AI visibility isn’t magic. It’s a method.

What Ask Hercules measures

Ask Hercules measures four different things and keeps them permanently separate. Blending them is the most common failure in AI visibility reporting, because each one carries a different level of proof.

• Grounded citation. The model ran its own live web search and cited a URL on your domain. This is the only class that counts toward your Citation Score.
• Ungrounded mention. The model named your brand from training memory, with no retrieved source behind it.
• Inferred recommendation. The model produced a ranked list of what it would recommend and named you in it. Useful signal, not a citation.
• No mention or provider error. A clean absence is recorded as a gap. Provider errors are excluded from both the numerator and the denominator of every score.

Alongside those response-level classes, the platform scores two structural layers: the Crawl Readiness Index™ (what your own site does) and the KSM™ pillars (how the outside world sees your brand). Those two are never merged into a single number either.

Crawl Readiness Index™

The CRI™ is an on-site, crawl-only scorecard. It is computed entirely from the pages we crawl on your domain — no external APIs, no third-party data, no model opinion. It answers one question: when an AI crawler reaches your site, is the evidence it needs actually extractable?

• Answerability — do pages answer real questions in extractable blocks rather than burying the answer in narrative?
• Visibility — is the substance present in the served HTML, or only after JavaScript runs?
• Trust Readiness — are authorship, entity identity, schema and verifiable claims in place?
• Crawl Health — status codes, canonicals, robots rules, sitemaps and duplication.
• Topical Cohesion — do pages cluster into a coherent topic and link to each other in a way a retriever can follow?

The CRI™ is a readiness score, not a visibility score. A perfect CRI™ does not mean an AI system cites you; it means nothing on your own site is preventing it.

KSM™ framework

Where the CRI™ looks inward, the KSM™ (Knowledge Structuring Model) looks outward. It scores how the wider information environment — language models, knowledge graphs and citation surfaces — currently represents your brand, across three pillars.

• Structured Extractability (E) — can your knowledge be lifted cleanly out of your public material and reused in an answer?
• Entity Salience (S) — is your brand resolvable as a distinct entity across independent reference sources?
• Citation Authority (C) — do retrieval systems actually reach for your material when answering questions in your category?

Each pillar is evidence-gated. If a pillar does not clear its minimum evidence threshold, the platform reports insufficient evidence and the pillar is excluded from the overall score rather than being filled in with a guess.

Citeability Engine™

Measurement without remediation is a dashboard. The Citeability Engine™ is the part of the platform that turns a diagnosed gap into a concrete, publishable change: a structured answer asset, corrected schema, an entity fix, a knowledge-base entry or a page-level rewrite.

• Every answer asset follows the same structure — definition, failure modes, evaluation criteria and FAQ — because that structure is what retrieval systems can lift.
• Structured data is generated, validated against the rich-result requirements and checked for the common failure of schema that only exists after JavaScript executes.
• Changes can be published through WordPress, an embed, an API push or an API fetch, so the fix lands on the live site rather than staying a recommendation.
• Remediation is human-in-the-loop. Nothing is published to a customer property without approval.

The engine is deliberately opinionated. It does not generate volume; it produces the specific evidence that a missing citation requires.

Hercules Circuit

The Hercules Circuit is the operating cadence that keeps the work compounding. It is a five-state weekly loop with a required artifact at each step, enforced by the software rather than by discipline.

• Monday — Select the question. One live buyer question per workspace. Focus is enforced, not suggested.
• Tuesday — Build the answer asset. A reusable unit of authority in the standard four-part structure.
• Wednesday — Define activation. The checklist, scorecard or assessment that converts attention into value.
• Thursday — Capture proof. At least one piece of real proof is mandatory before the week can advance.
• Friday — Reinforce and synthesise. Hubs updated, proof injected into the archive, metrics snapshotted, next week unlocked.

There are no backward transitions once a step is closed. The cadence is the product: it encodes the method into software so the loop survives a busy week.

