Intervals, not points. Every rate carries a 95% interval and its sample size.
≥7 runs per core prompt per day, pooled over 2–4 week windows, with the effective sample size reported per prompt cluster. Below that we label results “snapshot” or “insufficient” and draw no trend conclusions. Long-tail prompts run weekly and are only reported at window level.
We disclose our own biases first.
How intervals are computed
Runs of the same prompt are correlated, so we use a two-stage (prompt-clustered) bootstrap: resample prompts with replacement, then resample runs within each chosen prompt; 2,000 resamples, fixed seed. The estimate is the equal-weight average over prompts. In simulation, treating every run as independent produced intervals about half as wide as they should be, covering the true value only 51–72% of the time instead of 95%. We also publish the intraclass correlation (ICC): when it is high, uncertainty comes from which questions were asked, and more prompts help more than more repeats.
Why our intervals are wider than “±5 points”
Some tools quote tight margins by treating every run as an independent sample. Runs of the same question are not independent, so those margins are too narrow — in our simulation they covered the true value only about half to three quarters of the time. Ours are wider because they are honest about where the uncertainty comes from: mostly from which questions were asked. When we compare you with a competitor we compare you on the same answers (paired differences), which cancels much of that shared noise.
Mentioned is not recommended
We report whether an answer names you, cites a page about you, and recommends you — separately, plus whether it describes you correctly against a fact sheet you sign. In our sandbox experiment, a large share of “mentions” of a new brand were the assistant telling people to wait for reviews.
Sample sufficiency
Trend-grade = ≥14 days of data and a median of ≥7 runs per prompt per day. Snapshot = ≥7 runs per prompt but fewer days: shows structure, not change. Below that we do not draw conclusions.
Stability metrics
Same prompt, same day — source overlap (Jaccard)
Same prompt, same day — rank overlap (RBO)
Same prompt, consecutive days — source overlap
Different phrasings — source overlap
Published baselines: day-to-day citation Jaccard 0.34–0.42; same-moment repeats 0.32–0.43. When the phrasing overlap is lower than the repeat overlap, “ask it differently and you get a different market” — the largest, most ignored source of risk.
Known biases (ours included)
API ≠ the logged-in app. Consumer apps add memory, personalization, location and model routing. We sample programmatically with web search enabled; every report names the answer channel (official API or an API-compatible gateway) and the model that answered. The gap between the two has not been quantified by anyone in this category — including us — so we do not pretend it is zero.
Intervals are prompt-clustered. Repeated runs of the same prompt are correlated, so we resample prompts first, then runs within each prompt. Treating every run as independent (the common shortcut) produced intervals about half as wide as they should be in our simulations, covering the true value only 51–72% of the time instead of 95%.
The prompt set defines the market. A different set of questions is a different market. Our prompts are listed in full in the appendix so you can check them.
Brand detection is string matching on names and aliases. It misses purely semantic mentions and can be fooled by common words.
What we will not do
We do not guarantee rankings, citations or a share of AI answers.
We do not seed Reddit or any community, buy reviews, or post undisclosed content.
We do not sell llms.txt, 'AI schema' or chunked rewrites as results.
We do not charge per citation or per ranking — AI answers move ±30–200% for reasons unrelated to anyone's work.
We do not publish self-ranked “best of” lists, for ourselves or for clients.
We do not show logos or testimonials we do not have permission to show.