Reference rate
The share of completed AI answer runs, over a defined prompt set, in which a brand or its domain appeared — an observed frequency with a denominator, not a score and not a prediction.
Karl-Gustav Kallasmaa, Founder & CEOLast updated A reference rate is the share of completed AI answer runs, over a prompt set someone defined, in which a brand or its domain appeared in the answer. It is an observed frequency with a stated denominator. It is not a score, not a rank, and not a prediction of future mentions.
The definition is only useful if the denominator travels with it. A bare percentage is not a claim anyone can check; "named in twelve of fifty successful runs across two models over the last seven days" is. Everything else on this page follows from insisting on that second form.
The denominator, precisely
In Attensira the denominator is successful runs in the window: the rate is the number of those runs whose stored answer text named the brand or its domain, divided by all of them, reported as {value, n}. Failed runs are stored but excluded from n, so a day with collection problems yields a smaller sample rather than a distorted rate.
Two values that look adjacent are different facts, and a chart that draws them the same way is lying:
value: null— not measured. Nothing was observed, so nothing can be reported. A model markedtracked: falsewas never queried; rendering that as a measured zero invents a rejection that never happened.value: 0withn > 0— measured zero. Runs completed and none of them named the brand.
Why rates do not sum, and are not a share
Competitor reference rates use the same denominator as yours. They are independent rates, not slices of a fixed total. A single answer that names you and three competitors raises four rates from one run, which is why adding them can exceed the whole and tells you nothing.
This matters for a field name people misread. Attensira's ShareOfVoice is the brand's own mention rate under an older name — the same number, not a share of anything. Nor is AvgPosition a rank of your brand: it is the mean index of a URL within a model's own citation list, computed per source domain, answering "when this domain is cited, how far down the list does it sit". Attensira computes no visibility score and no sentiment metric, and no predicted probability of being referenced.
How the underlying detection actually works
A mention is a case-insensitive substring match of the brand name or domain against the answer text. There is no entity resolution and no word-boundary check. A brand called Arc matches "march" and "search"; a brand called Notion matches "notional". That error is systematic rather than random, so averaging does not remove it, and every competitor rate compared against yours inherits the same method — unfairly in favour of whoever has the more common name.
The mitigation is that verbatim answer text is stored for every successful run. The rate is a derived number; the answers are the evidence. If a brand name is a common English word, reading a handful of stored answers tells you within minutes whether the number measures the brand or the dictionary.
Reading a movement honestly
Models are non-deterministic: the same prompt asked twice can return different answers naming different brands. Sampling is the only defence, and Attensira documents one draw per prompt per model per reading on Starter and three on Growth and Business. At n=1, most week-to-week wobble sits under the noise floor.
Gating is the second defence. Every delta is tested with a two-proportion z-test at 95% significance, and a movement that fails returns {real: false, value: null} — meaning the change could not be proven, never that it was zero. A tool that instead prints every wobble as a trend is selling a story the sample does not support.
Failure modes
- Quoting the percentage without `n`. At three draws, one hit and two hits are adjacent outcomes, and a single different answer moves between them.
- Treating today as a full day. Each prompt × model × country combination runs at most once per workspace-local day, so today is partial until it closes. Compare whole days to whole days.
- Reading a rate as a share of a market. It is not; see above.
- Calling a rate a probability. It describes runs that happened on prompts you chose. See citation probability.
- Comparing rates measured on different prompt sets. Editing prompts mid-window changes what the denominator means, and no downstream test repairs that.
Frequently asked questions
What is the denominator?
Successful runs in the window, reported with the rate as {value, n}.
Is a measured zero the same as null?
No. Null means not measured; a zero with n greater than zero means runs happened and none named the brand.
Do competitor rates sum to the whole?
No. They share a denominator and are independent, so one answer can raise several at once.
Does it predict future mentions?
No. It is an observed frequency, and Attensira computes no predicted probability, visibility score or sentiment metric.
Terms related to Reference rate
The informal name for the chance an AI answer cites a page. No tool computes it; it is only estimated from an observed citation rate over a fixed prompt set.
Occurrences of a brand name in the text an AI assistant generates, and why a mention is a different measurement from a citation, a rank, or a visit.
How an AI answer attributes what it says to the pages it read, and why a citation is a distinct outcome from a click or a mention.
The practice of measuring how content fares inside AI assistants' answers — what can actually be observed, what has to be sampled, and what cannot be attributed at all.