An AI answer can name your brand without linking to your website.
It can cite your documentation without recommending your product.
It can recommend your product while using an independent review as the supporting source.
All three are forms of visibility, but they do not mean the same thing. Combining them into one score can hide whether people are seeing your name, your website or an actual reason to choose you.
Quick answer
Track mentions, citations and recommendations separately.
- Mention: The answer names the brand, product or company.
- Citation: The answer links to or identifies a source associated with a claim.
- Recommendation: The answer presents the brand as a suitable option for the user's need.
Then add context: prompt intent, position, sentiment, stated reason, cited domain, competitors, provider and observation date.
A mention shows presence. A citation shows source use or attribution in that response. A recommendation shows fit within a decision. None guarantees a click, sale or future appearance.
What is an AI brand mention?
A brand mention occurs when the generated answer includes a recognized brand, company or product name.
Examples:
- “Prerender Buddy can compare crawler-facing and rendered content.”
- “Other options include Company A, Company B and Prerender Buddy.”
- “Prerender Buddy is not necessary if the raw HTML already contains the important content.”
All are mentions, but their value differs.
The first describes a capability. The second is a list appearance. The third is a qualified statement that may improve trust even though it is not a simple endorsement.
For every mention, record:
- exact name detected
- sentence or paragraph context
- prompt and intent
- whether the brand is central or incidental
- positive, neutral, mixed or negative stance
- position within a list, when relevant
- competitors appearing in the same answer
Mention count alone cannot show how the brand is framed.
What is an AI citation?
A citation is a visible reference to a source used or associated with the answer.
The source may be:
- your official website
- competitor website
- independent publication
- review or comparison site
- documentation or repository
- customer, partner or marketplace page
- forum or community discussion
OpenAI's ChatGPT search documentation says web-search responses may include inline citations and a Sources panel. It also advises opening the sources because citations and results can be incomplete, outdated or incorrect.
That warning matters for monitoring. A tracker should preserve the URL and claim context, but a person should verify that the source actually supports the answer.
What is an AI recommendation?
A recommendation occurs when the answer presents a brand as an option that fits the user's request.
Examples include:
- “For a small agency, consider…”
- “The best fit if you need…”
- “Use this option when…”
- “A suitable alternative is…”
A recommendation should be tied to the prompt's constraints. A product recommended for an enterprise team is not automatically visible for a solo founder's question.
Record:
- recommendation language
- use case or audience
- reason given
- qualifications or warnings
- rank or grouping where visible
- supporting sources
This turns “we appeared” into “we were associated with this use case and reason.”
One answer can contain any combination
| Mention | Official citation | Recommendation | Interpretation |
|---|---|---|---|
| Yes | No | No | Brand awareness or incidental reference |
| No | Yes | No | Official page may support a claim without the brand being named clearly |
| Yes | Yes | No | Brand and owned source are visible, but no selection is made |
| Yes | No | Yes | Brand is recommended, possibly through general or third-party information |
| Yes | Yes | Yes | Brand, owned source and prompt fit appear together |
| No | No | No | No observed visibility in this answer |
An independent citation should also be tracked separately from an owned-domain citation. The brand can benefit from external corroboration even when its website is not the source shown.
Why a single visibility score is not enough
A composite score is useful for scanning trends, but it hides the underlying event.
Imagine two brands both score 60:
- Brand A is named often in long lists but is rarely recommended or cited.
- Brand B appears less often but is repeatedly recommended for the target use case and supported by its documentation.
The same score describes two different positions and two different actions.
Keep the raw components available. Use the summary score as a navigation aid, not as the evidence itself.
The core metrics
Prompt coverage
The percentage of monitored prompts with at least one valid observation during the reporting window.
This is a data-quality metric. If only half the planned prompts ran successfully, the visibility report should not imply complete coverage.
Mention rate
The percentage of valid answers in which the brand is mentioned.
Mention rate = answers mentioning the brand ÷ valid answers × 100
Report the numerator and denominator. “40% mention rate” is more useful as “8 mentions across 20 valid answers.”
Recommendation rate
The percentage of valid answers in which the brand is presented as a suitable option for the prompt.
This requires context classification. A negative warning or historical reference is not a recommendation.
Citation rate
The percentage of valid answers that cite at least one URL from the brand's owned domains.
Also track external sources that support a brand mention. These are important evidence even though they do not increase owned-domain citation rate.
