Agencies should report AI visibility to clients by separating brand mentions, positive recommendations, website citations and factual accuracy across a fixed set of buyer questions. Every report should state the collection surface, dates and denominator on page one, and connect findings to completed website work without presenting visibility counts as financial return.
Client stakeholders are rightly sceptical of opaque AI scores. When an agency presents a single percentage without explaining which prompts were tested or how citations were counted, account teams struggle to defend month-to-month changes. This guide provides a clear reporting structure and a reusable template for agencies delivering Generative Engine Optimisation (GEO) services.
Core rules for client reporting
State the prompt cohort, answer surfaces, dates and completed-run denominator before showing any percentages.
Report brand mentions, recommendations, domain citations and accuracy as separate metrics.
Keep unbranded category discovery results separate from branded verification checks.
Present analytics referrals and commercial enquiries alongside visibility data without conflating them.
Why single AI visibility scores confuse clients
Many early AI reports collapse different engines, question types and evidence records into one headline “visibility score”. While a single number looks tidy on an executive slide, it hides the operational facts an agency needs to explain performance and choose the next piece of work.
Consider what happens when a client’s educational blog post is cited in three general definitions, while their core service remains absent from every shortlist answer. A blended score may rise even though undecided buyers still do not see the client recommended. Equally, a client may be recommended while third-party reviews supply the cited links.
Academic work on generative search also highlights why visibility is multi-dimensional. The original KDD 2024 GEO research evaluated source visibility using multiple distinct measures, including word count and position within cited responses. Commercial client reporting benefits from the same clarity: define each metric plainly so the client knows what was observed.
Define the scope, surface and denominator on page one
Open every monthly or quarterly report with a compact methodology box. Clients should never have to guess whether an observation came from a consumer web interface, a search-enabled API call or a direct model prompt without live retrieval.
Include five scope statements at the top of the deliverable:
Prompt cohort and version: state how many unbranded discovery prompts and how many branded verification prompts are in the locked sample, and note whether the cohort changed since the last report.
Collection surfaces and settings: name each engine, model or mode tested, along with the language, geographic context and whether web search retrieval was active.
Observation window: record the exact dates when the checks were collected.
Completed-response denominator: state the total number of scheduled checks, the number of completed responses analysed and the number of failed runs or timeouts excluded from rate calculations.
Evidence retention: confirm where the client or account manager can inspect the full saved answer text and source lists behind any figure in the report.
Keeping the denominator visible prevents misinterpretation when collection errors occur or when an agency expands a prompt library. Our guide to interpreting AI visibility evidence explains how to document denominators and source provenance consistently.
Report four core visibility measures separately
Instead of a blended index, organise the main scorecard around four distinct questions that map directly to client value and agency actions.
Brand mention rate in completed answers
A brand mention records whether the client’s company, product or domain name appeared in a completed response. Report unbranded mentions separately from branded prompts, and verify that ambiguous company names have been checked against the surrounding answer text so false matches are excluded.
Positive recommendation rate for target tasks
A recommendation records whether the response presented the client as a suitable option for the buyer’s stated requirement. If an answer names the client only to say that its service does not cover the requested sector or region, record that as a mention with a negative qualification rather than a recommendation.
Website citation and source provenance
A website citation records whether the response attached a link to the client’s domain as supporting evidence. Distinguish between citations attached to passages in the final answer, wider sources consulted during search retrieval and plain URLs printed in the text. OpenAI’s web search documentation describes how inline citations differ from searched sources.
Factual accuracy of brand descriptions
Across both unbranded and branded prompts where the client appears, check whether the answer states core facts accurately: service scope, target customer size, geographic coverage (such as Wales or UK-wide delivery), pricing structure and integrations. Flag any recurring inaccuracies alongside the cited source that appears to supply the outdated statement.
How to present competitor share of voice fairly
Clients naturally want to know how often competitors appear in the same answers. Frame competitor comparisons strictly as share of appearances within your tested unbranded prompt cohort, not as total market share across all AI users.
For example, state: “Across 30 completed unbranded shortlist observations this month, Client A was recommended in 8 answers, Competitor B in 14 answers and Competitor C in 11 answers.” Note that multiple providers are usually recommended within a single response, so recommendation rates across competitors do not need to sum to one hundred per cent.
