Technical Methodology

Contents

• How CiteMetrix Queries AI Platforms
• Scan Pipeline Architecture
• Platform Coverage
• ModelScore™: The Composite Visibility Metric
• Brand Facts & Accuracy Detection
• Branded vs. Unbranded Intelligence & Share of Voice
• BYOK Architecture: Data Sovereignty & Key Isolation
• Remediation: From Finding to Fix to Verification
• Scope, Boundaries & Honest Limitations

How CiteMetrix Queries AI Platforms

CiteMetrix does not scrape AI chat interfaces. Every query is submitted programmatically through each platform’s published API. This matters for data quality, reproducibility, and terms-of-service compliance: the data is exactly what the platform’s API returns, captured in full, with no screen-scraping fragility in between.

Query execution

• Queries are submitted via authenticated API calls using each account holder’s own API keys (the BYOK model — see BYOK Architecture).
• Each query is sent as a fresh, isolated, single-turn message with no prior conversation context — identical to what a first-time user would see.
• Responses are captured in full: the complete response text, any source citations the platform returns, and metadata including the model used and a response timestamp.
• Queries run server-side in a managed pipeline — no browser, no tab, no manual interaction.

Query strategy: five intent tiers

CiteMetrix does not treat all queries as equivalent. It uses an intent-stratified query model, where each tier exposes a different layer of AI visibility:

Tier	Intent type	Example	What it reveals
1	Direct brand	“What is [Brand]?”	Raw brand knowledge — does the AI know who you are?
2	Branded category	“[Brand] vs. competitors”	Competitive framing — how the AI positions you against the field
3	Unbranded category	“Best [category] options”	Discovery gap — are you surfaced when not named?
4	Problem / solution	“How do I solve [pain point]?”	Consideration stage — does the AI route the problem to you?
5	Fanout	Sub-questions a model decomposes a complex query into	Hidden gaps — where competitors surface in implicit sub-queries

Scheduling and versioning

• Automated scans run on user-defined schedules; manual scans can be triggered on demand from the dashboard.
• Every result is timestamped and versioned — each scan is a point-in-time record, so visibility can be tracked as a trend rather than a single snapshot.
• Because AI visibility can shift with each model update and indexing cycle, the trend line is the product. A one-time snapshot is a single data point; CiteMetrix captures the movement.

Scan Pipeline Architecture

The scan engine runs as a set of managed cloud services, not inside the web application. This separation is deliberate: anything that makes a paid API call or runs longer than a few seconds executes in the pipeline, while the web application handles only the dashboard, configuration, and billing.

The pipeline is a two-stage, queue-decoupled design:

• Scheduler — a scheduled job identifies which domains are due to scan, and fans out one work unit per (query × platform) pair onto a queue.
• Fetcher — each work unit triggers exactly one AI-platform API call. The full raw response is stored before any interpretation happens.
• Analyzer — a second stage reads the stored raw response and derives the citation fields (cited or not, sentiment, position, mentions), marking the record complete.

Why one API call per work unit matters

Each work unit carries a unique idempotency key enforced by a database uniqueness constraint. If the same (query × platform) unit is ever delivered twice — a normal occurrence in any distributed queue — the second attempt cannot create a second paid call. Duplicate billing is therefore structurally impossible, not merely guarded against. This is verified behavior, not an aspiration.

Resilience

• A failure on one platform or one query does not affect the others — each work unit is independent.
• A work unit that fails transiently is retried; one that fails repeatedly is set aside for inspection rather than lost or retried forever.
• Raw responses are preserved, so analysis logic can be improved and re-run later without paying for the API calls again.

Platform Coverage

CiteMetrix monitors nine AI surfaces. Eight are queried directly through their providers’ APIs; Google’s AI Overview is captured through a search-data provider because it is a search-results feature rather than a chat API.

