In short: Four numbers get written into GEO plans that cannot be targeted honestly: a blended visibility percentage, a citation count, a share of voice, and a rank.
Each fails for a different reason, and knowing which reason matters, because the fix is different every time and three of the four have an honest version underneath them.
The one with no honest version is rank. There is no position to hold, so a target for one is a target for something that does not exist.
A Blended Visibility Percentage
The number a tool gives you: one figure across platforms and markets, often out of a hundred.
It fails because the things averaged do not belong in the same average. Platforms cite almost different sets of sources, and the study on this site found that the same question in two languages agrees on 4% to 11% of the sites it names. An average over environments that share almost nothing describes none of them.
The honest version underneath it: the same measurement, reported per platform and per market, unblended. It is more numbers and less comfortable, and it is the only form in which a change tells you where to act.
A Citation Count
"Grow citations from 40 to 60 this quarter." It sounds measurable, and the counting is the easy part.
It fails because the denominator is not fixed. Citations depend on how many questions you asked, on which platforms, how many times, and on which afternoon. Run the same set twice and the count moves. A target on an unstable count is a target you can hit by measuring more.
It fails a second way if the count mixes the four outcomes. A linked citation, a named source, a brand mention and unattributed use are not interchangeable, and a single number treats them as though they were.
The honest version: a rate rather than a count. Named in four of ten questions, in two of three runs. The denominator is fixed by the question set, so the number cannot be improved by asking more.
Share Of Voice
Your mentions as a proportion of all brand mentions in a category's answers. It is the most sophisticated-looking number in most GEO reports.
It fails because the denominator is unknowable. You cannot enumerate every brand an assistant might name across every phrasing of every question, so what gets computed is your share of the brands that appeared in your sample. Change the sample and the share changes without anything changing in the world.
- Visibility score of 34 out of 100Named in 4 of 10 questions on ChatGPT, 6 of 10 on Perplexity
- Citations up 50% this quarterNamed in 2 of 3 runs, up from 1 of 3, on the same ten questions
- Share of voice 12%Named alongside these four competitors; absent where these two appear
- Ranked third for our categoryPresent in the answer, not the first source named
A Rank
This is the one with no honest version, because the thing it measures does not exist.
A generated answer names a few sources and does not order them by merit. Being named third is not a position in a ranking; it is being named. There is no fourth place to lose, no first place to hold, and nothing that behaves like a gradient you can climb.
Tools that report a rank are constructing one, usually from how often you appear across their sample. That construction is not wrong so much as invented, and a target set on it is a target on the tool's method rather than on the world.
What to say instead when somebody asks for a rank: whether you are in the answer at all, and who is in it with you. Those are the two facts that exist, and the second one is more useful than any position would have been, because it names the competitors whose coverage you would have to displace.
Why These Survive Anyway
All four persist because a stakeholder asked for one number and somebody had to answer. That pressure is real and the answer is not to refuse.
It is to give the one number that is defensible, which is presence against a fixed set, and to bring the competitor list with it. A named list of who appears instead of you does more in a meeting than a score, because it turns an abstraction into a decision about specific companies.
Key Takeaways
- A blended visibility score averages platforms and markets that share almost no sources. Report it unblended.
- A citation count has no fixed denominator. Use a rate against a fixed question set instead.
- Share of voice needs a denominator nobody can enumerate, so it measures your sample rather than the market.
- A rank has no honest version. Generated answers have no positions, only inclusion.
- When one number is demanded, give presence against a fixed set and bring the list of who appears instead of you.
Check yourself
Before you move on
Not scored, not recorded, and not part of the certificate. Both answers are settled by a sentence in this lesson, and the reasoning appears whichever option you pick.
- 01
Why can share of voice not be targeted honestly?
- 02
Which of the four bad targets has no honest version underneath it?