1 / 270 ai recommendation rate · cold · about 0.4%
segmentcold
queries270
models_queried6
methodmeasured live
run_idmonth-0
date2026-05-18
statuspublished
register: machine · public record
The Recommendation Ledger.
The Recommendation Ledger is Dual Cognition's public record of measured AI recommendation rates: real queries to the major AI models, segmented cold vs branded, every number bound to a run id and date. We publish the result even when it stings, including our own.
ledger entries
The record, run by run.
Every entry below is a real measurement: live API calls, a stated date, a stated run id, and the count of models queried. Nothing is averaged across runs, smoothed, or projected. And full disclosure: SuperHired is a brand from our own portfolio, client zero. That's exactly why we can publish its ugliest number.
0 / 135 ai recommendation rate · cold · 0%
segmentcold
queries135
models_queried3
entity_consistency69.6 / 100 · 15 surfaces
methodmeasured live
run_idinitial-v8
date2026-05-22
statuspublished
0 / 180 ai recommendation rate · cold · 0%
segmentcold
queries180
models_queried4
methodmeasured live
run_idinitial-v11
date2026-05-26
statuspublished
our own first self-audit, run the week before our site launched. this is the starting line we sell from.
cold rate0% 0 of 219 cold queries
branded rate100% 60 of 60 · the query named us · echo, not proof
entity accuracy0.00 asked directly, AI doesn't know us yet
run_idself-month-0
date2026-06-09
modelsclaude · grok · perplexity · 279 live responses
statuspublished · measured live
download this run · self-month-0.json
data dictionary: in the raw artifact, the top-level recommendation_rate field is the blended figure across all segments, cold and branded together. the cold rate above derives from awareness_breakdown summed over the cold stages (unaware, problem-aware, solution-aware). that's why this entry reads 0 of 219 cold queries, never the blended figure.
Baselines, not highlights. Month 0 is where honest counting starts, and it's the same starting line we hold ourselves to.
how to read the ledger
Four terms carry every entry.
The Ledger has no fine print. It has definitions, and they don't move between runs.
- cold
- Queries from a buyer who's never heard of the brand. They describe the problem, not the company, and the models answer with whoever they actually know. This is the real test, and it's the number we lead with.
- branded
- The buyer already named the brand in the query. Near-tautological: of course the model can discuss a company you just handed it. We report it separately and never blend it into the cold number.
- measured live
- Real API calls to the named models on the stated date. No simulation, no synthetic panel, no replaying old answers. If a model wasn't actually queried in that run, it isn't in that run's count.
- run id
- Every figure on this page carries the id of the run that produced it. A number without provenance is an opinion in a suit. With a run id and a date, you can ask us for the receipts, and we can hand them over.
why publish the sting
A number you can verify beats a chart you have to trust.
Every agency keeps a folder of charts that make the story look good. The Ledger isn't that. It's the measured record, published before it flatters anyone: an about 0.4% here, two flat zeros there, and our own self-audit on the same page, held to the same format. We work this way because of a first principle of the practice: proof must be agent-discoverable. A claim without a published, traceable number is a promise only a human can take on faith, and machines don't take anything on faith. They read what's on the record. So we keep the record public, run ids and all, and let you check it.
first principle · proof must be agent-discoverable