How we measure your brand's AI visibility

Models

How we measure visibility

Every model we query answers against the live web. Each one reaches for sources differently, so we treat them as four independent readings of what the web says about you today. That independence is measurable. In our own study, 900 runs of the same 30 questions on ChatGPT, Gemini, and Perplexity, any two surfaces agreed on the top answer only 20 to 23 percent of the time (Jaccard 0.20 to 0.23).^[9] An independent 2026 study across many categories found a similar split: models agree on the top-recommended brand 41.6% of the time.^[1] A brand's standing on one surface says little about the others, so tracking a single model misses most of the picture.

Scoring

How your visibility score is calculated

One number that tells you how visible your brand is across AI. Here's exactly what goes into it.

1. 1

   Prompts sent to each model

   Your tracked prompts are sent to each AI model's API: ChatGPT, Claude, Perplexity, and Gemini. Each prompt is run multiple times per model because LLMs give different answers to the same question. Peer-reviewed studies show outputs vary even at "deterministic" settings, so a single response is not a reliable measurement.^[2],[3],[4]

2. 2

   Responses checked for mentions

   Every response is analyzed for your brand name using both an exact case-insensitive match and a fuzzy normalized check that strips non-alphanumeric characters. A mention is detected if either method finds your brand.

3. 3

   Score calculated

   Your visibility score is the percentage of queries where your brand was mentioned. Mention rate over repeated prompts is the measure independent research converges on: exact AI answers almost never repeat, but a brand's mention rate is stable across runs, which is also why we don't sell an "AI ranking position" metric.^[6],[7]

   Score = (mentions / total queries) × 100

4. 4

   Errors excluded

   If a query fails due to an API error or timeout, it's excluded from the denominator entirely. Your score only reflects responses that were actually received and analyzed.

Approach

Why we query models directly

We query each AI model's API directly, the same models that power ChatGPT, Claude, Gemini, and Perplexity.

Improving

What moves your score

Every model we query is doing the same thing under the hood: searching the web for sources that answer the prompt, then composing an answer from what it finds. This is measurable and moveable: the foundational peer-reviewed study on generative engine optimization found that targeted content changes boosted a source's visibility in AI answers by up to 40%.^[8] Four things consistently show up in the sources that get cited.

1. 1

   Fresh web content that answers the prompt

   Reddit threads, Quora answers, and recent articles that mention your brand in the context of what the prompt is actually asking. Relevance to the question beats generic brand mentions every time.

2. 2

   Authority on sources AI search weights heavily

   Wikipedia, major publications, and industry-specific subreddits that reliably surface in grounded searches. A mention on a domain the models already trust moves the needle.

3. 3

   Repeat mentions across independent sources

   One mention on one site is easy to pass over. Three independent sources corroborating the same claim is much harder to ignore. That’s when models start treating it as the default answer.

4. 4

   Prompt-term adjacency

   Your brand name appearing near the prompt’s core keywords on the source page. Proximity is how search-grounded models decide which mentions are relevant to the question being asked.

References

The research behind this methodology

The sources cited above, in full. We label each one honestly: peer-reviewed papers passed independent academic review; preprints and industry studies haven't, but publish their data and methods openly. Our own measurement is labeled as such: one vertical, descriptive, and reviewed by nobody but us. No study validates our exact run count. The research supports measuring over repeated runs as a practice, and three runs is where we balance statistical stability against querying cost.

1. [1]
   Preprint

   Who Owns the AI Recommendation? A Multi-Industry Empirical Map of Brand Category Ownership Across Large Language Models (opens in a new tab)

   Żatuchin, D. · arXiv:2606.23057, 2026

   Found only 41.6% agreement between models on the top-recommended brand. A top spot on one model doesn’t carry to another, so tracking a single model gives an incomplete picture.

2. [2]
   Peer-reviewed

   The Good, The Bad, and The Greedy: Evaluation of LLMs Should Not Ignore Non-Determinism (opens in a new tab)

   Song, Y., et al. · NAACL 2025

   Shows that judging an LLM from a single response per prompt is unreliable; the paper itself samples each prompt many times and reports averages across runs.

3. [3]
   Peer-reviewed

   Non-Determinism of "Deterministic" LLM Settings (opens in a new tab)

   Atil, B., et al. · Eval4NLP @ ACL 2025

   Across 5 LLMs, 8 tasks, and 10 runs each, accuracy varied up to 15% between identical runs. Even at temperature 0 with fixed seeds, no model produced repeatable outputs.

4. [4]
   Peer-reviewed

   Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference (opens in a new tab)

   Yuan, J., et al. · NeurIPS 2025 (oral)

   Traces run-to-run variation to the inference infrastructure itself (floating-point non-associativity, GPU batching), which callers of commercial LLM APIs cannot switch off.

5. [5]
   Peer-reviewed

   On the Reproducibility of LLM-Centric Empirical Studies (opens in a new tab)

   Angermeir, F., et al. · ICSE 2026

   A replication of 85 LLM studies ran up to 30 repetitions per study specifically because of non-determinism, observing up to 30% metric differences between repetitions.

6. [6]
   Industry study

   AIs Are Highly Inconsistent When Recommending Brands: Marketers Should Take Care When Tracking AI Visibility (opens in a new tab)

   Fishkin, R. (SparkToro) & O’Donnell, G. (Gumshoe.ai) · SparkToro Research, 2026

   Across 2,961 runs of 12 prompts, exact brand lists almost never repeated, yet per-brand mention rates stayed stable over many runs, concluding that visibility % across many prompts run multiple times is a sound metric, while "AI ranking position" metrics are not.

7. [7]
   Preprint

   Don’t Measure Once: Measuring Visibility in AI Search (opens in a new tab)

   Schulte, B., Bleeker, F. & Kaufmann, E. · arXiv:2604.07585, 2026

   Repeated identical AI-search queries shared only 32–43% of cited sources; brand visibility should be measured as a distribution over repeated queries, not a one-off observation.

8. [8]
   Peer-reviewed

   GEO: Generative Engine Optimization (opens in a new tab)

   Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K. & Deshpande, A. · KDD 2024

   The foundational generative-engine-optimization study (Princeton/IIT Delhi): targeted content changes (adding citations, quotations, and statistics) boosted source visibility in generative engine responses by up to 40% on a large multi-domain benchmark.

9. [9]
   Our measurement

   Repeatability of AI shopping-agent answers: 30 queries, 5 repeats, 3 surfaces, two collection days

   Turner, K. (Lumidian) · Lumidian working paper, September 2026 (Study 0; write-up on request)

   900 API runs of 30 headphone queries on ChatGPT (GPT-5.6 with web search), Gemini 2.5 Pro (Search grounding), and Perplexity (Sonar Pro), collected on two days nine days apart. Within-day repeatability of the top-3 answer: Perplexity 0.80, ChatGPT 0.62, Gemini 0.45 (median Jaccard). Agreement between surfaces on the same query: 0.20 to 0.23. Constrained questions (a budget, a use case) were not stable on any surface. One vertical, descriptive, not peer-reviewed.

↑ Back to top