AI visibility report template for clients (UK agency edition)

Clients have started asking whether they show up in ChatGPT, and agencies are being asked to report on it before the measurement is settled. In AgencyAnalytics' 2026 benchmark of 494 agency professionals, 66% said they see client demand for AI-search work and 41% reported problems measuring its return. The survey's publisher sells an AI search tracker, so read it as an indication rather than a census.

This template is for the document that goes to the client: the AI search visibility report you would be willing to defend line by line. It sets out the sections, shows a redacted example record, lists what to leave out and gives caveat wording you can paste. It is written for UK agencies, but nothing in it depends on the market.

What the client report is for

A client report on AI answers is a record of what named answer engines said, when, in reply to which questions, and what the agency decided to do about it. It is not a ranking report. Insites argues that agencies need two kinds of AI visibility report rather than one (a vendor blog). Whichever split you use, the client-facing report has the strictest job, because it is the one forwarded to a board or a sales director who was not on the call.

That job comes with one rule: every figure must be traceable to answers that were actually collected. If a reader asks where a number came from, the report should already show the questions, the engines, the dates, how many times each question was asked and who checked the answers.

The template at a glance

1. Scope and method

Put the method first, in a box, because every number below depends on it. List the questions exactly as they were asked (unbranded buyer questions, not "is [Client brand] any good?"), the target market, each engine with the exact model that answered, whether that engine searched the web, the dates of collection, how many times each question was asked and who reviewed the answers.

A review of how AI visibility tools collect data argues that disclosure matters more than whether a tool reads the API or the consumer interface. The same holds for an agency report: a client can weigh a method they can see.

For reference, this is what an AnswerLift record holds: the prompt and its version, the engine, the model, the search method, any cited sources, and the capture and review times. Answers are collected from the providers' APIs through the OpenRouter gateway. The models behind ChatGPT and Claude are asked with their vendors' own web search, Gemini is asked with no search grounding, Perplexity's Sonar always searches, and five comparison models answer from model knowledge. The consumer apps are never scraped. Where a client wants to see what a person sees, someone checks the ChatGPT app, Perplexity or Google AI Overviews by hand and files it as a hand-recorded observation. The full method is on the methodology page.

2. Named in k of n runs, with dates

The headline figure is a count with its denominator, for example "named in 2 of 3 runs, OpenAI GPT API with web search, 3 to 4 November 2026". Show it per engine and, where it helps, per question.

Repetition matters because a single answer is a weak sample. In a study reported by Search Engine Journal, SparkToro found the same list of brands in under 1 in 100 runs, and the same order in under 0.1%, across 2,961 runs of 12 prompts. The study was run with Gumshoe, which sells an AI visibility tool. SparkToro's follow-up argues that how often a brand appears across many runs is steadier than any one answer. Ahrefs, which sells AI visibility tracking, puts it this way: "Think of AI tracking less like rank tracking and more like polling."

Polling has a rule worth borrowing: report the sample size. Do not show a percentage when fewer than five answers were collected. "2 of 3" is honest. A percentage from three answers implies a precision the sample cannot carry, because each answer is a third of the figure.

A Sprint run in AnswerLift asks each question three times of each engine by default and reports "named in k of n runs", with no percentage when n is under 5.

3. Cited domains

For the engines that searched the web, list the domains their searches cited: the domain, the engine, how many of the collected answers cited it, and a flag for the client's own domain and for known competitors. This is often the most useful table for planning, because it shows which third-party sites the answers lean on.

Two cautions. First, engines asked from model knowledge produce no citations, and none should be shown for them; in AnswerLift's panel that includes Gemini, which is asked without search grounding. Second, citations move. An Ahrefs-sponsored article in Search Engine Journal reported that AI Overviews' wording changed 70% and their cited sources 45.5% between observations. A domain cited once is a lead, not a pattern, so show the count and let the reader tell the difference.

