Being mentioned is not being recommended
Two companies, two very different numbers, and a published claim that doesn't survive being checked. The distinction underneath it is the most useful thing you can learn about measuring this.
published figures, recomputed · June 2026 · July 30, 2026 · 5 min read
In this article
There is a measurement everyone in this field reports and a second measurement almost nobody does, and the gap between them is where most of the money is wasted.
The first is how often you get named. Ask a hundred buying questions, count the answers that contain your name, divide. Every tool sells this number.
The second is what happens when you are named. Being listed eleventh in a list of twelve is a mention. Being the one the model says to call is also a mention. They are recorded identically and they are worth wildly different amounts.
A June 2026 visibility analysis published a good example of the distinction, and then drew a conclusion from it that the numbers do not support. Both halves of that are instructive, so let’s do both.
The example
Two companies in the same category, measured across a thousand AI responses:
Apollo 645 mentions 21% of them were recommendations
Lavender 59 mentions 32% of them were recommendations
The published framing: the smaller company is more persuasive per mention, because its positioning is more specific. Fewer mentions, more influence.
The idea is right. The evidence for it, at these numbers, is not there.
Why the numbers don’t carry the claim
Fifty-nine is a small sample, and a percentage measured on a small sample has a wide range of values it is compatible with.
Run the standard interval on both:
Lavender 32% of 59 95% interval ≈ 20% to 44%
Apollo 21% of 645 95% interval ≈ 18% to 24%
Those intervals nearly touch. The true recommendation rate for the smaller company could plausibly be 20% — indistinguishable from the larger one. It could also be 44%, which would make the headline an understatement. The measurement does not tell you which, and reporting the point estimate as a finding hides that.
This is not a nitpick and it is not a gotcha. It is the difference between a number you can act on and a number that reads well. If you restructure your positioning because 32% beat 21%, you have made a decision on a coin that was only flipped fifty-nine times.
What would have made it hold? More runs. The same hundred questions asked five times each instead of once gets the smaller company from 59 mentions toward 300, and the interval narrows to roughly ±5 points. Then the comparison means something. The cost is a few hours of compute.
Note what that study’s own methodology page does not say: how many times each question was asked. It says ten models, a hundred prompts per category. It never says how many runs. On a system that produces a different answer to the same question on consecutive attempts, one run is not a measurement — it is a sample of size one, repeated.
The distinction is still the right one
Now the useful part, because the flawed example was pointing at something real.
Two businesses with identical mention rates can be in completely different commercial positions, and the treatment for each is different.
High mentions, low recommendation rate. The models know you exist. They list you. They don’t send anyone to you. You are furniture in the category — named because you’re one of the names, not because you fit the question. This is a positioning problem, and publishing more about yourself makes it worse, not better: more generic material about a generic company produces more generic mentions.
Low mentions, high recommendation rate. When you come up, you win. You just don’t come up. This is an awareness problem, and it is the more tractable of the two. The thing that works is being present in more of the places the models read, so that the questions where you already fit actually surface you. Written by somebody else has where those places are.
Low on both. Start at gate one. Something more basic is wrong.
High on both. Defend it. You are in the position everyone else in your category is paying to reach, and the answer sets are less stable than they look.
How to tell which one you are, honestly
Three rules, and they are the whole method.
Ask the question more than once. The same prompt to the same model on the same day produces different names. If your report was built from one run per question, it is a screenshot of noise. We use two independent runs in the first pass and a third at day seven, and we publish the spread, not just the average.
Separate the count from the outcome. Record, for every appearance: were you named, and were you the answer? Those are two columns, not one. Any report that gives you a single “visibility score” without showing you that split is hiding the more important half.
Look at the interval, not the point. If a number is built on fewer than about a hundred observations, treat differences under ten points as unresolved. This applies to our numbers as much as anyone’s — it is why our report shows you a range for anything measured thinly, and says so on the page rather than in a footnote.
The one place this gets genuinely misused
A recommendation rate is easy to inflate by choosing questions you already win. Ask a hundred questions shaped like “what’s the best tool for [the exact thing you do]” and your rate will be excellent and meaningless. The questions have to be the ones a buyer actually types, including the broad ones where you lose, and including comparisons with competitors you’d rather not name.
If a vendor won’t show you the question list, the score is unauditable. Ours is published with every report, verbatim, including the questions where you place badly.
