Sources, Citations, Mentions & Recommendations: What They Measure and Where They Come From
In the AI visibility category, the terms source, citation, mention, and recommendation are often used loosely or interchangeably. They are not the same thing. Each describes a different kind of visibility, produced at a different stage of the answer-generation process.
QueryBurst makes each of these stages inspectable. This page explains what each term actually measures, why the distinction matters, and why different failures at different stages require different fixes.
These terms are not interchangeable
In AI systems, a source, a citation, a mention, and a recommendation each describe a different kind of visibility at a different stage of the answer-generation process.
A page can be retrieved as a source without being cited.
A page can be cited without its brand being mentioned in the answer.
A brand can be mentioned without being recommended.
And a brand can be recommended without the sources that influenced that recommendation receiving visible citations.
These are different outcomes. They are produced at different stages. And when one fails, the fix depends on which stage failed — not simply on whether the brand appeared in the final output.
Most platforms in the AI visibility category report some combination of these metrics. Few distinguish between them.
A source is a page that was retrieved, not necessarily a page that influenced the answer
AI systems do not always retrieve external information when generating an answer. Some responses are produced entirely from the model's existing knowledge — which is itself a form of visibility that QueryBurst makes inspectable through its pre-retrieval prior. When retrieval does occur, it may involve live web searches, internal indexes, cached content, or a combination.
In practice, most retrieval-augmented responses involve queries that reach the web. QueryBurst models this retrieval stage because it is the part of the process that can be directly investigated. Internal indexing is not directly observable, though even an internal index is ultimately built from web content — so the information that exists on the web still matters, even when the system does not fetch it live at query time.
Being retrieved is the first stage of visibility in the answer-generation process. But retrieval alone does not mean a page influenced the answer.
A retrieved page may be:
- scored against the system's information need and discarded
- included in a shortlist but outcompeted by stronger evidence
- used to verify a claim the system already held
- used as primary evidence that shaped the answer
Some platforms in the AI visibility category report any URL that appeared in the system's search results as a "source." That conflates retrieval with influence. A page that was fetched during research is not necessarily a page that shaped the answer.
The distinction matters because the problems are different. A page that was never retrieved has a discoverability or relevance problem — the system's research queries did not find it. A page that was retrieved but not used has an evidence quality problem — the system found it but its content was not compelling enough to survive selection.
QueryBurst makes this distinction inspectable. Within a QueryBurst run, users can see which pages were retrieved, which survived selection, which were used as evidence, and which were discarded — and at which stage.
A citation is a visible reference, not evidence of influence
A citation is a visible link or attribution that appears in the final AI response. It is the metric most frequently treated as evidence of influence.
It is not reliable evidence of influence.
A page can be cited as a formality — referenced in a footnote or source list while the substantive content of the answer came from elsewhere. In these cases the citation is real but the influence is minimal.
A page can also deeply influence an answer without receiving a visible citation. Its claims, evidence, or framing may have shaped the system's reasoning during the research phase without being attributed in the final output.
Citation counting — treating each visible URL reference as equivalent evidence of influence — loses both of these distinctions.
QueryBurst treats citation as one stage in a longer chain: retrieval, selection, evidence extraction, grounding, and citation. A page that is cited can be traced back through those stages to understand what role it actually played. A page that is not cited can be investigated to determine whether it was retrieved, whether its evidence was used, and where it dropped out.
A mention is a brand name in the response, not a measure of how it was described
A mention is a brand name appearing in the text of an AI response. This is the unit that AI Share of Voice is built on.
But a mention carries no direction on its own. A brand can be mentioned as the lead recommendation, as one option in a list, as a comparison point, as an also-ran, or as a warning.
- "We'd recommend HubSpot for your use case" — mention as recommendation
- "HubSpot is one of several options" — mention as inclusion
- "Unlike HubSpot, Pipedrive offers more flexible cancellation" — mention as unfavourable comparison
- "We'd avoid HubSpot for teams of your size" — mention as warning
All four count equally in mention-based metrics. They are not the same visibility outcome.
In multi-turn conversations the problem compounds. A brand established by name in an earlier turn may be referred to as "their Growth plan" or "the platform we discussed" in later turns — the brand is functionally present but no longer mentioned by name. Conversely, a brand may be named repeatedly as a point of comparison without ever being the recommendation.
Mention counting treats presence as the unit of measurement. The context of the mention — how the brand was described, what role it played in the answer, and whether the user was being directed toward or away from it — is where the actual visibility outcome lives.
A recommendation is the outcome that matters — the question is where the path to it succeeded or failed
Being recommended is the outcome a business cares about. When a customer asks an AI system for advice and the system recommends your product, that is the point at which AI visibility produces a commercial result.
The distinction QueryBurst draws is not that recommendations don't matter. They are the whole point.
The distinction is between observing that a recommendation happened and diagnosing why it did or did not happen.
A prompt tracking platform can report: "Your brand was recommended in 12 of 50 tracked responses." That observation is accurate within the panel.
It cannot generally explain why the brand was recommended in those 12, or why it was absent from the other 38. The path from the model's prior knowledge, through research, retrieval, evidence selection, and grounding to the final recommendation is not observable from the collected output.
QueryBurst makes that path inspectable. Within a QueryBurst run, users can investigate:
- Whether the brand was already known to the model before retrieval began
- Whether the brand entered during the research phase
- Whether its evidence survived selection and filtering
- Whether its content influenced the grounded answer
- Where in the chain it dropped out, if it dropped out
When a brand is not present at any stage — not known, not retrieved, not considered — that itself is a diagnosis. It means the model had no prior association with the brand in this context and the research process did not surface it. That is a different problem from a brand that was retrieved and lost on evidence quality, or one that survived to grounding but was not recommended in the final answer.
A prompt tracker can express the absence. QueryBurst can diagnose it.
Different failures at different stages require different fixes
The reason these distinctions matter is not taxonomic. It is practical.
A brand that is never retrieved has a discoverability problem. The system's research queries — shaped by the model's prior knowledge and the user's conversation context — did not produce searches that found the brand's content. The fix involves the information the brand publishes and how it relates to the questions AI systems generate during research.
A brand that is retrieved but not used has an evidence quality problem. The system found the page but its content was not strong enough, relevant enough, or well-structured enough to survive selection. The fix involves what the page says and how clearly it says it.
A brand that is used as evidence but not cited has an attribution problem. Its content influenced the answer but the system did not surface a visible reference. This may be a function of how the system handles citations rather than a weakness in the brand's content.
A brand that is cited but not recommended may have a positioning problem. Its evidence was strong enough to be referenced but the system's overall assessment — shaped by all the evidence it collected, including from competitors — favoured a different recommendation.
A brand that is not known to the model at all has a prior knowledge problem. The model has no existing association between the brand and the relevant category, problem, or use case. The fix may involve the broader information environment rather than any single page.
Collapsing all of these into "appeared or did not appear" loses the information needed to decide what to do.