Start with the answer, not the score.
For each buyer question, paste the relevant answer as it appeared. Keep the wording intact. Record the engine, the question, whether the product was named, and whether the answer included a citation to your site.
Do not reduce a useful failure to “visibility: zero”. A client needs to see what the buyer saw.
A lost answer usually falls into a few practical groups. The product is absent, or a rival is named instead. The answer names no product but cites other sources, or the product appears while the useful page on your site does not.
Those are different jobs. A missing product may call for clearer product facts or more third party coverage. A missing citation may call for a better page. A rival comparison may call for a page that answers the same question in plainer language.
Use the exact question in the report. In the CiteGraph sample, the scans asked 507 distinct buyer questions. That is enough variation to show why a broad brand score cannot explain what happened in every answer. The source for that figure is CiteGraph scan data, read live on the research page.
For each lost answer, record:
The full buyer question and the scan date.
The answer verbatim, with the engine named.
Whether the product appeared, was absent, or was unclear.
Whether your page was cited, another page was cited, or no useful source appeared.
The next page or fact worth testing.
Do not write a recommendation before showing the evidence. The pasted answer is the shortest route from a report to a decision.
When you have many answers, group them by failure rather than by engine. A founder usually needs to know which buyer questions are exposed, not read a tour of every platform. You can still retain the engine field for diagnosis.
Naming rates ranged from 6.1% on ChatGPT to 10.7% on Claude in the sample. The table below shows the rates and denominators for each engine.
Cited sources
The second section answers a sharper question: which pages supplied the answer when your page did not?
List the cited URL, domain, page type, and the claim the page appears to support. Add your interpretation only after recording the source. This helps separate a page that explains a category from one that supplies a product fact, comparison, review, definition, or proof point.
Citations are often the part clients can use fastest. A cited page is an example of material an engine could retrieve and use. It is not a guarantee that copying the format will earn a citation. Still, the page gives you something more solid than a general instruction to “create better content”.
CiteGraph has recorded 64,667 citations across 9,791 AI answers. Those answers came from 77 scans of 44 sites in 48 categories. The engines cited 9,192 distinct domains and 21,465 distinct pages. These figures come from the live CiteGraph research data.
That spread matters. A visibility report should not treat the winning source as a single competitor to outrank. It should show the set of pages doing the work across the questions you care about.
For each cited source, record:
The URL and domain.
The page type, such as a product page, review, comparison or guide.
The claim the page appears to support.
Whether the source recurs across questions.
The owned page or independent coverage that could address the gap.
The number of sources per answer also changes by engine. In the sample, Perplexity cited 5.5 sources per answer on average, Gemini cited 4.2, ChatGPT cited 3.5, and Claude cited 4.9. The figures come from the CiteGraph measurement method.
A report can include a source gap beside every lost answer: “No owned page cited”, “owned page available but not selected”, or “third party source dominates”. Keep the labels operational. The recommendation should name the page, fact or coverage gap that needs attention.
Action plan
The plan should be shorter than the evidence section. If it contains every possible SEO task, it is a backlog wearing a tie.
Give each action a target question, a page or source to change, an owner, and a test. Tie it to a lost answer or a cited source. “Publish more content” is not an action until the report names the buyer question and the page type.
A useful plan covers practical work:
Repair answerable pages with clear product facts, limits, audience, pricing context or comparison criteria.
Build missing comparison or alternative pages when the lost answers repeatedly ask buyers to choose between options.
Improve access and page clarity when a relevant page exists but does not appear among the cited sources.
Work on independent coverage when the answer depends on evidence outside your own site.
Use the measurement method guide to keep the test consistent. A before and after check should use the same questions, engines and run design. Otherwise, the report may measure a changed sample rather than a changed page.
The report should also state what the plan cannot do. It can improve the material available to an engine and test whether the tracked answers change. It cannot command an assistant to name a product, and it cannot turn a sample into a forecast.
If you need a quick site check before assigning content work, the free AI readiness check runs fifteen checks without an account and shows the fix for each. Keep that separate from citation evidence. A readiness pass is a list of conditions, not proof that an engine will cite the site.
Margin
A report without a margin makes small differences look like decisions.
Show the sample behind every rate. If a product appears in 10.5% of 2,440 Perplexity answers, the reader can see the size of the evidence. If a result comes from a much smaller slice, label it accordingly. The denominator belongs beside the percentage, not in a footnote that nobody opens.
The engine comparison below is useful as context, not as a league table. More sources per answer does not automatically mean more product mentions, and a higher naming rate does not automatically mean more useful citations.
Engine
Answers sampled
Sources per answer
Scanned product named
Perplexity
2,440
5.5
10.5%
Gemini
2,437
4.2
9.4%
ChatGPT
2,434
3.5
6.1%
Claude
2,120
4.9
10.7%
The same applies to source counts. “We found 21,465 pages” sounds broad, but that figure covers the full CiteGraph scan data, not one client category. A client report should give the local count and explain how the wider figure provides context, if it provides any.
Record the run date, question set, engines and scan count. Keep the raw answers and URLs. If a result moves later, you need to know whether the page changed, the question changed, the engine changed, or the sample simply moved.
Dossier structure
Send the report in an order that lets a busy client move from proof to work:
Scope: questions, engines, scans, dates, and the naming or citation rule.
Lost answers: verbatim answers with the product and citation status marked.
Cited sources: URLs, domains, page types, and the claims each source supports.
Action plan: pages or signals to change, owners, and the next test.
Margin: sample sizes, rates, caveats, and what the report does not measure.
That is the AI visibility report a client can use in a meeting. It does not bury the failure under a single index, and it does not confuse a technical check with a citation.
CiteGraph's pricing page says:
“Every CiteGraph plan starts with a 7-day trial, opened with a $1 charge that confirms the card, that includes 2 scans; a single scan without a subscription is $12.”
That detail matters because the reporting method and the tool cost are separate decisions. You can copy the structure without buying a tool, then choose a plan by the number of questions, runs and projects you need tracked.
Starter is $29 a month, covering 2 projects, 10 questions each, 3 runs per question on 4 engines, 6 tracked rivals and 30 written pages a month. Growth is $69 a month, covering 3 projects, 10 questions each, 5 runs per question on 4 engines, 10 tracked rivals and 60 written pages a month. Scale is $149 a month, covering 6 projects, 10 questions each, 5 runs per question on 4 engines, 10 tracked rivals and 200 written pages a month.
Pick the tier by how many questions you need tracked, not by the discount. A single scan without a subscription costs $12, and every plan starts with a 7-day trial opened with a $1 card confirmation and includes 2 scans.
An AI visibility report is the lost answers, the sources that beat you, the plan, and the margin around the result.
Questions people ask
What should I send a client first in an AI visibility report?+
Send the tracked questions and the lost answers verbatim first. Then show the cited pages, so the client can see what the engine used instead of relying on a score.
How many questions do I need for a useful report?+
There is no single correct count in the available data. Use a defined question set, record its size, and keep the same questions for the next comparison so the result has a clear margin.
Can an AI visibility report prove that my next page will be cited?+
No. It can show which pages appeared in the measured answers and identify changes worth testing. The result describes a sample, not every future answer from an engine.