who does the model name when nobody suggests anybody?
companies one round of answers named that the declared rival list never carried
the list you declared is not the list the model uses
A round is collected against the names you tracked, and every tool in this category reports on those names and stops there. The answers do not stop there. In one paid round of twenty questions, written from a customer’s own site and asked four times each, the answers named 339 companies the declared list of thirteen never carried, and eleven of the twenty questions named none of the thirteen at all.
That second number is the one worth sitting with. More than half the questions a customer was measuring themselves on were questions where their whole competitive set was absent from the answer, and a report scored only against the declared list would have shown a clean sheet on all eleven.
how does it decide something is a name?
Two different jobs, kept apart. A tracked brand is matched by the literals you declared, with no model in the loop, so the thing being measured cannot be measured by a guess. Every other name in the answer is harvested afterwards by position in the sentence and by how the word is spelled in the rest of the corpus, and the round records which question and which arm each one arrived through.
lulu rivals --ledger ./ledger --snapshot ROUND --least 2
The result is a table of who came up and where, so the list you carry into the next round is corrected by what the model actually said rather than by what you remembered to type.
what it will get wrong
The harvest is positional, and two kinds of word slip through it. Type words like API, AI and ML sit where a company sits and are capitalised in the corpus the same way, so they appear in the table above real names and a reader strikes them out in a second. Going the other way, a name that never appears inside a sentence and leaves no trace in the surrounding spelling is dropped, which is how Twelve Data was lost from the round of 4 August 2026.
Both are written down here rather than left for you to find, because a table that hides its two failure modes is a table you cannot use as evidence.
see it on a real round
Four questions are published with every name in every answer, asked six times each of one model with no search tool attached, and each name carries the interval its own sample supports rather than a bare percentage.