Your LLM visibility data is probably lying to you
After returning from holidays, I found myself questioning the reliability of brand visibility trackers for AI-generated responses. I spent a weekend building a custom solution, pulling data from API providers and comparing fresh web results against logged prompts.
The core problem: probabilistic variance
The fundamental issue stems from how language models operate. Ask the same question ten times and you get ten slightly different answers. This means that observable changes in visibility metrics between measurement periods often represent statistical noise rather than genuine trends.
A brand's visibility score fluctuating from 34% to 41% monthly reflects variance, not growth.
However, patience yields dividends. After weeks and months of repeated observations, genuine patterns emerge from the noise. Consistency across different models and prompt variations becomes the true indicator of meaningful signals.
Prompt sensitivity
Minor wording adjustments can dramatically alter which brands surface in responses. When web search integration enters the equation — sometimes called "query fanout" — instability compounds further. The model operates probabilistically while simultaneously reading a probabilistic index, exponentially expanding the possible response surface.
Methodology requirements
Fixed prompt sets constitute sampling, not methodology. Trustworthy data demands volume, variation, and sustained repetition over time. Only by identifying patterns that remain stable across this noise can one distinguish genuine signal.
Competitive analysis
The productive question isn't "Did my visibility improve?" but rather "Where do competitors consistently appear across models where I don't?" This structural gap reveals actionable opportunities.
Content mapping
Understanding which topics surface a brand and which never do provides valuable signals about content and entity recognition.
Time matters
Data collected over two weeks warrants significant skepticism. Months of consistent prompt testing, conversely, likely contains insights awaiting discovery.