How we measure, and how we say what we could not

Which engines we ask

Every question is asked of each engine below, separately. The weights approximate each engine's reach in Turkey and normalise themselves: only engines that answered enter the denominator. When an engine does not answer a question, that answer leaves both the numerator and the denominator; it is never counted as "did not mention you".

ChatGPT

0.6

Gemini

0.25

Perplexity

0.15

Claude

0.05

Google AI Overview

0.5

Google AI Mode

0.3

The question panel

A default scan is 40 questions. They are written the way 5 researched buyer personas for the brand's sector would ask an AI, under the same category quota for every persona. The panel is frozen: the brand is measured on the same questions next time, so two periods can be compared.

Category questions

× 1

Comparison questions

× 1.2

Purchase questions

× 1.5

• Questions that already name the brand ("how do I use X?") leave the visibility rate: the answer describes the brand, but it cannot measure whether the engine picks the brand unprompted.
• With fewer than 5 distinct questions that do not name the brand, no visibility rate is stated. A rate over one question is not a rate.
• Questions where the engines disagree are asked 3 more times of each engine that answered (within the run's spending ceiling), and the majority verdict is used. Questions every engine agreed on are not repeated: the uncertainty is not there.

Is the brand named really you

A name appearing is not proof the brand was named: it may be a namesake company, a person, or an answer echoing the question back. Every mention is verified separately and weighted by how the answer placed it. An adverse mention does not count as zero; it counts against.

Named as the leading option

1

Named among the strong options

0.6

Named among the alternatives

0.35

Named with a caveat that would give a buyer pause

-0.5

Named as the option to avoid

-1

• A mention that could not be verified is not counted in the brand's favour: it leaves both the numerator and the denominator of the visibility rate.
• If more than 20% of mentions could not be verified, visibility is shown as "not measured". A rate computed across that much missing data is a guess, not a measurement.

How much the number moves

Next to the visibility rate we publish an interval. It comes from a 20,000-replicate bootstrap that resamples whole questions within their design strata (persona and category), and it is the 95% percentile interval. The unit is the question: five engines answering one question are five readings of one measurement, not five measurements.

• The interval is not a confidence interval for the brand's "true" visibility; the questions were written to a design, not drawn at random. What it says is narrower and still useful: how stable the number is under the panel that produced it.
• The interval is the same every time for the same measurement; it does not change between two renders.

When we call a change a change

Two measurements are compared only when they used the same question set and the same scoring rules. A difference is written as "up" or "down" only when it is larger than asking the same questions again could produce. Otherwise the report gives both numbers and the range the difference could lie in, with no direction word and no arrow.

Similarity between answers to one question (ICC)

0.362

• The range is a Newcombe interval for the difference of two proportions. Because one question's answers across engines resemble each other, the answer count is reduced by that similarity before the interval is computed.
• Because the report picks its largest changes out of many rows, each row's bar is raised for the number of rows compared (Šidák correction). A difference picked from many rows has to clear a higher bar.
• When the range includes zero, no change is claimed. "Unchanged" is not written either: that is a claim too.

How the score is built

The total is a weighted mean of five pillars summing to 100. A pillar that could not be measured is not scored as zero; its weight leaves the denominator. With one exception: if visibility could not be measured, no total and no grade are given, because what remains is mostly site readiness and would raise the total.

Visibility

30

Citation

20

Perception

10

Content

25

Technical

15

• Perception is not computed from fewer than 5 mentions, in answers to questions that do not name the brand, whose tone was read. A mention that could not be read is not counted as "neutral".
• Grades: A+ ≥ 90 · A ≥ 80 · B ≥ 65 · C ≥ 50 · D ≥ 35 · F ≥ 0

What low visibility does not mean

Visibility measures how often AI answers name the brand unprompted. It measures nothing else. Our reports never present a low rate as evidence of any of these:

• That consumer demand is low
• That brand awareness is low
• That market share is low
• That consumers do not prefer the brand
• That AI does not recommend the brand