LLM Citation Benchmark — the measurement framework for AI-search visibility.

Why a standardised benchmark is needed

The market for LLM citation analysis in 2026 is fragmented. Individual tools measure in different ways, advisory firms report contradictory numbers, and academic research uses methods that are rarely replicable in commercial advisory work. The fragmentation has practical consequences: brands receive reports stating "42 per cent citation rate" or "0.73 share of voice" without any comparable basis. Board presentations are built on metrics that are not collected consistently between quarterly reports.

The benchmark framework documented here proposes a reproducible measurement methodology that satisfies four criteria. First, transparency — every methodological decision is discussed openly; no black-box scoring. Second, replicability — the prompt matrix is documented, the execution cadence is standardised, the analysis logic is codified. Third, comparability — industry baselines are published so that individual brand values can be contextualised. Fourth, robustness — statistical significance is tested explicitly, confidence intervals are reported, known measurement errors are named.

The framework in detail

The framework defines seven primary metrics and five secondary metrics, broken down by model and by industry.

Primary metric 1 — citation rate. The share of prompts in a matrix in which the brand appears explicitly as a source or as a content reference in the answer. Operationalisation: a brand counts as cited if either (a) the brand URL appears as a source link in the LLM output, (b) the brand name appears in the answer text alongside an unambiguous identifier (product name, CEO name), or (c) Perplexity displays a source card with the matching domain. Measured per model, aggregated over 7-day periods.

Primary metric 2 — AI Answer Rate (AAR). The relative position of the brand within the LLM answer, operationalised on a 3-point scale: top (first 30 per cent of the answer), middle (30–70 per cent), tail (final 30 per cent). The AAR top-rate is particularly valuable because it approximates user perception probability.

Primary metric 3 — source-origin breakdown. Classification of cited sources into Own (the brand's own domain), Third-Party (trade media, Wikipedia, third-party reviews) and Hybrid (brand mention inside a third-party source). A high Own share signals strong owned-content infrastructure; a high third-party share signals strong external corroboration; a healthy mix is typically the most stable foundation.

Primary metric 4 — competitor share of voice. The brand's citation density relative to a defined competitor list (typically three to five direct competitors). Calculated per prompt and aggregated.

Primary metric 5 — entity-resolution rate. The share of citations where the brand is resolved correctly as an entity (right name, right category, right location reference). A low rate signals entity ambiguity or weak disambiguation.

Primary metric 6 — hallucination rate. The share of citations where the LLM attributes incorrect attributes to the brand (wrong founder, wrong product description, wrong pricing, wrong features). Verified manually against ground-truth sources.

Primary metric 7 — source freshness. Average publication date of cited sources relative to the current date. Low freshness (older sources) signals content-aging risk.

The five secondary metrics are: citation length (average length of brand-related text segments in answers), context accuracy (whether the brand is cited in a thematically appropriate context), link presence (whether citations carry a clickable URL), consistency score (consistency of attributes across different prompts) and temporal drift (change of metrics over time).

Prompt-matrix architecture

The prompt matrix is the central measurement instrument. It is constructed across five dimensions, each mapping a distinct user intent.

Dimension 1 — brand queries. Prompts that name the brand explicitly and reflect typical informational user intents: "What is [brand]?", "Who is the CEO of [brand]?", "What is [brand] known for?", "Pros and cons of [brand]". Typical matrix size per brand: 50–100 prompts. These prompts measure primarily entity resolution and the consistency of the model's answer.

Dimension 2 — category queries. Prompts that name the category without the brand: "Best CRM platforms for B2B", "Leading cloud providers 2026", "Reliable insurers in Germany". Matrix size: 100–300 prompts. These measure primarily unprompted recall — is the brand mentioned spontaneously?

Dimension 3 — use-case queries. Prompts that describe concrete problem scenarios: "How do I optimise email marketing for an online shop with under 100 employees?", "What are solutions for supply-chain transparency in the Mittelstand?". Matrix size: 100–200. These measure association strength — is the brand correctly placed in the problem context?

Dimension 4 — comparison queries. Prompts that trigger direct comparisons: "[Brand A] vs. [Brand B]", "Alternatives to [dominant competitor]", "Which is better — [Brand A] or [Brand B] for [use case]?". Matrix size: 50–150. These measure relative positioning and competitive narratives.

Dimension 5 — long-tail queries. Specific, rare queries with concrete details: "How does [brand] integrate with [third party]?", "Pricing of [brand] for a 500-employee company". Matrix size: 100–300. These measure deep content coverage.

Execution protocol

Execution follows a standardised protocol that minimises measurement noise.

Time windows. Each model combination is sampled in at least two discrete 7-day windows, with at least 14 days in between. This filters out short-term model drift (daily rerouting decisions, A/B tests).

Geographic control. All prompts are executed via defined geo-IP proxies — typically a Frankfurt IP for DACH benchmarks, Paris plus Amsterdam for EU, New York plus San Francisco for the US. Identical prompts across different geos show substantial differences in LLM answers, and these must be reported explicitly.

