Loading
Published in full — a score nobody can check is a score nobody should trust
An answer naming a competitor but not the tracked brand still counts in the denominator — being absent is a result. An answer naming no brand at all is excluded, because nobody lost it.
What counts as an answer about this category — the rule is published because it changes scores
A category restricts three things: which brands are eligible, which market the answers come from, and which topics count. The third matters most. Without it a brand's whole footprint in a market is credited to every category it is listed under — measured directly, HP had 9,861 appearances under Mobile Phone Headsets, and its most common topics there were printer-setup questions. Topics are matched to a category using published German keyword rules that read the topic text and know nothing about brands, which is the property that makes them usable: any rule that infers a category from which brands appear in it will keep the very generalists it needs to exclude.
Categories without a scope
Some categories have no topic vocabulary that fits yet — there are no pet, drink or tyre topics in the data. Those are left unscoped rather than mapped to a near-miss, and a category with no topics matches no answers: every brand returns zero observations and the category is marked low data. It is listed rather than hidden, because "we have not asked about this yet" and "this category does not exist" are different statements. There is deliberately no fallback to a broader scope — a number built from a neighbouring category would be wrong exactly where the reader most needs to be told there is nothing here.
Comparability
The scope applies to the numerator and the denominator alike, so reach stays a share of the answers that were actually in scope. Because this changes almost every category score, reports carry a scoring version and only reports on the same version are comparable.
The prompt library is much broader than this tree. Of 128,266 answers, 46,186 — 36% — are on a topic that can be placed in a category at all; the rest are about products the taxonomy does not cover. This is a consequence of how prompts were selected: a brand was chosen because it sells something in a target category, and most of that brand's prompts are about its main business. Only in-scope answers are scored, in the numerator and the denominator alike, so a brand's strength in adjacent products never leaks into a category it is weak in. bosch is the clearest case: 14,590 answers across the campaign, 6,676 of them on a topic this tree can place — the other 7,914 are about its other businesses and count toward no category's score.
The headline number
Reach = answers naming the brand
÷ all answers in scope
Prominence = mean attention, where present
Quality = mean quality gate, where present
Score = 100 × ( R^0.4 × P^0.3 × Q^0.3 )^1Quality gate
One verifiable sub-gate today: whether the brand's own domain was among the sources the model used. 1.0 if yes, 0.6 if no. Spec correctness and market availability need retailer and catalogue data this export does not carry.
λ = 1
Presentation only — mathematically cannot change ranking order. Published and frozen because it is the most gameable constant here.
1 / log₂(rank + 1) — the standard information-retrieval discount, borrowed not invented
Every answer naming two or more brands becomes a set of matches
P(a > b) = e^θa / (e^θa + e^θb) θ fitted by MM iteration to reproduce the observed record, normalised Σθ = 0. Strength = 1500 + 400θ / ln(10)
No chosen weights
Nothing here is tuned — the rating is whatever best explains the match record. That is why the fit check is published beside it on the Head to head page.
Half-win prior
Each observed pair carries a half-win each way. Without it a brand that never loses has no finite rating and the fit silently degenerates.
Bootstrapped over 300 resamples, seeded so results are reproducible
Two brands whose intervals overlap are not meaningfully different, and the overview says so in words rather than presenting a ranking the data does not support.
A dash is more honest than a precise-looking number built on nothing
Below this the score shows — instead of a number.
Below this the score shows — instead of a number.
Quality is half-built
Two of four intended gates — spec correctness and market availability — need retailer and catalogue data not in these exports.
No time dimension
The primary export carries no date, so this is a snapshot rather than a trend.
Rank inferred on the model source
Reliable for ranked lists, rougher for flowing prose.
Short brands need an allowlist
LG, HP, 3M and GE are two characters. A length filter silently deleted them during development, so the allowlist is explicit and visible.
Entity list is not exhaustive
Retailers and marketplaces are separated from brands by an explicit list — auditable, but incomplete.
Subcategories are inferred
The data source does not supply them. Keyword rules and a model decide, both are published, and neither is perfect — treat a single surprising subcategory placement as something to check rather than a finding.
Not externally validated
Reproducible and defensible. Whether the scores predict commercial outcomes is still open.
Ingested 2026-10-06T14:45:35Z · Haushaltselektronik-results-2026-09-21-under-1gb