Beyond SparkToro: What Actually Predicts LLM Visibility

The study used hundreds of volunteers running identical prompts in November and December 2025. The methodology matters: these were real users on real accounts with personal browsing histories and geographic variation, not sterile API calls. SparkToro measured both the composition of brand lists (which brands appeared) and the ordering (the sequence in which they appeared). Composition repeatability was under 1%. Ordering repeatability was closer to 0.1%.

AirOps' 2026 State of AI Search report reached a convergent finding through different methodology: only 30% of brands stay visible from one AI answer to the next, and just 20% remain present across five consecutive runs of the same prompt. Brands earning both citations and mentions were 40% more likely to resurface across multiple runs than citation-only brands.

The instability is real and measurable. SparkToro framed it as a warning against rank-based tools. The more consequential insight is what remained stable when individual responses did not.

Fishkin's own data supports this distinction. While the specific list of brands changed between runs, the brands that appeared most frequently were consistent across the full sample. A brand showing up in 70% of runs for a given category held roughly that rate across measurement windows. The variance was in composition and order; the statistical distribution of appearances was stable.

This is the same principle behind media measurement in every other channel. Television advertisers do not track their exact position in an ad break. PR professionals do not measure their rank in a news cycle. Both track share of voice: the proportion of total category mentions their brand captures over a sustained period. That aggregate frequency metric smooths the noise that makes any individual instance unreliable. PR's measurement evolution from clipping services to layered evidence frameworks took 15 years. LLM visibility is compressing that timeline.

Sill's own data across 7,442 AI responses, 139 brands, and four platforms confirms this pattern. The median AI Share of Voice is 15 out of 100. The metric is not volatile at the aggregate level; individual responses are. The distinction matters: 55% of brands have a 10+ point SOV spread between their best and worst platform, but those spreads are stable over multi-week measurement windows. What SparkToro proved about instability at the instance level reinforces the case for frequency-based measurement at the aggregate level.

Three independent studies published between November 2025 and February 2026 converge on the same finding from different methodologies, sample sizes, and verticals. Microsoft Clarity analyzed over 1,200 publisher sites and found LLM traffic grew 155.6% over eight months, with sign-up conversion at 1.66% versus 0.15% for search traffic: an 11x difference. Subscription conversion followed the same pattern at 1.34% for LLM versus 0.55% for search.

Adobe's analysis of one trillion retail visits during the 2025 holiday season measured AI referral traffic surging 693% year over year. AI-referred shoppers converted 31% higher than other traffic sources, with revenue per visit up 254% and bounce rates 33% lower. These are not users casually browsing; they arrive with intent shaped by the AI recommendation.

Search Engine Land's 13-month analysis from January 2025 through February 2026 measured LLM referral conversion at approximately 18%: higher than paid shopping, SEO, or PPC. Per-platform conversion data shows the advantage is not uniform: Claude leads at 16.8%, ChatGPT at 14.2%, Perplexity at 12.4%. The platform variation is another reason single-platform LLM visibility measurement produces structurally misleading conclusions.

Loamly's 2026 State of AI Traffic benchmark quantified what SparkToro's dark traffic research had theorized: standard analytics tools miss the vast majority of AI-referred visits. The mechanism is straightforward. Mobile AI apps do not pass referrer headers. Users who receive a recommendation in ChatGPT or Claude frequently copy the URL and paste it into a browser, stripping the referral chain entirely. The visit registers as direct traffic in GA4.

The 0.13-1% AI traffic figures that appear in most analytics dashboards represent only the visits with intact attribution. The actual volume is five to ten times higher. Previsible's State of AI Discovery report, analyzing 1.96 million LLM sessions, found AI traffic concentrates on high-intent pages: industry pages at 1.14% visible share, tool pages at 0.95%, pricing pages at 0.46%. These are not casual browsing sessions; they are visitors arriving with a specific recommendation and comparing options.

The dark traffic problem compounds every other measurement challenge. A marketing team reviewing GA4 sees AI referral traffic as a rounding error. The same team's branded search trend, product demo requests, and direct traffic spikes may all contain AI-influenced visits that are structurally invisible to standard attribution. This is why the attribution gap is a measurement infrastructure problem, not a channel performance problem. The channel is performing; the analytics stack cannot see it.

SparkToro's partnership with Datos produced the Q4 2025 State of Search report, and the findings sharpen the urgency beyond what the brand recommendation study alone conveyed. Google desktop searches per US user declined nearly 20% year over year. AI tools nearly tripled their usage share over the same period. Google still accounts for 73.7% of desktop searches across the 41 domains analyzed, but the trajectory is directional and accelerating.

Pew Research confirms the behavioral shift from the consumer side: 39% of US adults now use ChatGPT weekly for decision-making tasks. Similarweb's global tracker shows Gemini growing 643% year over year while ChatGPT has reached 883 million monthly users generating two billion queries daily. Conductor's November 2025 benchmarks found AI Overviews now appear in 25.11% of Google searches, up from 13.14% in March 2025; 87.4% of identifiable AI referral traffic comes from ChatGPT alone.

The volume is there. The growth is there. The conversion advantage is documented across independent studies. The measurement infrastructure is what has not kept pace. Fishkin himself noted on the Near Media podcast in February 2026 that while AI responses are inconsistent at the individual level, aggregate brand presence can be measured statistically. The qualifier matters: it is the difference between a broken metric and a misapplied one.

AirOps' 2026 report found that the overwhelming majority of AI brand visibility originates from content the brand did not create. Listicles, comparison guides, community discussions on Reddit and Quora, and niche publisher reviews drive 85% of brand discovery in AI search. This finding reframes LLM visibility strategy: optimizing your own site is necessary but nowhere near sufficient.

BrightEdge's citation analysis reinforces this from a different angle: only 17% of sources cited in AI Overviews also rank in the organic top ten. Five out of six AI citations pull from pages not on the first page of traditional search results. SE Ranking's study of 2.3 million pages found that domains with Quora and Reddit mentions have approximately four times higher chances of being cited by AI platforms. Citation volume can differ by 615x between platforms for the same domain, with Grok and Claude at opposite ends of the range.

AirOps also found that pages not updated quarterly are three times more likely to lose their AI citations, adding a temporal dimension that static SEO content does not face. Sequential headings and rich schema correlate with 2.8x higher citation rates. The brands with the highest LLM visibility have not just optimized their own properties; they have built a presence across the third-party ecosystem that AI models draw from. That ecosystem is dynamic, platform-specific, and requires continuous measurement to track.

SparkToro established what not to measure: point-in-time rank on any single platform. The research published in the three months since points toward what to measure instead. Frequency-based Share of Voice, tracked daily across every major AI platform, captures the stable signal within the noise that Fishkin's own data validated. Branded search trend in Google Search Console provides a downstream proxy for AI influence on discovery behavior. Content change attribution, using a within-brand control group of queries unaffected by the change, isolates whether specific interventions moved the needle or whether the shift was platform noise.

This is the same multi-layered evidence structure that PR, television, and podcast advertising have used for decades. None of those channels have perfect attribution. All of them justify billions in annual spend through frameworks that are rigorous without being complete. The four-metric framework for AI search ROI operationalizes this approach for marketing teams facing Q2 budget reviews.

The brands that assemble this evidence infrastructure first will have the measurement foundation their competitors are still deferring. Forrester projects 25% of planned AI search spend pushed to 2027 for lack of proven ROI. That deferral is not a channel judgment; it is a measurement gap. The data to close it exists today. The question is whether the tools and frameworks are in place to collect it.