Brand Drift Score — Methodology

Version 1.0 · April 2026 · Hublo Technologies Pvt Ltd

AskLLM consumer audit datapoints: N=20 prompts × 4 LLMs = 80 datapoints per completed standard audit (Tribunal #12 disclosure; CoreTestCases T1.1.7).

1. Why BDS exists

Bottom line: The industry talks about “AI visibility” with no shared definition. Vendors publish proprietary scores that cannot be reproduced or compared. BDS fixes that with a fully specified formula, a reference implementation, a public benchmark, and a permanent DOI — so any practitioner, agency, or research team can compute the same number independently.

LLMs have become primary recommendation engines for buyer-intent queries. When a prospective buyer asks ChatGPT “what’s the best APM for Node.js?” or asks Gemini “which antimicrobial coating is safest for hospitals?”, the brands that appear in the answer — and the brands 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.

2. The formula

Bottom line: BDS is a weighted sum of four components. Weights are fixed at 0.4 / 0.3 / 0.2 / 0.1 for v0.1.x and will not change without a major-version bump and a new whitepaper DOI.

BDS = 0.4 · citationRate
    + 0.3 · rankQuality
    + 0.2 · sentiment
    + 0.1 · sourceLinkPresence

where  each component ∈ [0, 1]  and  BDS ∈ [0, 1]

3. The four components

3.1 Citation Rate (weight 0.4)

Bottom line: The fraction of prompts in the benchmark cluster in which the brand is named at all. Largest weight because being named is the first-order outcome.

With a 10-prompt cluster and the brand appearing in six, citationRate = 0.6. A brand with a citation rate below 25% is functionally invisible to AI-assisted buyers — no amount of rank quality or sentiment compensates for a structurally absent mention.

3.2 Rank Quality (weight 0.3)

Bottom line: The average inverse-ordinal rank of the brand when named, using a piecewise discount schedule. Being mentioned first and being mentioned fifth are very different outcomes; the schedule is non-linear by design.

position 1      → 1.0
position 2      → 0.7
position 3      → 0.5
positions 4–5   → 0.3
position 6+     → 0.1

Empirically, rank beyond position 3 produces little downstream click behaviour in LLM-mediated flows; the schedule reflects this by flattening the lower tail.

3.3 Sentiment (weight 0.2)

Bottom line: Rescaled polarity of the mention sentence, measured at clause level. Raw polarity in [-1, +1] is mapped to [0, 1] via (p + 1) / 2. Neutral mentions map to 0.5.

The weight (0.2) is deliberately lower than the structural components because sentiment in LLM-generated text skews positive; it is more useful as a negative-flag tiebreaker than as a first-order signal.

3.4 Source-Link Presence (weight 0.1)

Bottom line: The fraction of answers that contain a hyperlink to a domain controlled by the brand (or a verified subdomain). Captures whether the brand converts a mention into a click.

The low weight (0.1) reflects that this component is heavily model-dependent — some models cite links aggressively, others almost never — and therefore should not dominate the composite. Two brands mentioned equally often can have very different source-link rates, and that difference is real commercial signal, but it is refinement, not foundation.

4. Weight rationale

Bottom line: The weights were chosen to impose a strict rank order on the four components, not tuned against an external target. BDS is a defined metric in the same sense as NDCG’s log discount is defined, not a learned one.

Citation frequency matters most, rank quality next, sentiment third, source-link presence as a refinement. The specific numeric values were chosen to sum to 1.0 with a clean stride and to produce interpretable arithmetic in worked examples. Changing the weights silently would break every downstream citation in the whitepaper, the README worked example, the dataset card, and every third-party tool that embeds BDS. The discipline is: never patch weights in place; bump to v1.0.0, publish a new whitepaper DOI, and let v0.1.0 remain as-is.

5. Measurement protocol

Bottom line: BDS 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 have too much variance.

• Cluster size: 8–20 prompts. Smaller clusters admit excessive variance; larger clusters straddle sub-categories and dilute the signal.
• Model coverage: at least two of {gpt-5, gemini-2.5, claude-sonnet-4.5, perplexity} per window. Scores are computed per model and then averaged with equal weight. Per-model BDS is retained separately for drift diagnosis — divergence between models is itself informative.
• Sampling frequency: weekly. Daily sampling over-weights stochastic model rollouts and retrieval jitter; monthly sampling misses inflection points.
• Mention extraction: span-level named-entity recognition against a canonical alias list per brand (legal name, common trading names, accepted abbreviations). Clause-level sentiment; a single answer can contribute positive and negative mentions for the same brand.
• Rank calculation: 1-indexed position in the model’s enumeration (explicit lists) or order of first mention (narrative answers).

6. Drift detection

Bottom line: ΔBDS is the week-over-week delta. The interesting threshold is a rolling-8-week z-score of |z| > 1.5, which filters retrieval jitter while catching real category re-ranking events.

A brand’s time series is the full sequence of weekly BDS values, not any single value. Brands with low volatility (stable weekly BDS) are structurally embedded in their category’s retrieval corpus. Brands with high volatility are position-fragile — usually the mid-tier challengers who bounce in and out of the top 3 as reviews, forum threads, and news cycles churn.

7. Validation

Bottom line: BDS 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.

Observed BDS values spanned 0.33–0.82, consistent with qualitative expectation (category leaders in the top quartile; challenger and regional brands in the lower quartiles). Week-over-week ΔBDS values within single brands were contained within ±0.05 at 4 of 5 boundaries, supporting the protocol choice of weekly cadence.

The seed benchmark is insufficient to compute 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.

8. Limitations

Bottom line: BDS is a descriptive metric, not prescriptive. Four specific things it does not capture, all by conscious design trade-off for reproducibility.

• Paraphrased references. Answers that describe a brand without naming it (“the vendor that pioneered agentless observability”) score zero.
• Voice-assistant answers. Alexa, Siri, and Google Assistant voice surfaces are fragmented and not straightforward to sample at scale in v0.1.
• Visual artefacts. Product thumbnails, logos, and embedded images in AI Overviews are not scored.
• Non-English prompts. BDS v0.1 operates on English-language models and English prompts; multilingual validation is future work.

9. Related work

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

• Peec (peec.ai) — per-prompt visibility metrics against customer-selected prompts, subscription-gated.
• Goodie (goodie.ai) — multi-model coverage positioned as category-level share-of-voice.
• Bluefish (bluefish.ai) — product framed around Generative Engine Optimization (GEO), ranked leaderboards per vertical.
• Profound (tryprofound.com) — enterprise reporting and alerting for reputation-monitoring.

10. Availability

Bottom line: All artefacts are public and permissively licensed.

• Reference implementation (TypeScript, MIT): github.com/hubloai/brand-drift-score
• Published NPM package: @askllm/brand-drift-score
• Seed benchmark dataset (CC-BY-4.0): askllm/brand-drift-benchmark-q2-2026
• Whitepaper (Zenodo, CC-BY-4.0): DOI to be minted on upload — this page will be updated with the live DOI once it exists.

11. How to cite

BibTeX:

@techreport{hubloai2026bds,
  author      = {{Hublo Technologies}},
  title       = {Brand Drift Score: 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}
}