LLM Visibility Index

The industry discusses AI visibility without a shared, reproducible definition. Vendors publish proprietary scores that cannot be independently verified or compared. The LVI addresses this with a fully specified formula, a reference implementation, a public benchmark, and a permanent DOI, enabling any practitioner, agency, or research team to compute the same number independently.

LLMs have become primary recommendation engines for buyer-intent queries. When a prospective buyer asks ChatGPT what is the best APM for Node.js, or asks Gemini which antimicrobial coating is safest for hospitals, the brands that appear in the answer, and those that do not, directly affect pipeline and revenue. This surface differs structurally from organic search: results are synthesised rather than ranked, attribution to source documents is weak, and the same prompt can yield different answers across models, across weeks, and sometimes across consecutive invocations of the same model.

LVI = (0.35 × Visibility) + (0.2 × Favorability) + (0.15 × Factuality) + (0.15 × Actionability) + (0.15 × Consistency). Factuality is not measured: its weight is reserved and contributes 0, so the attainable maximum is 85.

The stored weights are visibility 0.35, favorability 0.2, factuality 0.15, actionability 0.15, and consistency 0.15. They are not renormalised when factuality contributes 0. LVI is a defined metric in the same sense as NDCG log discount is defined, not a learned one.

Visibility carries the largest stored weight. Factuality stays in the vector so the scale can include it later, and today it contributes 0. The attainable maximum is 85. Changing these weights would be a methodology change and is not done on this page.

LVI is computed over a prompt cluster: a semantically coherent bundle of buyer-intent prompts for one narrow category. The cluster is the unit of measurement; single prompts carry excessive variance.

• Cluster size: 8–20 prompts. Smaller clusters admit excessive variance; larger clusters straddle sub-categories and dilute signal.
• Model coverage: production consumer audits use ChatGPT (OpenAI) and Gemini (Google) — two models per window, equal weight when both run. Per-model LVI is retained for drift diagnosis; divergence between models is informative.
• Sampling frequency: weekly. Daily sampling over-weights stochastic rollouts; monthly sampling misses inflection points.
• Mention extraction: span-level NER against canonical alias list per brand. Clause-level sentiment; a single answer can contribute positive and negative mentions for the same brand.
• Rank calculation: 1-indexed position in explicit lists or order of first mention in narrative answers.

ΔLVI is the week-over-week delta. The operational threshold is a rolling-8-week z-score of |z| > 1.5, filtering retrieval jitter while catching real category re-ranking events.

A brand time series is the full sequence of weekly LVI values, not any single value. Brands with low volatility are structurally embedded in their category retrieval corpus. Brands with high volatility are position-fragile, typically mid-tier challengers who bounce in and out of the top three as reviews, forum threads, and news cycles churn.

LVI was validated against a 10-row seed benchmark spanning two verticals (consumer displays; industrial antimicrobial coatings) over five consecutive weeks in Q2 2026. The dataset is published on Hugging Face under CC-BY-4.0.

That seed reported values on a 0–1 illustration scale (0.33–0.82). It is not the stored audit composite, whose attainable maximum is 85 because factuality is not measured. Week-over-week ΔLVI within single brands was contained within ±0.05 at four of five boundaries, supporting weekly cadence.

The seed benchmark is insufficient for category-level baselines; a Q3 2026 release will expand to ≥100 rows with B2B SaaS coverage (APM, observability, mid-market CRM) aligned with the public benchmark page.

LVI is a descriptive metric, not prescriptive. Four specific exclusions follow, each a conscious trade-off for reproducibility.

• Paraphrased references. Answers describing a brand without naming it score zero.
• Voice-assistant answers. Alexa, Siri, and Google Assistant surfaces are fragmented and not sampled at scale in v0.1.
• Visual artefacts. Product thumbnails, logos, and embedded images in AI Overviews are not scored.
• Non-English prompts. LVI v0.1 operates on English-language models and English prompts; multilingual validation is future work.

The commercial LLM-visibility landscape in early 2026 is dominated by proprietary tools. To our knowledge, none has published a formal metric definition, reference implementation, public dataset, and archival paper. LVI is the first open artefact in the category.

• Peec (peec.ai) — per-prompt visibility metrics, subscription-gated.
• Goodie (goodie.ai) — multi-model category share-of-voice positioning.
• Bluefish (bluefish.ai) — GEO framing with vertical leaderboards.
• Profound (tryprofound.com) — enterprise reporting and reputation monitoring.

All artefacts are public and permissively licensed.

• Reference implementation (TypeScript, MIT): github.com/hubloai/brand-drift-score
• NPM package: @askllm/brand-drift-score
• Seed benchmark (CC-BY-4.0): askllm/brand-drift-benchmark-q2-2026
• Whitepaper (Zenodo, CC-BY-4.0): DOI pending mint on upload.

BibTeX

@techreport{hubloai2026lvi,
  author      = {{Hublo Technologies}},
  title       = {LLM Visibility Index: A Composite Metric for Measuring
                 Large-Language-Model Visibility of Named Brands},
  institution = {Hublo Technologies Pvt Ltd},
  year        = {2026},
  month       = {April},
  doi         = {10.5281/zenodo.PENDING},
  url         = {https://askllm.io/methodology}
}