Personalisation isolation. All sessions run in isolated incognito browsers or fresh API sessions. No persistent user history, no accumulated preferences. This rules out personalisation effects that may be active on pro accounts (ChatGPT Plus, Claude Pro).

Model-version pinning. Where possible, specific model versions are targeted (GPT-4o vs. GPT-5, Claude Sonnet 4.6 vs. Claude Opus 4.6). Model updates can shift citation rate by double-digit percentage points; pinning enables clean tracking of update effects.

Redundancy sampling. The same prompt is executed three to five times per window. Internal variance in LLM answers to the same query is measurable; redundancy yields the best estimate of the "stable" answer pattern.

Attribution logic

The classification of when a brand counts as cited follows hierarchical criteria with decreasing confidence.

Tier 1 — explicit citation with URL. The LLM answer contains a clickable link to the brand domain or a numbered source card (Perplexity). Highest confidence, unambiguously classifiable.

Tier 2 — explicit brand mention plus unambiguous identifier. The brand name is named alongside a product name, CEO name, founding context or other disambiguating information. Medium confidence; human validation on edge cases.

Tier 3 — indirect reference via description. The LLM describes characteristics that unambiguously fit the brand without naming it ("the German market leader for industrial automation, based in Stuttgart"). Lower confidence; reported separately in some reports and not aggregated into the primary metric.

Non-citation. A pure name mention without context coherence (e.g. as a homonym association) does not count as a citation. This classification requires entity disambiguation, typically handled by an NER pipeline plus manual validation.

Industry baselines

From 12,000 prompts across 150 brands we derived industry baselines that serve as a reference point for individual brands. The baselines are not normative targets — they are averages that show where an industry stands statistically.

B2B software: median citation rate 24%, 75th percentile 41%, top decile 58%. Legal & professional services: median 31%, 75th percentile 48%, top decile 64%. Healthcare: median 18%, 75th percentile 34%, top decile 52% (YMYL conservatism in the models). E-commerce retail: median 22%, 75th percentile 39%, top decile 56%. Automotive: median 26%, 75th percentile 42%, top decile 60%. Finance: median 19%, 75th percentile 33%, top decile 51%. Travel: median 28%, 75th percentile 45%, top decile 61%. Industrial manufacturing: median 14%, 75th percentile 28%, top decile 44%.

These figures refer to citation rate over the standardised matrix, averaged across all six LLM systems. Model-specific variations are substantial and should be reported separately in operational reports.

Quality assurance and known limitations

Quality assurance runs through four mechanisms. First, all automation is calibrated quarterly against a manually coded sample of 500 prompts. Where deviations exceed 5 per cent, the automation is revised. Second, code review of the extraction and classification code by external reviewers, documented in internal review reports. Third, outlier analysis of implausible measurements (citation-rate jumps of more than 20 percentage points within a week are flagged as anomalies and investigated manually). Fourth, quarterly re-calibration of the attribution heuristics against updated model behaviour patterns.

Known limitations include: (a) attribution for tier-3 indirect references remains subjective and is scored conservatively. (b) LLM hallucinations can produce false citations that cannot be distinguished methodologically from correct citations — the hallucination-rate metric addresses this, but imperfectly. (c) Personalisation in real user scenarios is stronger than in isolated measurements — actual citation rates may diverge. (d) The benchmark focuses on publicly accessible LLM interfaces; internal enterprise copilot contexts are not covered.

Openness and replication

The benchmark framework is explicitly designed as an open reference framework. Methodology documentation is shared with research partners and agencies on request, including prompt-matrix templates, classification heuristics and statistical analysis code. The hope is that the market develops a shared measurement basis over the coming year, analogous to the historical evolution of established SEO metrics (organic impressions, CTR) — with clear operational consequences for board reporting and investment decisions.

Operationalising it for your own brand

For brands that want to use the benchmark framework productively, onboarding is structured in three phases. Phase one (months 1–2): construct your own prompt matrix, adapt it to your industry, collect a baseline across two 7-day windows. Phase two (months 2–4): implement interventions — entity work, passage engineering, schema graph — and run a post-intervention measurement. Phase three (from month 4): continuous weekly monitoring with drift alerts for citation-rate declines beyond defined thresholds.

Structurally this matches our LLM Citation Monitoring service, but it can also be built with comparable tools and internal resources. The key point is methodological consistency — one-off measurements without repeatability are anecdote, not benchmark.

Conclusion: benchmark as discipline

The value of a benchmark lies not in individual numbers but in disciplined repeatability. Brands that measure their AI-search visibility systematically against a standardised framework build a dataset that holds up over years — and that stays robust against model updates, industry shifts and organisational change. The framework documented here is a proposal, not a prescription. Our hope is that independent replication and constructive critique will improve it over time, and that the market will converge on a shared measurement basis that puts board communication and investment decisions on a more rational footing.