AI visibility measurement

The instability is partly mechanical. Production inference endpoints are non-deterministic even at temperature 0, because server load changes the batch size and batch-size-dependent numerical kernels send identical requests down different floating-point paths^[2]. It is partly the search surface. An empirical study of 11,500 queries across Google Search, AI Overviews and Gemini Flash 2.5 found an AI Overview on 51.5% of queries, drawing on sources with under 0.2 Jaccard similarity to the organic results, and measurably less consistent across two runs of the same query^[3]. The vendor Profound tracked 10,000 randomly sampled prompts over 14 days in spring 2026, against ChatGPT, Perplexity and Copilot, and reported stable results, with day-of-week variation small next to the differences between engines^[4]. A 2026 measurement paper draws the opposite conclusion from the same behaviour: answers vary across runs, prompts and time, so a one-off observation is unreliable and visibility belongs in a report as a distribution rather than a single-point outcome^[5].

A study of more than 100,000 prompt responses covering over 100 brands between March and May 2026 found that on a brand's first visibility run, global household names appeared in 73% of relevant AI answers, established mid-market and regional brands in 44%, and niche and small brands in 11%^[6]. The same work found a brand's measured sentiment flips between runs roughly 6.7 times more often than whether the brand is mentioned at all^[7], so presence tracking is far steadier than tone tracking and should be reported separately from it.

An audit of roughly 37,000 commercial recommendation runs puts the failure at the bottom of the prominence ladder rather than spread evenly across it: among specialist and regional brands, 48% to 52% never surfaced in any run of the set^[8]. A visibility programme for a brand in that band is measuring absence, and needs enough runs to distinguish it from sampling noise.

Measurement also has to allow for citations that do not exist. A per-entity bias study of 100 Hungarian B2B entities, using 1,400 probe runs over 2,062 sources, found well-known tier 1 brands produced a fabricated-citation rate of 52.69% against 37.87% for obscure tier 3 entities (p = 1.67e-11)^[9]. Model familiarity with a brand yields confident wrong citations rather than fewer of them, so a visibility score needs per-entity calibration before it can be set beside another brand's. Citation accuracy covers the failure mode itself.

A 12-language study of 66 brands across three models and 35,640 responses found that moving from an English query to a brand's home-market language raised recommendation share by 0.80 for local champions but only 0.15 for global multinationals. Response stability varied more with which model was asked than with which language was used, at an eta-squared of 0.32 against 0.01^[10]. A programme that samples one model in one language has measured that combination and nothing wider, which is the practical argument behind multilingual GEO.

A controlled experiment of 252,000 trials across six models, testing 18 content factors with brand names anonymised so content effects could be separated from position bias, found topical relevance and position in the retrieved list were the strongest predictors of being cited first^[11]. The founding GEO paper reported content changes could raise visibility by up to 40%, with the efficacy of the strategies varying across domains^[12]. Both results argue for measuring per engine and per market rather than publishing one visibility figure; SEO and GEO covers how much of that overlaps with rank tracking.