4. Verbatim answer excerpts

Paraphrase is where a report starts to drift from the evidence. Quote the stored answer and put the provenance next to it: the question and its version, the engine and model, the date and time, whether the engine searched, the sources it cited, and the reviewer who checked it, with the date of review. An illustrative, redacted record looks like this:

Question (version 2): "Which accountants would you recommend for a UK e-commerce seller?"

Engine: Claude API (web search) · model [model id as recorded] · [date and time collected]

Excerpt: "… [Competitor A] and [Client brand] are both … "

Cited sources: [comparison site], [client domain]

Classification: cited · Reviewed by [reviewer name], [review date]

Hand-recorded records follow the same pattern, labelled as observed in the app by a named person on a stated date.

5. What changed since last time

Trend is what clients ask for, and it is where reports overclaim most. Compare like with like: the same question, the same engine and model, the same search mode and the same number of runs. Show earlier and current counts side by side, for example "named in 1 of 3 runs in October, 3 of 3 in November", and flag any change that sits within the variation the repeated runs already show as "may be noise".

Never say that a change was caused by the work. Answers drift without anyone touching the site, as the figures above show. Wording you can use is in the caveats below.

6. Remediation backlog

The report should end in work, not in a number. List each action decided in response to the evidence, with an owner, a priority, a due date and a status, and link it to the records that prompted it: "Publish a comparison page that answers [buyer question 3]; prompted by [record references]; owner [name]; due [date]." State the action and the reason. Do not state an expected change in the answers.

7. Caveats you can paste

Edit these to match what you actually collected:

These answers were collected from the providers' APIs, not from the consumer apps. The apps add their own instructions, memory and personalisation, so a person asking the same question may see a different answer.

Each figure counts answers that were actually collected. "Named in 2 of 3 runs" means three answers were collected and two named the brand. Where fewer than five answers were collected we show the count, not a percentage.

AI answers vary between runs, days and users. Changes marked "may be noise" are within the variation we observed and should not be read as a trend.

[Surfaces] were not measured automatically. Where they were checked by hand, the record says so and names the person.

Nothing in this report shows that a change to the website caused a change in an answer.

The first caveat is not a formality. An academic study that asked 80 questions 20 times each found that the API, even combined with Google results, recovered only about 26 to 28% of the URLs ChatGPT cited. API answers are evidence of what the model says. They are not a screenshot of the app.

What to leave out

• A single score. A number out of 100 hides which questions, engines and runs produced it. Stellar Search, a London agency, calls single percentages of this kind false precision.
• Share of voice without a denominator. Share of which questions, on which engines, across how many runs? If you show it, show the counts it was made from.
• Position as a KPI. Where a brand sits in a list changes from run to run, and SparkToro's follow-up treats position as noise. Report whether the brand was named, not where.
• Causal claims. See the section on what changed.
• Unreviewed records. If nobody has read an answer, it should not be in front of a client. In AnswerLift, unverified records are excluded from the client report by the query that builds it, not by the page.

PDF, print and link expiry

Many agencies report through dashboards: in a sitechecker survey of 73 agencies (vendor-authored), 67% used Looker Studio. If AI visibility figures go into a dashboard, keep k, n and the dates in the labels, because a chart tile without them is a score by another name.

For the record, send a dated PDF, and if you share a link, let it expire. An evidence report is client-confidential, and a link that lives for ever outlasts the context it was written in.

AnswerLift's share report works this way. It shows verified records only, it is read-only, it carries no AnswerLift analytics, it can carry your agency's logo, it can be printed or saved as a PDF from the page, and the link expires 30 days after it is issued.

Start with real data

The quickest way to test this template is on one of your own clients. The free AI Visibility Snapshot puts three buyer questions to the models behind ChatGPT, Claude, Gemini and Perplexity and returns what each said and which sources it cited, with every answer checked by a person. The Visibility Sprint produces the full verified ledger and one client report. For why answers change between runs, and how to report that honestly, read why AI answers change every time.