The Question Nobody Is Asking
Most brands entering the AI visibility conversation are asking the wrong question. They want to know how to get mentioned by ChatGPT or Perplexity. The question they should be asking is why AI models choose to cite some sources and systematically ignore others, and whether that decision is something a brand can influence or something that simply happens to them.
The answer is that it is influenceable. But not through the mechanisms most brands assume.
What AI Models Actually Do With Sources
Large language models do not retrieve information at the moment of inference the way a search engine retrieves pages at the moment of a query. They build a compressed representation of the world from the corpus of content they were trained on, and they update that representation continuously through retrieval-augmented systems that pull current information from indexed sources.
In both cases, the question of which sources get incorporated into a response is a function of how the model has learned to weight authority, trust, and relevance. That weighting is not random and it is not purely based on domain authority in the traditional SEO sense. It reflects a more complex signal: how consistently has this source produced content that human readers engage with deeply, return to, and treat as reference material rather than as content to be scanned and forgotten.
The behavioral signal underneath that judgment is attention. Not declared attention. Not survey-measured attention. Behavioral attention: the aggregate pattern of how real human readers have physically interacted with a source's content over time.
Trusted, Neutral, and Toxic Sources
BAXindex classifies sources into three categories based on their behavioral attention profile and their track record in AI response outputs.
Trusted sources are those whose content consistently generates Deep Reader and Active Explorer behavioral patterns at scale. Their content earns re-read loops, sustained attention retention, low drop-off rates, and high citation frequency in LLM outputs. A brand that places its content on trusted sources, or that builds its own properties to trusted source standards, is contributing positively to its AI citation probability.
Neutral sources generate average attention profiles — moderate engagement, standard decay curves, neither building nor damaging citation authority. The majority of publisher inventory falls into this category. Appearing on neutral sources at volume is not harmful, but it is not building the behavioral corpus signal that drives AI citation.
Toxic sources are those whose content generates systematically low attention profiles: high Headline Skimmer and Distracted Browser session rates, steep decay curves, low re-read signals, and poor attention retention. Appearing on toxic sources does not just fail to build citation authority. It actively degrades it, because the AI model's representation of a brand becomes associated with low-attention content environments.
This is the source problem. A brand can spend significant budget placing content on high-reach, low-attention sources and generate impressive impression counts while simultaneously training LLMs to treat it as low-authority. The analytics dashboard shows strong performance. The AI citation rate declines.
Why Traditional Brand Safety Misses This
Brand safety frameworks were built to protect brands from association with harmful content: extremist material, misinformation, illegal activity. They are necessary but they address a different problem.
A source can be entirely brand-safe in the traditional sense — no harmful content, no reputational risk, fully viewable placements — and still be a toxic source in the attention sense. A low-quality content farm that produces brand-safe articles nobody reads deeply is not flagged by any existing brand safety tool. It generates clean placement reports and damages citation authority simultaneously.
The attention audit fills the gap that brand safety cannot. It asks not whether a source is harmful but whether it is cognitively valuable: does content placed here generate the behavioral engagement patterns that build AI citation probability, or does it generate the patterns that erode it?
The Trust Sphere
BAXindex maps sources against what we call the Trust Sphere: a classification of publisher and platform inventory by behavioral attention profile, cross-referenced against AI citation frequency data for brands that have placed content across those sources.
The Trust Sphere is not a static list. Source quality changes as editorial standards shift, as audience behavior changes, and as platform algorithms alter the cognitive context in which content is consumed. A publisher that produced strong attention profiles in 2022 may have degraded significantly by 2026 as it scaled content production to meet SEO volume targets and lost the editorial density that made Deep Reader sessions common.
For brands, the practical implication of the Trust Sphere is that source selection is now an AI visibility decision, not only a reach decision. A smaller placement on a trusted source with a high Deep Reader rate contributes more to citation authority than a larger placement on a neutral or toxic source with ten times the reach.
What an Attention Audit Reveals
An attention audit maps a brand's existing content placements against the Trust Sphere, identifies the proportion of placements that are generating trusted, neutral, and toxic attention profiles, and calculates the net effect on the brand's AI citation probability.
The typical audit reveals three things that surprise most marketing teams.
First, a significant proportion of placements that appear high-performing in conventional analytics are neutral or toxic sources in attention terms. The reach numbers look strong. The behavioral signal is weak.
Second, the brand's own properties often outperform its paid placements on attention quality, but receive a fraction of the investment. A brand's editorial blog, its methodology pages, its technical documentation — these frequently generate Deep Reader sessions because users who arrive there have high intent. They are systematically underfunded relative to the paid channels that generate more sessions but less cognitive engagement.
Third, fixing the source mix does not require spending more. It requires reallocating from toxic and neutral sources toward trusted sources and owned properties. Brands that complete this reallocation typically see AI citation rates improve within two to three quarters, as the LLM's representation of the brand's authority updates to reflect the improved behavioral signal.
The Actionable Layer
BAX connects source audit findings directly to deployment. The platform identifies which specific placements are generating toxic attention profiles, which owned content sections are underperforming their potential, and what editorial changes would shift the behavioral signal in the right direction.
This is the four-layer architecture that distinguishes BAX from every other platform in the attention measurement space: Visibility (where the brand appears in AI responses), Attention (what behavioral signal those appearances generate), Sources (which sources are building or eroding citation authority), and Action (what specific changes fix the gap).
The source problem is solvable. But it requires measuring the right layer.