The Citation Report / Model Intelligence

How Pendium calculates your AI visibility score

When a prospective buyer asks a conversational assistant for software recommendations, traditional search engine rankings offer zero insight into whether your brand makes the shortlist. Pendium calculates a brand's AI visibility score by executing structured prompt matrices across seven major conversational platforms, including ChatGPT, Claude, Gemini, and Perplexity. Rather than counting simple keyword mentions, the system parses the semantic context of each generated response against specific buyer criteria to determine direct recommendations. The resulting data weights third-party citations, earned media, and competitive displacement to generate a unified recommendation metric that shows marketing teams exactly which commercial queries they win.

Scanning the seven major AI engines

A single language model cannot represent the broader generative search market. A brand might dominate answers in ChatGPT while remaining invisible in Claude, or appear prominently in Perplexity citations while getting filtered out of Google AI Overviews. To build an accurate baseline, the Pendium AI visibility platform conducts automated scans across seven core engines:

  • OpenAI ChatGPT (GPT-4o and reasoning-focused variants)
  • Anthropic Claude (Sonnet and Opus architectures)
  • Google Gemini (Advanced and search-grounded models)
  • Perplexity (Online search index integration)
  • xAI Grok
  • DeepSeek
  • Google AI Overviews (Enterprise search SERP integrations)

Capturing this data requires systematic query orchestration rather than simple web scraping. Pendium interfaces directly through dedicated API pipelines and its Model Context Protocol environment. By running specialized diagnostic commands like scan_visibility across an array of more than 80 purpose-built tools, the system queries each model with standardized parameters.

AI Platform Primary Retrieval Method Citation Handling Scan Focus
ChatGPT Search index plus internal weights Markdown footers and inline tags Conversational shortlists
Claude RAG integration and long-context analysis Inline textual attributions Architectural comparisons
Gemini Google Search index integration Direct web links and Knowledge Graph Live market data
Perplexity Multi-source real-time search synthesis Prominent numbered references Source verification
Google AI Overviews Core web index and SGE modules Structured snippet carousels Top-of-page zero-click share
Grok Real-time social data and web feeds Dynamic inline links Immediate trend velocity
DeepSeek Deep reasoning token processing Structured synthesis tags Complex logic matching

To prevent caching anomalies and personalized search bias, each scan isolates prompt delivery. The system strips user history, resets contextual threads, and normalizes temperature settings across engines. This isolates the actual retrieval behavior of the model. By inspecting raw conversational output through programmatic endpoints outlined in the Pendium agent integration, the engine catalogs how each model surfaces vendor alternatives for identical enterprise queries.

Measuring context rather than just mentions

Counting the raw volume of brand mentions is an outdated practice carried over from traditional social listening tools. In conversational AI search, an unranked brand mention carries very little commercial value. A model might mention your platform as a historic legacy tool, cite you as an expensive alternative, or bundle your name into an indiscriminate twenty-vendor list. None of those instances qualify as a commercial recommendation.

Pendium evaluates the semantic role your company plays inside each generated answer. The platform separates output text into four distinct classification tiers:

  1. Primary Recommendation: The model explicitly names the brand as a top-two choice for the primary task.
  2. Qualified Alternative: The model suggests the brand only under specific edge cases or constrained budgets.
  3. Neutral Co-mention: The brand appears in a broad list with no evaluative preference.
  4. Deprecated Reference: The model cites the brand as outdated, difficult to use, or unsuitable for the prompt's criteria.

This parsing relies on natural language processing pipelines that evaluate sentence syntax, entity proximity, and conditional clauses. When an assistant states, "Vendor A is the ideal choice for SOC 2 compliance, whereas Vendor B lacks native audit trails," the system assigns positive recommendation equity to Vendor A and negative equity to Vendor B.

An open laptop sitting beside a contract document, pen, and potted plant on a white desk.

Mapping answers to buyer personas

Enterprise software purchases rarely follow a single evaluation path. A Director of Security evaluates software on compliance, data residency, and cryptographic isolation. A VP of Growth looks for implementation speed, automated workflows, and CRM synchronicity.

The evaluation framework within Pendium AI Visibility tests how models respond when prompting from distinct buyer archetypes. The system runs identical functional queries through varying buyer lenses:

  • Enterprise compliance officers prioritizing auditability and data governance
  • High-growth startup operators demanding rapid time-to-value
  • Mid-market procurement teams focusing on seat pricing and support contracts
  • Technical implementation architects reviewing API limits and documentation depth

If Claude consistently recommends your platform for small engineering teams but excludes you whenever an enterprise buyer specifies single sign-on constraints, the scoring model reflects that gap. Visibility is calculated as a distribution across buyer segments, not a flat global average.

