The Short Answer
A mention shows that a brand or product appears in a sampled answer. A recommendation suggests it for a stated need. A citation identifies a source referenced by that answer. Referral traffic records visits attributed to a source in your analytics. These observations can overlap, but they answer different questions. Read the actual answer, record how it was collected and keep the sample size visible. A brand name appearing in a paragraph is useful evidence; it has not yet made a sale.
Key takeaways
- Separate the signals. Mentions, recommendations, citations, accuracy and visits describe different aspects of discovery.
- Keep the measurement conditions visible. Questions, provider, surface, dates and successful sample counts determine what a comparison means.
- Connect observations to decisions. Use the evidence to choose a content, accuracy or publishing action, then evaluate what happens next.
Why does one visibility number leave so much unexplained?
An overall score can help summarise a report, provided its calculation is clear. It becomes less useful when the team cannot tell what changed beneath it. More mentions could reflect more relevant answers, a larger sample or a new set of questions that includes the brand name.
The commercial meaning also varies. A product can appear as an unsuitable option, receive a qualified recommendation or be described using an outdated specification. All three might contain the same brand name. Their implications for the marketing team are rather different.
Begin with the underlying observation and work towards the summary. Ask what appeared, in response to which question, on which surface and with what supporting sources. A dashboard should help you find that information. It should not require an act of faith in a coloured circle.
For video campaigns, the useful connection is between the information you have published and what can be inspected in relevant answers. That makes the reporting specific enough to inform the next piece of work.
What counts as each kind of observation?
Agree your definitions before collecting results. In the reporting approach below, these are separate labels that can apply to the same answer. They are not stages every reader necessarily passes through.
| Observation | What you have observed | What it does not establish |
|---|---|---|
| Mention | The answer names the brand or product | That it endorses the product or describes it correctly |
| Recommendation | The answer suggests the product for a stated need | That the advice fits every customer or caused a purchase |
| Citation | The answer references a particular source | That the source supports every nearby claim or received a visit |
| Accuracy check | A statement has been compared with an applicable source | That every statement in the answer has been verified |
| Referral visit | Analytics attributes a visit to a source under its reporting rules | The full path that influenced the visitor |
| Commercial action | A tracked enquiry, signup or purchase occurred | That a particular AI answer caused the action |
A citation may point to your website, a retailer, a publisher or a discussion. Record the destination. A reference to somebody else's product review is a different publishing observation from a reference to your own campaign page, even if both mention your brand.
How can the same campaign produce several different stories?
Imagine checking a touring-car question across twelve planned answer samples. Ten return usable responses under the chosen test conditions; two checks fail. Within the ten successful samples, six mention the car, three recommend it for the question and two cite the brand's campaign page.
For this example, the mention rate is 6 out of 10, the recommendation rate is 3 out of 10 and the campaign-page citation rate is 2 out of 10. The labels can overlap, so the counts should not be added together. Report the two failed checks separately.
Now read those answers. One recommendation may concern a different market. Another answer may cite the correct page while confusing two variants. The arithmetic remains correct, but a useful report will highlight the representation problem alongside the count.
These are illustrative numbers, not AdCyte results. They show why a denominator and the original answer matter. “We appeared in 60%” leaves out almost everything someone needs to judge the result, including what the other 40% means.
What exactly is being sampled?
Record the provider, the surface and how the response was obtained. A monitored API response and a captured answer in a consumer application are different observations. They may involve different settings, retrieval behaviour and surrounding context.
Do not label an OpenAI API sample as what customers see in ChatGPT. Likewise, a Gemini API response is not a capture of Google AI Overviews or AI Mode. A report can be useful within a defined scope without claiming to represent every experience carrying a related brand name.
Keep the exact question and any relevant settings with the result. Distinguish a fresh question from a follow-up conversation. Record the time and any available information about the model or search mode. Where a detail is unavailable, say so rather than filling it in from an assumption.
The scope of the observation sets the scope of the conclusion. Ten inspected answers can tell you about those ten answers. They cannot establish how an entire market encountered the brand that week.
How should failed checks and missing data appear?
A successful answer without a brand mention is a valid negative observation for that measure. A failed request does not tell you whether the brand would have appeared. Keep those outcomes separate so missing data does not masquerade as poor performance.
