Fix entity records before chasing volume.<\/strong> Nine of the eleven fastest sweeps in the panel involved a Wikipedia or Wikidata edit. If your entity record is wrong or missing, external coverage propagates to fewer engines than it should.<\/p>\nWhere this diagnostic breaks down<\/h2>\n
Four honest limits, because a diagnostic you cannot falsify is not a diagnostic.<\/p>\n
Small samples fake patterns.<\/strong> Single-prompt mention rates carry a \u00b121pp sampling band at typical run frequencies. Diagnose on prompt clusters, never on one query.<\/p>\nUnexplained changes dominate.<\/strong> We reached a documented cause for about one in three step changes. The rest had no traceable artifact. Anyone claiming near-total attribution in AI search is overstating it.<\/p>\nPersonalisation and locale contaminate comparisons.<\/strong> Engines vary answers by region, account history and interface surface. Consensus computed across inconsistent conditions measures your test setup, not the engines.<\/p>\nEngines run their own experiments.<\/strong> Model updates and retrieval A\/B tests produce movement that looks causal and is not. A solo move with no traceable source and no index-family echo is often exactly this.<\/p>\nThe diagnostic is a prior, not a verdict. It tells you where to look first, which on a panel this size is worth roughly a week of investigation per event.<\/p>\n
Frequently asked questions<\/h2>\n
How do you calculate cross-engine AI visibility consensus?<\/strong>
\nDivide the number of tracked engines that name your brand for a prompt cluster by the number of engines tracked for that cluster, then multiply by 100. Six of eight is 75%. Compute it at cluster level over a fixed window, never on a single prompt.<\/p>\nHow many engines need to move before I call it consensus?<\/strong>
\nFive of eight is a reasonable threshold, but only after collapsing into index families. Five engines spanning at least three of the five families is a genuine signal. Five engines from two families is one index refreshing, and should be treated as a partial move.<\/p>\nWhat is a normal consensus score?<\/strong>
\nDepends entirely on prompt type. Panel medians: 88% on branded prompts, 31% on category prompts, 22% on comparison prompts, 12% on problem-led long tail. Above 50% on a category cluster puts you in the top decile.<\/p>\nIs high cross-engine AI visibility consensus always good?<\/strong>
\nNo. Consensus concentrated on a single source domain is fragile. In our panel, brands whose consensus later collapsed had a median Consensus Source Concentration of 61%, against 28% for brands whose consensus held for 26 weeks. Measure diversity alongside agreement.<\/p>\nWhy does Claude usually move last?<\/strong>
\nIts median lag from third-party source publication to sustained mention lift was 17 days in our panel, against 1 day for Perplexity. Most sweeps therefore register as six or seven of eight engines rather than all eight, and a Claude-first move is strong evidence the cause is not a fresh external page.<\/p>\nCan I use this if I only track three engines?<\/strong>
\nPartly. Three engines from three different index families gives a usable synchronous-versus-solo read. Three engines from one family gives almost nothing, because they co-move for infrastructural reasons regardless of what happened to your brand.<\/p>\nWhat is the minimum tracking frequency for this to work?<\/strong>
\nDaily runs on a stable prompt set, with core prompts sampled multiple times per day and aggregated to cluster level. Weekly sampling cannot resolve a seven-day synchronous window, which is precisely the discrimination the whole method depends on.<\/p>\n
![\"Line](\"https:\/\/maxaeo.ai\/blog\/wp-content\/uploads\/2026\/07\/1784554351894-10-51904-1.jpg\")
<\/figure>\n![\"Diagram](\"https:\/\/maxaeo.ai\/blog\/wp-content\/uploads\/2026\/07\/1784554351894-10-51904-2.jpg\")
<\/figure>\n