Share of mentions
The brand's appearances relative to all tracked competitor appearances within the defined prompt set.
Always state which competitors and prompts are included. This is not universal market share.
Source coverage
The distinct domains and URLs appearing across the monitored answers, with frequency and prompt association.
This reveals whether visibility depends on one source, a group of independent references or the official website.
Context distribution
The proportion of mentions that are positive, neutral, mixed or negative, plus the use cases associated with the brand.
Automated sentiment can help group observations, but review commercially important or ambiguous cases manually.
Track the denominator carefully
Visibility percentages become misleading when failed or irrelevant runs are silently included.
Separate:
- valid answer
- refusal or no answer
- technical failure
- answer without web search when sources are required
- off-topic result
- duplicate or retried run
Define the denominator before calculating the metric. If citation tracking applies only to web-search answers, do not describe absence of citations in a non-search response as an owned-domain failure without qualification.
Branded and non-branded prompts answer different questions
Branded prompts
Examples:
- “What is Prerender Buddy?”
- “Does Prerender Buddy work with Webflow?”
- “What are the limitations of Prerender Buddy?”
These test whether the system can identify and describe the brand accurately.
Non-branded discovery prompts
Examples:
- “How can I check whether AI crawlers can read my JavaScript site?”
- “What tool monitors stale crawler-facing pages?”
- “How can a small agency track brand visibility in ChatGPT?”
These test whether the brand appears when a user has a problem but does not name a provider.
Do not combine both categories into one percentage without showing the split. A high branded mention rate can coexist with no category discovery.
Separate prompts by intent
Group prompts by the decision they represent:
- educational: “What is prerendering?”
- diagnostic: “Why can crawlers not see my page?”
- solution discovery: “What tools solve JavaScript crawler visibility?”
- comparison: “Prerender Buddy vs another option”
- purchase: “Which plan fits ten sites?”
- support: “How do I configure a CNAME?”
A recommendation on a solution-discovery prompt has different commercial meaning from a mention in a definition.
The reporting unit should therefore be the prompt cluster and intent—not a single global number.
Track platforms and modes separately
Do not average ChatGPT, Gemini, Perplexity, Claude and other providers into a number that conceals their differences.
Record:
- provider
- model or mode when visible
- web search enabled or observed
- region or language where relevant
- account or personalization conditions when controlled
- clean or continuing conversation
- date and time
A platform comparison can be useful, but each platform's sample size and method should remain visible.
Repeated runs and prompt variations
Generated answers can change between runs and with small wording changes.
Use two separate tests:
- Repeat stability: Run the same prompt under the same defined conditions.
- Intent variation: Run a small set of natural phrasings that preserve or deliberately change the intent.
Do not hide this variation by reporting the most favorable answer. Preserve every valid observation and aggregate over a defined window.
For small teams, a compact, stable prompt set is better than hundreds of synthetic questions no customer would ask.
Add referral traffic, but keep it separate
Visibility is an upstream observation. Referral traffic is a downstream behavior.
OpenAI's publisher FAQ says ChatGPT referral URLs include utm_source=chatgpt.com, which can be tracked in analytics.
Track:
- visits from AI referral sources
- landing pages
- qualified actions or conversions
- engagement and bounce behavior where appropriate
Do not turn “no tracked referral” into “no AI influence.” A person may see a brand recommendation and later search for the brand directly. Conversely, a citation does not guarantee a click.
A practical reporting table
| Prompt cluster | Valid answers | Mention rate | Recommendation rate | Owned citation rate | Main competitors | Common source domains |
|---|---|---|---|---|---|---|
| Crawler diagnosis | 20 | 35% | 15% | 10% | A, B | official docs, review site |
| Monitoring | 18 | 22% | 11% | 6% | B, C | comparison sites, forums |
| AI visibility | 20 | 40% | 20% | 15% | C, D | vendor blogs, publications |
Under the table, include:
- method and reporting window
- exact prompt list or link to it
- provider split
- failed-run count
- important context changes
- source and competitor observations
- recommended actions
This makes the report auditable instead of presenting a number without its evidence.
How to interpret common patterns
High mentions, low recommendations
The brand is known or included in lists, but the answers may not connect it to a specific use case.
Investigate positioning, audience clarity, product evidence and comparison context.
