Reporting checklist
A good AI visibility report shows citation, mention and recommendation rate plus share of voice, broken out per engine and traceable to the exact prompt, answer, source and timestamp behind each number. Anything that reports a single blended score, hides the prompts, or substitutes website traffic for AI visibility is not measuring the channel — it is decorating it.
Why the report design matters
The same prompt returns different answers across runs and engines, which is why what a report measures — and how honestly it admits its own uncertainty — decides whether the numbers mean anything.
First principles
An AI visibility report is the artifact that tells you whether Answer Engine Optimization work is changing how AI engines mention, cite and recommend a brand. Because AI answers are synthesized fresh from many sources and shift with every model update, a report is not a leaderboard snapshot — it is a sampling instrument. The most important thing to inspect is not the headline number but the method underneath it: which prompts were sampled, across which engines, how many times, and whether you can trace any figure back to the exact answer that produced it.
The honesty of a report shows up first in how it treats variability. The same prompt returns different answers across runs: in one study of 22.5 million ChatGPT Shopping offers, 95% of product titles appeared in fewer than 30% of repeated runs of the identical prompt. A report that presents a single answer, or a single week, as proof of anything is mistaking noise for signal. Credible reporting samples a stable prompt portfolio repeatedly over time and shows the trend, with its uncertainty, rather than a lucky screenshot.
Many tools compress everything into a single AI visibility score. That can be useful as a headline, but it is the least informative line in the document. Citations behave like a power-law distribution with large sample-to-sample variability, so a blended score can move for reasons that have nothing to do with your work. The sections that follow — rates, share of voice, per-engine breakdowns, source grounding and the prompt-level drilldown — are what let you explain why the score moved, which is the only thing that makes a report actionable. If a report has the score and not the sections, treat it as a black box.
Section by section
Read a report in this order. Each section answers a different question, and a report missing any of the first five is incomplete rather than concise.
Before any chart, the report should state what it measured: the prompt portfolio (which buyer-intent questions, grouped by intent), how many prompts, how often each was sampled, which engines were covered, and over what window. Without this, every later number is unfalsifiable. A method note is the precondition for trusting the rest of the document.
These are three different outcomes, not synonyms. Mention rate is how often the brand name appears in answers; citation rate is how often the brand's own content is cited as a source; recommendation rate is how often the brand is actively recommended. A brand can be mentioned without being cited, and cited without being recommended. A report that collapses them into one 'visibility' number is hiding the most decision-relevant distinction.
Absolute rates mean little without a denominator. Share of voice — your share of the citations and recommendations in a category versus named competitors — turns 'we appear sometimes' into 'we hold X% of the answer space, and competitor Y is gaining.' Require a competitive delta: who is gaining or losing share against you, by category, period over period.
ChatGPT, Google AI Overviews and AI Mode, Gemini, Claude and Perplexity cite different sources, in different volumes. In one study of 127,198 citations, 69.6% were made by a single engine and only 2.7% of domains were cited by all five; the average sources per answer ranged from 3.7 in ChatGPT to 11.0 in Gemini. Aggregating across engines masks exactly the differences you need to act on. Insist on a section per engine.
A serious report shows not just that you were cited but which of your pages were pulled, and how often an answer linked a source versus synthesized without attribution. Because most citations are third-party, this section should also surface the external sources — reviews, directories, press — that engines read about you, not only your own domain.
Appearing is not the same as being described well. A credible report captures the tone and framing AI engines use when they mention the brand, and flags factual errors or 'brand hallucinations' in the answers. Visibility without credibility is a liability, so the context of a mention belongs in the report, not just its count.
Finally, the report must let you drill from any metric down to the underlying prompt, answer, source and timestamp, and should track drift and volatility over time so a model update reads as a step-change, not an unexplained dip. If you cannot open the evidence behind a number, the number is decoration.