Pair the competitor tally with a short breakdown of the third-party domains and competitor pages most frequently cited in those answers. Showing a client that a competitor is being cited via a detailed comparison guide or an industry regulator directory turns a frustrating comparison into a clear editorial or digital PR task.
Connect visibility observations to traffic and actions without overclaiming ROI
Account teams build long-term trust when they keep visibility observations, technical readiness, website analytics and commercial pipeline in distinct sections of the report. Each layer answers a different question and has its own measurement limits.
Answer observations: show whether the brand is named, recommended and cited in the monitored prompt sample.
Crawler and technical checks: show whether verified search crawlers can fetch priority pages, and report indexing status separately where the search platform provides it.
Website referral analytics: show measurable visits arriving from identifiable AI surfaces or referrers in your analytics platform.
Commercial enquiries and conversions: show tracked leads, trials or sales recorded in the client’s analytics or CRM system, including self-reported attribution where available.
Explain platform reporting boundaries clearly to clients. Google now provides a Generative AI performance report in Search Console with impressions for AI Overviews and AI Mode; it may not appear when there are too few impressions. Those data are also included in the Web search performance report. Keep this first-party data separate from your own sampled prompt observations.
Equally, many buyers read an AI summary and later visit a site through a branded search or direct bookmark. Never relabel unattributed direct traffic as proven AI referrals, and never convert a rise in mention counts into a fabricated revenue figure.
An annotated fictional client report example
The following report summary is entirely fictional. It demonstrates how an agency can present monthly findings with clear denominators and honest limits, rather than reporting a real NeuraCite client result.
Suppose a Welsh digital agency prepares a monthly report for a fictional commercial insurance broker, “Cambrian Risk Advisory”. The locked cohort contains fifteen unbranded buyer prompts for UK and Welsh SMEs, plus five branded verification prompts. Each prompt is observed once per month across two search-enabled surfaces, scheduling forty checks in total.
Scope and run accounting: 40 checks scheduled on 10 October; 39 completed responses analysed (29 unbranded, 10 branded); 1 unbranded check failed due to a temporary provider timeout and is excluded from percentages.
Unbranded discovery results (29 completed responses): Cambrian Risk Advisory was mentioned in 7 of 29 responses, positively recommended as a suitable broker in 5 of 29 responses, and received an attached website citation in 3 of 29 responses.
Branded verification results (10 completed responses): the broker was mentioned in all 10 responses, with 9 of 10 responses accurately describing its FCA regulation status, Cardiff office and SME sector focus. One response cited an archived PDF with an old telephone number.
Competitor context (29 completed unbranded responses): National Broker X was recommended in 16 of 29 responses, supported primarily by citations to its sector-specific policy pages and an independent UK insurance directory.
Completed actions this period: published a dedicated cyber insurance scope and claims-process page; removed the outdated PDF from the server and redirected its URL to the current contact page.
Next period priorities: expand the professional indemnity service page with clear policy limits and sector eligibility; review the independent UK insurance directory profile for missing regional credentials.
Because every figure carries its denominator and links directly to a completed or planned task, the account manager can walk the client through the report in ten minutes without defending a black-box score.
A reusable monthly agency reporting template
Use this seven-part structure when assembling your agency’s monthly or quarterly AI visibility deliverable:
Executive summary: three bullet points covering key observations in the unbranded cohort, work completed during the period and the primary focus for next month.
Methodology and denominator table: prompt cohort version, unbranded vs. branded prompt counts, observation surfaces, collection dates, completed responses and excluded failed runs.
Unbranded discovery scorecard: mention counts, positive recommendation counts and domain citation counts across completed unbranded responses, compared on a like-for-like basis with the previous period.
Branded accuracy and source audit: verification of how the client’s offering, pricing and location are described in branded checks, noting any inaccurate claims or outdated third-party sources.
Competitor and cited-source breakdown: competitor recommendation counts within the same unbranded sample, plus the top owned, competitor and third-party URLs cited.
Technical and referral context: confirmed crawler access status from Agent Analytics or server logs, alongside verified referral sessions and conversions where analytics tracking permits.
Action log and next decisions: status of technical, page and evidence improvements from the previous report, plus the prioritised tasks for the upcoming cycle.
If your agency is formalising its GEO reporting workflow, review our guide on how to choose a generative engine optimisation platform, explore NeuraCite Answer Engine Insights and Visibility Reports, and check NeuraCite agency pricing to match workspace allowances to your client portfolio.