High recommendations, low owned citations
The brand is selected, but third-party sources may be carrying the evidence.
Inspect those sources and verify that official product or documentation pages clearly support the same facts. Do not try to replace legitimate external corroboration with self-published claims.
Owned citations, low brand mentions
Your content may support a general answer without the entity being clearly associated with it, or the citation may concern a narrow fact.
Inspect the cited passage, brand clarity, page title and organization context. Do not assume that adding more brand repetitions will solve it.
Negative mentions
Visibility is not automatically favorable. Check whether the criticism is accurate, outdated or unsupported. Correct owned information and legitimate external profiles where possible; do not attempt to suppress truthful criticism.
No visibility and weak crawler access
Fix the technical eligibility problem first. Then re-establish the baseline before interpreting the content gap.
What not to put in an AI visibility report
Avoid:
- a single score with no raw observations
- percentages without denominators
- combining branded and non-branded prompts
- mixing providers without showing the split
- counting every mention as a recommendation
- calling every cited URL an endorsement
- reporting only favorable reruns
- treating correlation as proof of why an answer changed
- promising future mentions or citations
The report should make uncertainty visible enough to guide a decision.
Who this is for
- SaaS founders with already-shipped JavaScript websites
- React, Vite, Vue, Lovable, Bolt, or Base44 users
- SEO freelancers checking crawler-readable HTML
- Agencies maintaining client sites without rebuilding them
Where Prerender Buddy fits
Prerender Buddy's AI Visibility layer should preserve the answer-level evidence behind the dashboard:
- prompt and intent group
- provider and observation time
- brand and competitor mentions
- recommendation context
- owned and external citations
- source URLs and domains
- repeat-run history
- technical health of the mapped owned page
That information can feed the Improve layer. A high mention rate with low recommendation context creates a different brief from a crawler-access failure or a third-party citation gap.
The wider workflow remains:
- Make important content technically accessible where needed.
- Monitor whether the pages remain healthy and current.
- Measure mentions, citations, recommendations and competitors.
- Recommend a specific improvement based on the observed pattern.
What these metrics cannot prove
AI visibility metrics do not prove:
- a provider's private selection logic
- that a cited source caused a mention
- that a mention produced a visit
- that a citation produced revenue
- that the next run will repeat the answer
- that a page update caused a later visibility change
They are structured observations. Their value comes from consistent collection, preserved evidence and careful decisions—not certainty they cannot provide.
You may not need Prerender Buddy if
- Server HTML is already complete.
- Static pages crawl correctly.
- You are already rebuilding with SSR or static generation.
- You only need an audit, not a rendering fix.
Final recap
Mentions, citations and recommendations describe different outcomes.
Track each separately, add context and preserve the answer and source evidence. Split branded from non-branded prompts, group by intent, show denominators and keep provider differences visible. Use referral traffic as a related downstream metric, not as a substitute for visibility.
A useful AI visibility report should answer more than “Are we visible?” It should show where, how, why the appearance matters and what can be investigated next.
When the metrics reveal a meaningful gap, turn the evidence into a content brief, then decide whether to update an existing page or publish a new one.
Want a visibility baseline that separates names, sources and recommendations? Start a Prerender Buddy project and track the evidence behind each answer.
Frequently asked questions
Is a ChatGPT mention the same as a citation?
No. A mention names a brand or product. A citation links to or identifies a source. An answer can contain one without the other.
What counts as an AI recommendation?
Count an answer as a recommendation when it presents the brand as a suitable option for the prompt's need. Preserve the stated reason and any qualification. A list appearance or negative warning is not automatically a recommendation.
Can my website be cited without my brand being mentioned?
Yes. A page may support a fact while the brand is absent or not clearly named in the answer. Track the cited URL and the claim it supports.
Is every brand mention positive?
No. Mentions can be favorable, neutral, mixed, negative or incidental. Store the surrounding context and review important cases manually.
What is AI citation rate?
For an owned-domain citation rate, divide valid answers citing at least one owned URL by the valid answers in the defined prompt group. State the denominator, provider and reporting window.
Should I use one AI visibility score?
A summary score can help scan changes, but it should not replace mention, recommendation, citation, source and competitor data. Keep the components and observations accessible.
Does higher AI visibility guarantee more traffic?
No. Some citations generate clicks, some mentions influence later branded searches and some appearances produce no measurable action. Track referrals and business outcomes separately.