Every number in RavenBI carries a label. We separate what we saw from what we calculated, and we never present one as the other.
This page is not a disclaimer. It is how the product works.
Observed — recorded directly from a source. An advertisement was live on a given date. A product had a given price. A claim appeared in a given text. If we say we observed it, we saw it.
Reported — provided by a third party that we did not verify independently. Platform-declared reach falls here.
Estimated — produced by a RavenBI calculation from observed inputs. Advertising spend is estimated from reach and modelled CPM. Estimates are ranges, not figures, and they carry their assumptions.
Derived — an indicator RavenBI computes from several inputs. Winner, Momentum, Opportunity and Threat scores are derived. They are proprietary indicators, not industry standards, and they mean what this methodology says they mean.
AI interpretation — a language model read a text and classified it. Where available, RavenBI exposes the evidence behind the classification: the exact sentence from the source. Our advertising-argument classifications carry this evidence today, and we are extending it to the remaining classifiers.
1. We distinguish facts from estimates. Every figure is labelled. An estimate never appears with the visual weight of an observation.
2. No percentage without its base. A percentage without the absolute number behind it is a rhetorical device, not a measurement. RavenBI shows both, and suppresses percentages where the sample is too small to support them.
3. No conclusion without sufficient evidence. Below a minimum sample size, RavenBI shows direction and confidence instead of a precise figure. We would rather say "rising, low confidence" than invent a number.
4. Missing data stays missing. When a source does not tell us something, we leave it empty. We do not fill gaps with plausible defaults, and we do not infer a value from a product category or a brand name.
5. Every signal has provenance. Each data point records where it came from, when it was observed, what transformation was applied, and which classifier version produced it. If a figure is ever challenged, we can reconstruct exactly how it was produced.
RavenBI reports what it observes, labels what it estimates, and identifies what it infers.
RavenBI combines information from publicly accessible advertising transparency libraries operated by the platforms themselves, licensed third-party data providers, publicly accessible brand websites and marketplaces, and official regulatory sources.
On top of that raw material we build our own layer: entity resolution between brands, products, advertisements, landing pages and ingredients; classification against controlled vocabularies; historical snapshots; and derived indicators.
That second layer is what makes RavenBI more than a search box over public data — and it is also where our estimates and interpretations live, which is why we label them.
We do not claim complete coverage. RavenBI observes what its sources expose. The absence of an advertisement, brand or product in RavenBI does not establish that it does not exist.
We do not claim that estimates are audited figures. Advertising spend estimates are models, not advertiser-reported expenditure.
We do not claim that our classifications are infallible. They are produced by language models, measured against a human-arbitrated reference set, and improved when measurement says they need improving.
We maintain a reference set: several hundred advertisements, labelled independently by multiple evaluators who cannot see each other's answers or the production system's answer. Disagreements are resolved by a human.
Every change to a prompt, a model or a taxonomy is measured against that reference set before it reaches customers. We track precision, recall and — most importantly — how often the classifier correctly stays silent when a text simply does not contain the information.
That last measure matters more than it sounds. A classifier that always produces an answer looks confident and is frequently wrong. We would rather show you a gap than a guess.
If you believe a figure or a classification in RavenBI is wrong, tell us. Every brand, advertisement, product and ingredient page carries a reporting link, and we check reports against the original source.
Correcting an error is part of the method, not an exception to it.