How Audit, Fix and Track relate

The three phases are one circuit, not three products. Each phase produces the input the next one needs, and Track feeds back into Audit so the next cycle starts from evidence rather than opinion.

AUDIT                 FIX                    TRACK
CRI™ crawl scorecard  Citeability Engine™     Multi-model measurement
KSM™ pillar scoring   answer assets, schema   grounded vs. ungrounded
                      entity + page fixes     position and share of voice
   │                        │                        │
   └── gaps ────────────────┘                        │
                            └── published evidence ──┘
                            ┌───────────────────────┐
                            └─ measured change ─────┘  → next audit

• Audit establishes where you are missing and why — on-site readiness and off-site representation, scored separately.
• Fix removes the specific evidence gap the audit named, and publishes the change.
• Track re-runs the same questions against the same models and reports whether the change moved anything, with the raw response kept as proof.

Measurement methodology

Question set

Measurement starts from the questions your buyers actually ask, derived from your website and ICP rather than from keyword volume. Each question is scored for relevance to your category, and the relevance threshold is exposed as a control so you can see exactly which questions are in and out — and why each excluded one was excluded.

Models queried

Each question is asked of multiple frontier systems concurrently — ChatGPT, Claude, Gemini and Perplexity — using each provider's own web retrieval where it exists: Perplexity's native retrieval, Gemini's search tool, the OpenAI Responses hosted web search and Claude's web search tool. Providers are queried independently, so one failing provider degrades the run instead of voiding it.

Provenance

Every model result stores the provider, model identifier, retrieval mode, returned source URLs, the exact prompt sent, sampling settings, locale, timestamp and the raw answer text. Any score can be traced back to the response that produced it.

Scoring

For each question, Citation Score is the share of answering models that returned a grounded citation to your domain; Mention Score is the share that named you without a retrieved source. Scan-level scores are the mean of the per-question scores. Provider errors are removed from the denominator so an outage cannot read as a decline.

Repetition

Single runs are anecdotes. Scans can be scheduled to repeat on a fixed cadence, and every run is stored so a diff view can show what changed since the previous run — score movement, new wins, new gaps and changes to the excluded-question set.

Grounded citations versus ungrounded mentions

This distinction is the core of the measurement truth layer, and the reason our numbers are sometimes lower than tools that report a single "cited" figure.

One hard rule governs the whole system: a URL that we place into a prompt can never produce a citation. If our own context supplied the link, the retrieval mode is recorded as injected context and the result is disqualified from the Citation Score. The scan is not allowed to feed itself.

Limitations and model variability

We publish the limits of the method because a measurement you cannot interpret is worse than no measurement.

• Non-determinism. The same question asked twice can return different answers. Trends across repeated runs are meaningful; a single run is a sample.
• Model and index change. Providers update weights, retrieval indexes and safety behaviour without notice. A score movement can reflect a provider change rather than a change on your side.
• Retrieval availability. Grounding tools are not always available or triggered. When a provider answers without searching, the best possible result for that response is an ungrounded mention.
• Context differences. Locale, account history, personalisation and prompt phrasing all affect answers. Our runs are unpersonalised and will not match every individual user's experience.
• Sampling. Scores describe the question set that was measured, not the entire universe of possible questions.
• No guarantees. No vendor controls frontier models. We improve the signals you control and report honestly on the ones you do not.

Research sources and revision history

Primary sources

• KSM™ (Knowledge Structuring Model) — ksmmodel.ai, Dr. Anthony Q. Bowen.
• Provider retrieval documentation for OpenAI, Anthropic, Google and Perplexity web search tooling.
• Schema.org vocabulary and Google's structured-data requirements, used for rich-result eligibility validation.
• Ask Hercules internal specifications: the Citeability Engine specification, the measurement truth layer and the CRI™ scoring definition.
• Related reading: the loop-based growth whitepaper, our AEO compliance statement and the field notes on AI visibility.

Revision history

Material changes to scoring or classification are recorded here with the date they took effect.

See the method run against your brand.

Audit what AI systems currently say about you, fix the evidence holding you back and measure whether it moved.

7 days free · No credit card required