How we run, and how we count.
Written down before the first run so that month three can be held against day zero. What each number can say, and what it cannot.
You can repeat any run yourself in a fresh chat. That is why it is written.
Six rules, in the order they apply.
-
01
Questions
Ten to twenty questions per city and practice area, built from what people type rather than what sounds reasonable, agreed with you, then frozen for the term.
-
02
Runs
Each question more than once, in a temporary chat with memory off, on the assistants named in the agreement. One run is a data point. Repeats make it a measurement.
-
03
Sources
Every page the answer cites or draws on is captured and grouped: directories and ranking sites, the firm's own pages, Google profile signals.
-
04
Counting
A firm counts as named when its name is printed in the answer or in a map card. Name matching rules are written before the run, not after.
-
05
Comparison
Later months use the same questions, the same session state and the same counting rules as day zero. A changed rule means a new day zero.
-
06
Absence
We never say a firm was not named on the strength of one run. If two runs disagree, you get the disagreement, not a conclusion.
Four things no run on this site can tell you.
Location
Our connection is in Germany with the city written into the question. Map cards can differ from a phone in your city. Client runs are checked from a US address.
Interface, not API
We measure the consumer interface, the screen a person sees, not the developer API. The two answer differently and only the first one matters to your client.
Volatility
The same question can return two different lists a minute apart. That is why a rate needs five runs and why a single run is a direction, not a number.
Demand
None of this says how many people ask a given question. What it says is who is named when they do.
Asked and answered.
Which AI assistants do you measure?
The ones named in your agreement. The published record to date is ChatGPT, and we do not put an assistant's name on this site until we have a record on it.
Why five runs before you state a rate?
Because one run of a probabilistic system is one draw. In our own testing a name that appeared in one run was missing in the next often enough that anything under five runs would be a guess with a percent sign.
Can I check a result myself?
Yes, and the guarantee depends on it. Open a temporary chat with memory off, type the agreed question as written, and read the answer. Your screen is the one that counts, not our screenshot.
What happens when the method changes?
It gets a new version date, and every record shows the version it was run under. Older public records stay readable; nothing is silently re-counted.
Watch it run on your city.
We come prepared. Before the call we have already asked AI about lawyers in your city, many times over. You see who it names, who it leaves out, and why.