Brand Rank
- Recommendation lists are ordered surfaces — first and ninth are different commercial events.
- Top-3 share on valid lists is the operative shortlist KPI, not average rank.
- Read rank against list length and check for copied rankings before inferring model preference.
- As answers compress, rank loss converts into presence loss — the distribution warns first.
Consensus definition
Position has been part of generative-engine measurement from the start: the founding GEO benchmark weights visibility by where in the answer a source appears, not merely whether it appears.1 Order is also not an accident of style — language models themselves exhibit position effects, weighting earlier-listed items differently from later ones,2 and answer formats that present options as ranked or curated lists transfer that ordering directly to the reader. Brand Rank applies the lens at brand level: within answers that enumerate providers or products, record the brand's list position and derive thresholds — first mention, Top-3 presence as the shortlist cut, Top-5 as the longlist.
rhinegold operator refinement
Rhinegold reads rank distributions, not rank averages. A brand that is sometimes first and sometimes absent is in a different situation from a brand that is reliably fifth — the mean hides exactly the difference that matters. Three disciplines apply. First, rank only counts within valid lists: answers that actually enumerate alternatives, not every answer that names brands. Second, rank must be read against list length — third of three and third of twelve are different events. Third, watch for ordering artifacts: some answers sort alphabetically or copy a retrieved ranking verbatim, which says more about the grounding source than about the model's preference.
Operational use
Track the Top-3 share — the fraction of valid list answers in which the brand appears among the first three — as the primary rank KPI on a stable prompt set. Pair it with Brand Recommendation Share: BRS tells you how exclusive the recommendation is, rank tells you where in the order you sit when it is not exclusive. Movement in rank often precedes movement in Mention Rate — being pushed from Top-3 into the tail is visible before being dropped entirely.
Measurement boundary
Rank exists only where lists exist: discovery and comparison prompts produce them, informational prompts mostly do not — so rank coverage depends on the prompt set (Discovery Prompt design). Cross-provider rank comparison is weak because list length and list frequency differ by platform (Platform Divergence). And rank carries no direction: a first-position mention can still be a Negated Mention.
Distinct from
Against classical SERP position — superficially the same idea, but SERP rank is a deterministic index over documents, while Brand Rank is a distribution over stochastic, regenerated answers. Against Brand Recommendation Share, which weights by how many competitors share the answer, not by order. Against Mention Intensity, which measures depth of treatment rather than position in an enumeration.
Operational note
Common mistakes
- Averaging rank across answers with different list lengths — normalize by list length or report threshold shares instead.
- Counting rank in answers that merely name brands in prose — rank is only defined where the answer enumerates alternatives.
- Reading a stable first position as model preference when the answer copies a retrieved ranking verbatim — check the grounding sources before celebrating.
Where consensus is missing
No standard exists for what constitutes a valid list, how to treat ties and groupings ("top picks" vs numbered items), or how to normalize across list lengths. Position-bias research establishes that order matters inside models, but no published study yet quantifies how AI list position converts to buyer consideration.
Sources & deeper reading
- rhinegold list-presence methodology — valid-list detection, Top-3/Top-5 thresholds and rank distributions across providers, shared with engaged clients and partners.