The State of AI Answers, July 2026

For every prompt with six or more archived runs, we score run-to-run agreement of the full named-brand set per engine and grade the prompt by its worst engine. The result is the headline of this edition: on 37 of 57 prompts (64.9%), identical prompts returned a materially different brand roster more than half the time. 17 more were mixed. The set repeated perfectly on 3 prompts (5.3%). The remaining 9 of the archive's 66 prompts had fewer than six runs and are excluded here rather than graded on thin data.

The practical reading: a single screenshot of an AI answer is an anecdote, whichever way it flatters. One brand can sit stably in an answer while the rest of the roster churns around it, which is why presence counts and set agreement are separate measurements on the methods page. This instability is also why nothing here reports a per-query rank.

Across all 66 prompts, youtube.com appeared in the cited domains of 56 (84.8%) and reddit.com in 53 (80.3%). By run volume they beat every vendor site and every publisher: no other domain comes close. The median prompt drew on 28 distinct cited domains (min 5, max 66), pooled across every run of that prompt. The grounding pool per question is wide, and video plus forum content sits in most of those pools.

The rest of the list is what you would hope: the actual vendors in the measured categories, plus the horizontal utility players. UGC platforms at the top, brand sites underneath. A domain counts once per run here even if an answer cites it repeatedly.

Both regimes exist, cleanly separated. In 22 of 28 small verification baskets (5 runs each, run as gate checks), a single brand appeared in 80% or more of runs: the answer is already decided, and it is usually the incumbent, not the newest entrant. Meanwhile the large baskets are contested. In our own category the leader, Profound, appears in 69 of 226 runs (30.5%), with Otterly at 26.1%. The AEO-agency basket is even flatter: the leading agency shows in 12 of 45 runs (26.7%). Nobody selling AI visibility, ourselves included, owned their own answer on this snapshot. Ours is the number under a standing public test: the archive tracks whether our own playbook moves it, on the evidence log, nulls included.

The 28 verification baskets sit in categories currently in our outreach pipeline, so they publish in aggregate only: we never name companies we might be talking to (the data-handling rules are on the methods page). One large basket's tracked property is unnamed for a different reason: it has outside stakeholders, and that redaction is itself a logged entry (#007) on the evidence log. The container-runtimes basket shows concentration is not only a small-sample artifact: across 75 runs, Docker Desktop appears in 94.7% of answers, with OrbStack close behind. In a fragmented category, placement work can still change the answer. In a decided one, the honest first deliverable is the concentration reading itself, because it tells you which fight you are in before you fund the wrong one.

Everything above is computed from our append-only archive: 721 runs, 66 distinct buyer prompts, three engines, archived with timestamps from run #1 and never rewritten. What we query and how baskets, variance bands, and the brand-set metric work is fully disclosed on the methods page, including what we refuse to measure: per-query rank does not appear in this study and never will, because it does not survive repetition. Prospect and client readings are excluded from publication by rule.

If you want this instrument pointed at your own category, GapCheck runs the same basket methodology on the prompts your buyers ask, with the same variance bands on every number.