Cited Is Not Recommended: What AI Visibility Metrics Actually Tell You

A citation shows that an AI answer displayed a source. It does not show that the answer mentioned your brand, preferred it, or influenced a buyer. In our latest sample, 810 answers contained citations but no tracked-brand mention. A useful AEO or GEO report must keep those outcomes separate.

---

The Practical Takeaway

• An impression means your page appeared somewhere in an AI search experience.
• A citation means a source was displayed; a mention means the brand appeared in the answer.
• A recommendation, referral, and conversion are later outcomes that need their own evidence.

The reporting mistake is collapsing this ladder into one “AI visibility” score. A rising citation count can look encouraging while the brand remains absent. A brand mention can matter even when no page from the brand’s site is cited. A recommendation can happen without a measurable click.

What Each Metric Can Establish

Layer	What it tells you	What it does not tell you
Discovery	The engine could find or retrieve a page	The page shaped the answer
Impression	A URL appeared in an AI search feature	The user read or trusted it
Citation	A page was displayed as a source	Ranking, authority, recommendation, or influence
Brand mention	The answer named the tracked brand	Positive sentiment, preference, or buying intent
Recommendation	The answer presented the brand as a suitable choice	A click, lead, or purchase
Referral	A person arrived from an AI surface	Business value
Conversion	The visit produced a defined outcome	One citation caused the outcome

Google’s generative-AI Search Console reports show impressions and appearing pages, not whether a response recommended the business. Google introduced the reports as a limited rollout. Bing reports cited pages and sampled grounding queries, while warning that citation counts do not indicate ranking, authority, placement, or page role. See Bing’s documentation.

OpenAI documents ChatGPT search referrals through utm_source=chatgpt.com. This measures visits, but no-click mentions remain outside ordinary web analytics. See OpenAI’s publisher guidance.

What the Sample Showed

The clearest gap was between source activity and brand visibility.

Observation	Answers	Share of all answers
Contained at least one captured citation	928	73.2%
Mentioned the tracked brand	164	12.9%
Contained citations but no tracked-brand mention	810	63.9%
Mentioned a competitor but not the tracked brand	628	49.6%

Among answers with citations, 87.3% did not mention the tracked brand. This does not make the citations useless. It means citation volume and brand visibility answer different questions.

The same audit should then examine whether a mentioned brand was merely listed, described as a possible fit, clearly recommended, or rejected. Recommendation was not automatically classified in this dataset, so the report does not infer it from citation or mention counts.

Citation Selection Is Not Citation Influence

Even the citation layer can be split again. A page may appear in the source list without materially shaping the response.

A recent pre-submission paper calls this the difference between citation selection and citation absorption. Absorption means the source appears to contribute language, evidence, structure, or factual support. The authors frame the results as descriptive. Read the framework on arXiv.

For a working audit, ask:

1. Is the cited page relevant to the buyer’s question?
2. Does the answer reuse a definition, comparison, fact, or procedure from it?
3. Is your brand present in that source, the answer, both, or neither?
4. Does the answer describe the brand neutrally or recommend it for a use case?

A Better AEO and GEO Scorecard

Use a small scorecard instead of one composite score.

Question	Record	Next action when weak
Can the engines access the relevant page?	Yes, no, or unknown	Check indexing, robots controls, and page availability
Is the page appearing?	Page, provider, and prompt group	Improve topic fit, clarity, and extractable evidence
Is the brand named?	Present, absent, or inconsistent	Check entity wording and third-party descriptions
Is the brand recommended?	Recommended, listed, rejected, or unclear	Improve use-case differentiation and comparisons
Which competitor appears instead?	Competitor and prompt intent	Inspect the sources and claims supporting it
Did the answer create value?	Referral and business outcome	Connect analytics with CRM evidence and state uncertainty

Weight the rows according to the business decision. A founder seeking qualified signups should not value a citation and a conversion equally. An agency investigating description errors may care more about accuracy than traffic.

How PromptScout Makes This Repeatable

For one question, you can inspect the answer manually. A useful audit repeats the process across prompt groups and providers.

In PromptScout, keep citations, tracked-brand mentions, competitor mentions, and provider results as separate fields. Review the answer before marking a recommendation. Group findings by buyer intent, record one next action for each repeated gap, then run the same prompt set in the next monitoring cycle.

That creates a practical chain: prompt, cited source, brand outcome, competitor outcome, recommended fix, and follow-up observation.

How to Verify a Change

Keep the prompt group, provider scope, and classification rules consistent. Repeat the observation because AI answers vary. GEO measurement research recommends treating visibility as a distribution instead of a snapshot. Read “Don’t Measure Once” on arXiv.

Report movement by layer: citation, brand appearance, recommendation context, referral, and conversion.

Monitoring can show that outcomes changed together. It cannot establish that one page edit caused the recommendation.

Notes on the data

This report uses anonymized monitoring data from July 10 to August 9, 2026. The sample contains 86 buyer-style prompts, 1,267 successful AI answers, and 6,454 captured source citations across Bing Copilot, Gemini, Google AI Overviews, OpenAI, and Perplexity. Brand mentions, competitor mentions, and source citations were classified separately. The analysis is observational. Prompt coverage differed by provider, recommendation status was not automatically classified, and citation capture may be incomplete.

PromptScout supports this workflow: monitor the questions, inspect the sources, separate mentions from recommendations, and check the next change in a later cycle.