How to read it
The same sections, done honestly or done for show. Watch the method, not the polish.
| Section | Measures the channel | Decorates the channel |
|---|---|---|
| Headline | A score plus the sections that explain it | A single blended score with nothing underneath |
| Method note | States prompts, engines, sampling, window | No method; numbers cannot be reproduced |
| Rates | Citation, mention and recommendation reported separately | One 'visibility' number conflating all three |
| Competitors | Share of voice vs named competitors, with deltas | Your numbers only, no denominator |
| Engines | Broken out per engine | Aggregated across engines |
| Evidence | Drill to prompt, answer, source, timestamp | Polished charts, no prompt-level access |
| Traffic | AI visibility kept distinct from web traffic | Referral-traffic graphs passed off as AI proof |
Questions to ask the report
Whether the report comes from an agency or a tool, these questions separate measurement from theater.
"Show me the prompt portfolio, the engines, and how often each prompt was run."
"Is this mention, citation or recommendation? They are not the same outcome."
"What is our share of voice against named competitors, and who is gaining?"
"Can I see this per engine? A blended number hides where we actually appear."
"Which of our URLs and which third-party sources were pulled into answers?"
"Open the exact prompt, answer, source and timestamp behind this number."
Definition
A periodic, prompt-portfolio-based document that measures how often AI engines mention, cite and recommend a brand — reported per engine, with share of voice and prompt-level evidence behind every number.
An AI visibility report is a sampling instrument, not a scoreboard. Because AI answers are probabilistic and assembled mostly from third-party sources, a credible report samples a stable prompt portfolio over time, separates mention, citation and recommendation outcomes, breaks results out per engine, and lets the reader trace any figure to the prompt, answer, source and timestamp that produced it. A report that shows only a blended score is describing visibility it cannot evidence.
Where this fits
Knowing what a report should contain is half of an evaluation; the other half is checking that the agency or tool producing it can show this evidence on request. The companion guide on how to evaluate an AEO agency turns these reporting sections into discovery-call questions, and the directory criteria explain the published-methodology and measurement-based standards an agency is expected to evidence. Together they let a buyer tell a real reporting practice from a polished dashboard with nothing behind it. Our own methodology page describes the prompt-portfolio approach the AI Visibility Index uses to report at the category level.
Disclosure, because this touches the directory's neutrality: the operator of this portal also runs the agency Blobic, which appears in the directory under the same public criteria as every other firm, with a disclosure badge and no preferential ranking. No placement is paid and no position can be bought. We publish reporting standards as an independent observatory, not as a vendor pitching its own dashboard — companies looking for a provider are pointed to the directory, never sold to here.
FAQ
At minimum: a method note (the prompt portfolio, engines, sampling frequency and window); citation, mention and recommendation rate reported separately; share of voice against named competitors; a per-engine breakdown; source grounding showing which URLs were pulled; sentiment and answer context; and a prompt-level drilldown so any number traces to the exact prompt, answer, source and timestamp behind it.
The ones that map to different outcomes: mention rate (does the brand appear), citation rate (is the brand's content cited as a source), recommendation rate (is the brand recommended), share of voice (your share versus competitors), and sentiment (how the brand is described). Drift and volatility over time matter as much as the levels, because AI answers change with every model update.
Because engines cite different sources in different volumes. In one study of 127,198 citations, 69.6% were made by a single engine and only 2.7% of domains were cited by all five, with average sources per answer ranging from 3.7 in ChatGPT to 11.0 in Gemini. A blended cross-engine score hides where a brand actually appears, so reports should include a section per engine.
No. A single blended score is the least informative line in a report. Citations follow a power-law distribution with large sample-to-sample variability, so a score can move for reasons unrelated to your work. The per-rate, per-engine and prompt-level sections are what explain why a score moved, which is the only thing that makes a report actionable.
No. Referral traffic from AI tools is a separate, heavily undercounted signal and should never be presented as proof of AI visibility. A report that substitutes a web-traffic graph for citation, mention and recommendation data is measuring the wrong thing. Keep AI visibility and web traffic in distinct sections.
Next step
Use these sections as your acceptance criteria, then start your shortlist from agencies listed against public criteria. Agencies that report this way can apply to be listed.