Platform	Access	Notes
Perplexity	Perplexity API	Returns source citations natively — the cited URLs are captured
ChatGPT (OpenAI)	OpenAI API	Web-search-enabled model; configurable version
Claude (Anthropic)	Anthropic API	Used both as a monitored platform and as the analysis engine
Google Gemini	Google Gemini API	Search-grounded responses with citations
Microsoft Copilot	OpenAI-compatible API	GPT-family model served via Microsoft’s stack
Grok (xAI)	xAI API	Web and X/Twitter grounded responses
DeepSeek	DeepSeek API	Direct API
Mistral	Mistral API	Direct API
Google AI Overview	Search-data provider	AI Overview detection and citation capture where present

Additional platforms can be registered through the Platform Registry — an internal configuration layer where a new API-accessible model can be added with its own endpoint, authentication, and request format without re-engineering the pipeline.

ModelScore™: The Composite Visibility Metric

ModelScore is a composite AI-visibility score from 0–100. It combines four weighted components into one number that reflects not just whether a brand is cited, but what drives or suppresses that visibility. The weights below are the values published on the CiteMetrix methodology page and applied uniformly across accounts today.

Component	Weight	Primary data source	What it measures
Citation Score	45%	CiteMetrix scan data	How often the brand is cited across platform and query combinations
Brand Demand	20%	Google Search Console (OAuth)	Branded search demand — the signal AI systems respond to
Authority Transfer	20%	GA4 / Adobe Analytics (OAuth)	Referral sessions actually arriving from AI platforms
Technical Readiness	15%	CiteMetrix crawler + PageSpeed API	How well the site is structured for AI access and indexing

Citation Score carries the largest weight deliberately: cross-platform citation is the signal most directly tied to AI visibility and the one least available from any other tool. The weighting is a stated methodology choice, not a hidden formula — it is published, and it is the same for every account.

Why a composite, not a single number

A brand can be cited often yet see little referral traffic (high Citation, low Authority Transfer). Another can have strong branded demand that AI platforms are not amplifying (high Brand Demand, low Citation). A single citation-frequency number hides these patterns; a weighted composite with named, separately-reported components surfaces which lever to pull first. Each component is reported on its own as well as in the composite.

Brand Facts & Accuracy Detection

CiteMetrix includes a structured accuracy-verification system that compares AI-platform responses against a verified library of brand statements — the Brand Facts library.

How Brand Facts work

• Brand Facts are structured, categorized assertions about a brand: company information, product detail, pricing, certifications, legal statements, competitive positioning.
• Facts are entered manually, imported, or AI-suggested from public brand content, then reviewed.
• Facts can be marked critical, and critical facts are prioritized in every accuracy check.
• Facts are versioned — when one changes, the prior version is retained with its timestamp for audit purposes.

The accuracy check

On each scan, AI-platform responses are compared against the Brand Facts library using a language model as the comparison engine. The comparison is semantic rather than literal — it is designed to catch paraphrased inaccuracies, implied falsehoods, and misleading omissions, not only verbatim mismatches. When a discrepancy is found, the system records the platform, the triggering query, the verbatim AI claim, the conflicting verified fact, an issue type, a severity rating, and a recommended remediation.

Daily re-verification

Open accuracy issues are re-checked on a schedule. When a platform later corrects its response — because the model updated or a content change was indexed — the issue is resolved and the resolution is timestamped. The result is an auditable trail: when an issue was detected, what was done, and when it was confirmed resolved.

Branded vs. Unbranded Intelligence & Share of Voice

CiteMetrix separately tracks citation performance on branded queries (those naming the brand) and unbranded queries (category, problem, and solution queries where the brand is not named). The distinction is operationally important and routinely overlooked.

Branded citation rate tells you whether AI knows who you are. Unbranded citation rate tells you whether AI recommends you to people who do not yet know you — usually the more important number, and usually the lower one. A brand with a high branded rate and a low unbranded rate is effectively invisible at the consideration stage, where new customers are formed.

