Loading this page…
Loading this page…
Methodology
AI answers are probabilistic. Any product that hands you a stable rank position is telling you something the underlying system cannot support. This is what Sivlo does instead, in full, including the parts that make our numbers less impressive.
We crawl a handful of pages and construct the market around your company: category, products and services, buyer types and roles, geographies, languages, decision dimensions and competitors. There is no prebuilt category library behind it, and every field records whether it was inferred from your website, supplied by you, or generated from a fixture.
Not keywords, and not a pile of prompts. A buying situation is a real circumstance: a stage, a buyer, a use case, a constraint, a comparison. The questions underneath it are measurement variants of that situation, never its identity.
We favour strategically important intents and realistic question variations, while retaining a smaller number of identical repeats in order to observe the model’s own randomness.
We record which companies were put forward, in what order, with what tone, and which sources the engine used to get there.
Every rate retains its numerator, denominator and eligible observations. Failed provider calls are operational missingness and never become brand absences. There is no 0-to-100 Sivlo Score: a composite invented from weighted parts hides the sample sizes that make the parts mean anything.
Your buying situations measure every competitor named in the same answers, and all of their outcomes are retained. The company is the paying project; the market is the unit of intelligence.
Recommendation Presence includes relevant, successful recommendation-oriented observations. Mention Rate includes all relevant, successful observations. Citation Presence includes only observations where the measured engine or surface exposed source data. Every aggregate can be filtered by the engines, surfaces, markets, languages and buying situations that were actually measured.
Suitable rates include a Wilson 95% interval and a sample size. The dashboard leads with the plain percentage; the interval remains available as context rather than forcing statistical language into the headline.
Every diagnostic insight carries exactly one of three classes, and the product shows them differently on purpose. A class is never raised by assertion: it is decided by counting, and an insight that misses a threshold keeps the weaker class along with the reason it missed.
A relationship is only called Associated when it holds across at least 3 brands, 2 buying situations and 2 measured engines or surfaces, on at least 12 eligible observations, with a difference of at least 10 points. Association is a property of the measured market. It is not causation, and Sivlo does not claim it is.
Recurring measurement is only worth paying for if a difference between two runs means something. Two measurements are comparable only when they share a market-model version, a buying-situation set, a sampling strategy and a provider configuration. When any of those changed, Sivlo says the runs are not comparable instead of drawing a line between them.
To estimate how much a rate moves on its own, we repeat identical observations under one fixed configuration. With at least 3 repeats we derive a noise threshold from the variation we actually measured, never below 3 points. Movement inside that band is reported as indistinguishable from noise. Without a calibration run we fall back to non-overlapping 95% intervals, which is weaker and is labelled as such.
Every comparison ends in one of five verdicts, and only two of them are news: a meaningful increase, a meaningful decrease, movement within expected variation, insufficient evidence, or an incomparable configuration. Below 5 eligible observations on either side we report a measurement gap rather than a result — a rate built on a handful of answers swings on one of them. Competitors and buying-situation clusters are judged by the same rules the company is, so a small or thinly sampled competitor movement raises no alert.
Sivlo monitors automatically according to the subscribed plan; monthly is a historical cadence label. History only reports movement when comparable measurements provide sufficient evidence.
You can mark a recommendation done, and later measurements are displayed after that point in time. That is a sequence, not a cause, and Sivlo will not describe it as one.
The list below costs us sales. It stays because every item on it is something a competitor is willing to imply and we are not.
Official model APIs and consumer search surfaces are different measurement types. Gemini API is never labelled Google AI Overview or Google AI Mode. A surface appears only when Sivlo actually measured that surface.
Nobody can. What we can do is change what the engines have to work with, measure what happened, and tell you honestly when a change did nothing.
We show the interval and the sample size rather than rounding uncertainty away.
Internal intelligence used for classification, summaries or recommendations is never counted as a customer-facing AI visibility observation.
While no real answer-engine provider is connected, no output is presented as a real observation of a real model, anywhere it appears.
Sivlo does not publish to your website, generate content for it, connect to your CMS, or predict how much a change will lift your numbers.
What works is being genuinely findable, clearly described, and corroborated by sources other than yourself. Everything Sivlo ships is built on that.
If you disagree
A methodology nobody can check is a marketing claim. If a threshold looks wrong for your market, or a denominator excludes something you think it should count, tell us which one and why.