What Agencies Should Report to Clients About AI Visibility
A practical client reporting framework for AI visibility, covering scope, mentions, citations, crawler access, traffic, and the work completed.
An AI visibility report should help a client decide what to do next. A page of charts without scope, interpretation, or a record of completed work leaves the client to guess. We structure reporting around a simpler question: what changed in the parts of AI search that the engagement can influence, and what evidence supports the next action?
That starts with discipline about what the report can prove. A monitored answer is an observation from a selected model, prompt, location, and date. It is not a census of everything an AI system has ever said about the brand. Referral traffic is measurable when the browser sends a referrer, but it does not capture every visit influenced by an AI answer. We put those limits near the front rather than hiding them in a footnote.
State the measurement scope
Every report needs a scope block. We list the monitored models, markets, languages, prompt groups, competitors, and reporting period. We also note any setup change that breaks comparability, such as adding a brand alias or replacing a prompt halfway through the period.
This sounds administrative, but it prevents bad conclusions. A visibility increase after doubling the number of navigational prompts is not comparable with the prior period. A blended score can also conceal a gain in one model and a loss in another. We keep the overall view for orientation, then split the analysis by model and topic where the difference would change the recommendation.
The comparison set deserves the same care. Share of voice only means something relative to the named competitors. Adding distant brands can make a client's percentage move without any change in actual answers. Our report names the set so the reader knows what the percentage describes.
Separate mentions from citations
A brand mention and a citation are different events. The model can name a company without linking to its site, or cite one of its pages without making the brand prominent in the answer. We report them separately because the work behind them differs.
For mentions, we look at coverage, position in the answer, share of voice, and sentiment. For citations, we look at cited domains, the client's own cited pages, citation position, and self-citation rate. Promptwatch's metrics overview documents how its measures are defined. We use those definitions when the chart comes from the platform, rather than presenting a product metric as a universal industry standard.
A client often needs examples alongside the averages. We include a small set of real responses that explain a pattern, with the model, prompt, and observation date attached. One response should not overrule the trend, but it can show why a brand receives a weak score despite being mentioned.
Report access before content performance
Content cannot earn a fresh citation if the relevant retrieval system cannot fetch it. The technical section should therefore say which important pages received successful crawler requests, which returned errors, and which expected crawlers were absent during the period.
We do not collapse every AI bot into one count. Training crawlers, search index crawlers, and live citation fetchers serve different purposes. A decision to block a training bot does not automatically imply that the brand should block a search or user-triggered fetcher. Any recommendation also needs a check against the provider's current official documentation because names and controls can change.
Promptwatch's agency GEO guide connects crawler readiness with visibility and client reporting. That is product guidance, so we treat it as a workflow reference, not independent proof that a given technical change caused a ranking movement.
Connect visibility to visits carefully
The traffic section should answer which AI referrers sent measurable visits, which pages received them, and how the current period compares with a valid prior period. It should not imply that every AI-influenced visit carries a detectable referrer.
We call referrer-attributed visits a measured floor. Some browsers or answer experiences may omit the referrer. A person may also see a recommendation and return later through a direct visit or branded search. We can discuss those paths as possible influence, but we do not estimate a hidden traffic total without evidence.
Conversion and revenue remain in the client's analytics and CRM. If those systems can segment AI referrer domains, we report conversions from that segment at landing-page level. We keep the distinction visible: the visibility platform tracks answers, citations, crawls, and identified referral visits, while the client's systems hold the commercial outcome.
Put shipped work next to movement
Clients need to see what the agency did during the period. We maintain a dated change log of pages published, pages revised, access issues fixed, offsite work completed, and measurement changes. Each item links to the affected URL or ticket where possible.
The change log does not prove causation. It gives the reader a sensible timeline. If citations rose for a revised page after a successful crawl, that sequence supports further investigation. If every monitored prompt moved on the same day, a model change may be a better explanation than one article we published.
We report unresolved work too. A 403 that remains with the client's infrastructure team belongs in the report, as does a content brief waiting for legal approval. This keeps delivery conversations factual and stops the next period from reopening old diagnostic work.
Match the format to the decision
An executive page should be short: scope, material movement, business evidence, completed work, and next actions. Specialists can use an appendix with prompt and page detail. Different readers need different depth, but they should read from the same underlying data.
Promptwatch supports downloadable project and monitor reports. Its PDF report documentation explains which charts and filters appear, while the sharing guide covers scheduled delivery and its access rules. These are useful delivery mechanisms, not a substitute for agency interpretation. We still add the scope notes, work log, caveats, and recommendations that turn a snapshot into a client report.
We run client programs on Promptwatch because it keeps monitored answers, citations, crawler logs, and identified AI referral visits close enough to investigate as one chain. Our recommendation is to use that joined evidence as the reporting base, then keep conversions and revenue in the client's analytics and CRM. The final page should end with owners and due dates for the next actions, since that is where reporting becomes operating work.