What is reported

• Branded vs. unbranded citation rate, per platform and in aggregate.
• The specific unbranded queries where the brand is absent — and which competitors are cited there instead.
• Share of Voice: the share of AI responses mentioning the brand versus its configured competitors, by platform, by query tier, and in aggregate.
• Sentiment and framing: whether the brand is cited positively, neutrally, or negatively, and whether it is recommended, listed as an alternative, or used as a comparison point — tracked over time, since a sentiment shift often precedes a citation-rate change.

BYOK Architecture: Data Sovereignty & Key Isolation

CiteMetrix uses a Bring Your Own Key (BYOK) model. Each account supplies its own API keys for each AI platform. Keys are stored encrypted (AES-256-CBC), and scans execute using the account holder’s own keys — never a shared pool.

What this means in practice

• Per-account key isolation is enforced and verified. A given account’s scans use only that account’s keys. If an account lacks a key for a platform, that platform is skipped — the system does not fall back to any shared or administrative key. This was confirmed by direct test: a scan for an account holding a key for one platform but not another used the held key and skipped the other entirely.
• Because the API call is made with the account’s own key, the usage bills to that account’s own provider relationship. Billing follows the key by construction.
• Accounts retain ownership of their scan data and can export it. Keys can be rotated at any time without losing historical data.

Security posture

• AES-256-CBC encryption on all stored API keys; the encryption key is held in a managed secrets store, separate from the database.
• Role-based access control (Owner, Admin, Analyst, Viewer) with permission-gated key management.
• Two-factor authentication (TOTP plus backup codes).
• Comprehensive audit logging of user actions and data-access events.
• SOC 2: controls and documentation are in place; a formal third-party audit is planned and has not yet been completed. This is stated plainly rather than implied.

Remediation: From Finding to Fix to Verification

Monitoring identifies problems; CiteMetrix is built to act on them. Findings connect to specific remediation tools, and the daily re-verification described above confirms whether a fix actually changed what the AI says.

Tool	What it does	Addresses
Content Optimizer	Analyzes pages against the queries they fail to be cited on; recommends specific changes	Citation
FAQ Generator	Generates content targeting unbranded queries with zero citation rate	Citation
Schema Advisor	Recommends structured-data schema based on actual page content, not URL patterns	Technical
llms.txt Generator	Builds and maintains an llms.txt file	Technical
E-E-A-T Audit	Evaluates content against Experience, Expertise, Authoritativeness, Trustworthiness	Citation / Authority
Brand Alignment Auditor	Checks on-site content against Brand Facts for internal contradictions	Citation
Content Brief Generator	Produces briefs targeting specific unbranded citation gaps	Citation
On-Page SEO Analyzer	Technical analysis oriented to AI indexability	Technical
AI Content Audit	Bulk evaluation of existing content against AI-citation criteria	Citation / Technical

The loop is closed: a gap is identified, a tool produces the fix, and re-verification confirms whether the platform’s response changed. Identifying a problem and acting on it are treated as one workflow rather than two products.

Scope, Boundaries & Honest Limitations

A briefing meant for a skeptical reader should be as clear about boundaries as about capabilities. The following are stated deliberately:

• It is API data, not panel data. CiteMetrix measures what the platforms’ APIs return for the configured queries. It does not claim to observe what real end-users are privately asking AI assistants; it does not publish proprietary query-volume data.
• Citations are first-turn and context-free by design. Results reflect a fresh single-turn query. Multi-turn conversational behavior is a different thing and is not represented as such.
• The accuracy engine uses a language model to compare. Semantic comparison is far more capable than string matching, but it is an AI evaluating AI; findings are surfaced for human review, with the verbatim claim and the conflicting fact shown so a person can verify.
• SOC 2 is documented, not yet audited. Stated in the security section and repeated here intentionally.
• Some scoring inputs depend on customer integrations. Brand Demand and Authority Transfer require the customer to connect Search Console and analytics. Without them, those components are reported as unavailable rather than estimated.

Everything described on this page either runs in production today or is identified as planned. Where a number or weight appears, it matches what the live product applies. Where a capability is on the roadmap rather than shipped, it is labeled as such.