Identifying the questions you win

Winning an AI recommendation requires understanding the structure of user intent. Pendium categorizes prompts into commercial intent tiers, isolating the exact categories where a brand holds clear authority:

  • Direct replacement queries: "What can replace Competitor for enterprise billing?"
  • Feature-dependent questions: "Which marketing analytics tools support real-time warehouse sync?"
  • Greenfield discovery prompts: "What are the best customer data platforms for healthcare?"
  • Constraint-driven filters: "What are the top open-source alternatives to Tool?"

By clustering the outputs of these queries, the platform records your win rate per question type. The system identifies whether you capture broad discovery searches or strictly niche edge-case prompts, showing your team exactly where brand messaging has succeeded and where assistants default to competitor recommendations.

Weighting third-party citations and earned media

Conversational engines do not invent facts; they synthesize data retrieved through local context windows and vector indexes. To understand why an assistant selects one vendor over another, you have to track the source documents retrieved during the generation process.

Large language models prioritize authoritative, un-gated third-party validation over self-published corporate claims. According to the Muck Rack Generative Pulse report (Q4 2025) referenced in research on building a B2B brand identity that AI engines cite, over 85% of non-paid AI citations come from third-party editorial coverage, not the brand's own website. The model treats vendor websites as biased marketing material, relying on external editorial analysis to verify operational claims.

Pendium tracks external citations by extracting the real URLs referenced by retrieval-augmented generation modules across Perplexity, Gemini, ChatGPT, and Google AI Overviews.

Car dashboard touchscreen display showing a GPS navigation map while driving at night.

Tracking the sources AI actually trusts

The scoring mechanism breaks citations down into distinct source classes, assigning weight based on authority and historical frequency:

  • Independent technical publications and niche industry journals
  • Peer-review platforms and software evaluation registries
  • Top-tier financial and tech press
  • Community discussions on specialized engineering and operations forums
  • First-party vendor documentation and product pages

When an engine constructs an answer, citations from high-authority independent trade publications carry higher weight in the visibility calculation than product landing pages. If an assistant cites an independent comparative review confirming that your platform handles data processing faster than an alternative, that citation provides a strong positive signal in the scoring engine.

Competitor displacement metrics

A core function of the Pendium metric is measuring competitor displacement. This evaluates how often your product displaces an established competitor when buyers run comparative head-to-head prompts.

The displacement ratio measures three distinct data points:

  1. Direct preference share: The percentage of head-to-head queries where the engine recommends your brand over a named rival.
  2. Negative displacement: Queries where the user asked about your product, but the model redirected them to a competitor.
  3. Feature parity wins: Answers where your product is credited with specific technical capabilities that competitors lack.

Measuring displacement allows brands to move past raw visibility tracking. It exposes the exact qualitative friction points where an AI assistant advises prospective buyers to look elsewhere.

Calculating the unified recommendation score

The final output is a normalized 0-to-100 metric: the Pendium AI Visibility Score. Rather than an unweighted average of mentions, this score represents a mathematical synthesis of contextual placement, source validation, and query win rates.

AI Visibility Score = (R_dir * W_r) + (C_auth * W_c) + (P_cov * W_p) + (D_comp * W_d)

The model applies four core weighted factors:

  • Direct Recommendation Rate (R_dir, 40%): The percentage of category queries where the assistant places your platform in the primary recommendation position.
  • Citation Authority and Distribution (C_auth, 30%): The quality, domain diversity, and editorial credibility of external URLs cited by the models.
  • Persona and Query Coverage (P_cov, 15%): The breadth of unique buyer personas and search intent tiers where the brand earns recommendations.
  • Competitive Displacement Ratio (D_comp, 15%): The win rate achieved in direct comparison prompts against core market competitors.

To understand how to read and act on these numbers, teams can review the tactical breakdown in what to look for in an AI visibility score. The score updates dynamically as new scans run, providing a continuous timeline of performance.

When visibility shifts, the platform correlates the movement against real-world changes. If a new high-authority comparative review goes live or an agency deploys optimized technical documentation, the system marks that event directly on the historical trajectory chart. Teams can see the exact date an external article was indexed and how many points that single publication added to their Claude or ChatGPT recommendation rates.

Scan data moves directly into operational execution. Through tools like get_recommendations and complete_action on the platform's MCP server, engineering and growth teams turn score gaps into prioritized tasks. The software highlights specific third-party publications that require outreach, missing technical specs that models struggle to confirm, and target questions where competitors hold uncontested ground.

To establish your baseline visibility across ChatGPT, Claude, Gemini, and Perplexity, you can test your brand metrics instantly. Run a free visibility scan preview at Pendium.ai to inspect your current recommendation share, review cited sources, and identify the exact conversational queries your brand needs to win.