Define eligibility before calculation. In the worked example, an eligible sample is a successful, inspectable response collected under the intended conditions. An incomplete response that cannot be assessed should be recorded as unavailable, with the reason where known.
Show coverage alongside the result: planned samples, eligible samples and failures. If one provider supplies very few usable responses, avoid presenting its percentage as equally informative as a much larger sample from another. A decimal point can make a thin result look remarkably well nourished.
Which comparisons are actually useful?
Compare compatible observations. Keep the question set, provider, surface and collection method consistent where possible. Record changes that affect the comparison, such as adding new questions or moving from an API workflow to a consumer-interface study.
Keep branded and unbranded questions distinguishable. “Tell me about this named product” and “Which products suit this need?” test different things. A higher mention rate caused by adding the brand name to every question is hardly a discovery breakthrough.
Third-party metrics also have their own definitions. Microsoft's explanation of Citation Share in Bing Webmaster Tools describes a share of citations for a grounding query and treats it as observational, rather than traffic share. It should not be substituted for the answer-level citation rate in our example.
Keep those definitions attached to the numbers in your report. Two percentages can look comparable while describing different populations and different events.
Where do website visits and commercial outcomes fit?
Look at visits and actions alongside answer observations, using the appropriate analytics and attribution records. A citation can exist without a recorded click. A visitor may arrive after several influences, including paid advertising, earlier research and a direct recommendation from someone else.
State what the data supports. You might report a recorded citation, a set of attributed referral visits and several enquiries during a period. Connecting them causally requires more evidence than their appearance on adjacent slides.
For campaign teams, that separation is useful. It lets you discuss publishing coverage, observed representation and commercial behaviour without forcing them into a single claim. Each can reveal a worthwhile action even when the complete customer journey is unavailable.
What should the team do with the result?
Start with the most concrete issue. An incorrect product statement calls for an accuracy investigation. A relevant question with no clear public answer may justify new or improved content. A citation to an outdated page suggests checking that destination and its current wording.
An absent mention alone does not establish a content defect. Inspect the question, the answer and your available evidence before deciding what to change. Our article on responding to incorrect product descriptions covers one particularly actionable case.
Keep the work connected to the campaign. Analyse every ad through AdCyte, use its findings and approved facts to prepare useful material, and record where that material is published. Choose monitoring questions with a clear connection to customer needs; this question-selection approach can help.
Finish the report with a small number of specific actions, their owners and the evidence behind them. A useful result might be “correct the variant explanation and repeat the same checks”. It is considerably easier to act on than “increase visibility”.
See what sits behind the number
Use AdCyte to connect campaign knowledge with useful published content and inspectable AI-answer observations. Explore AI Visibility and read the methodology to understand the scope of the results before deciding what to do next.
Frequently asked questions
Is a mention the same as a recommendation?
No. A mention records the presence of the brand or product name. A recommendation suggests it for a particular need or circumstance. Read the surrounding wording: the product could be mentioned as an unsuitable option, in a comparison or as background context without being recommended.
Does a citation mean someone visited our website?
No. A citation identifies a referenced source in an answer. A visit is a separate event recorded through the available analytics. Inspect the cited destination and report referral data separately. Neither the citation alone nor a nearby increase in visits proves a purchase or complete customer journey.
Can one answer be counted in several categories?
Yes. An answer can mention a product, recommend it and cite a supporting page. Report each measure with its definition and eligible sample count. Do not add overlapping categories together as though they represent different people or mutually exclusive outcomes; they describe properties of the observed answers.
Should a failed check count as a missing mention?
No. A failed or unusable check does not provide an answer you can assess. Report it separately from a successful answer without the signal being measured. Define sample eligibility before calculating rates, and show how many planned checks produced usable observations so readers can judge the coverage.
Does a cited campaign page prove the AI watched our video?
No. A page citation does not establish which page elements were processed or how the answer used them. The page may contain text and other supporting information. Keep the claim limited to the observed reference unless you have specific evidence for how the video itself was processed.
Can these observations prove improved ROAS?
They can inform campaign discussions, but they do not independently establish return on advertising spend. ROAS needs appropriate spend, revenue and attribution evidence. Keep content work, sampled visibility, website behaviour and commercial outcomes identifiable so the team can understand the contribution and limits of each measure.




