Buyer FAQ
A. How accurate is a Snapshot?
Three things are independently auditable, and none of them depend on trusting the analysis node.
- Raw response fidelity. Every datapoint is the verbatim LLM response, with a timestamp and a raw JSON reference in the delivered
raw-data-index.mdand evidence bundle. You can inspect the package and confirm that the report's claim matches the raw response. Verifiable, not estimated. - Mention and coverage math is deterministic. The per-endpoint and per-category counts are computed by the orchestrator from the raw responses, not regenerated by an LLM. The same raw JSONs always produce the same numbers. Reproducible.
- Interpretation is locked to 6 rules. The analysis node is constrained by methodology.html §9 (no "best", no "guaranteed", no invented citation URLs, no competitor rankings, no percentage headlines, no missing Data Provenance footer). A violation forces a re-draft before delivery.
The AI-readable Markdown brief's YAML frontmatter carries a confidence: high | medium | low field, computed by the orchestrator from endpoint completion. The score is deterministic and reproducible; it is not generated by an LLM. We do not claim guaranteed accuracy in the sense of 100% right about your future rankings. We claim verifiable raw responses, reproducible math, and a fail-closed delivery QA gate. For the full breakdown, see methodology.html §11.
B. Which model endpoints are covered?
Today, Snapshot covers 8 production model endpoints, all measured through vendor-hosted APIs: Doubao, Kimi, DeepSeek, Qwen, ChatGPT, Grok, Claude, and Gemini with Google AI Overviews. Public-weight availability is disclosed only where it applies to specific model families or checkpoints: DeepSeek and Qwen have public-weight model families; Kimi has specific public checkpoints such as Kimi K2, but Clarivy does not claim every Kimi API model is open-weight.
All 8 run via documented API methods with the exact model id, data-retention posture, and training-control posture recorded per datapoint. This is an API audit, not a product-surface audit: we measure what each underlying model returns on a controlled prompt, not what a signed-in user sees inside the ChatGPT or Gemini app. See methodology.html §6 for what is in scope and what is not.
Why not cover every model? Adding an endpoint is a real commitment: API access, data-retention controls, training-control posture, DPA terms, a tested adapter, a reproducible request payload, and a per-datapoint raw JSON to keep the methodology contract honest. We will not silently swap endpoints that lack a reviewed data-control posture or a tested adapter: that would force customers into a privacy posture they did not sign up for.
The current production list does not include ERNIE, MetaSo, or Perplexity Sonar; the historical Day-1 self-audit at /audit/self-audit-01.html used a broader experimental matrix and is kept for transparency. "All AI models" is not a stable target: new models appear weekly, vendors retire models without notice: so adding an endpoint is a versioning event with a methodology changelog entry, not a marketing update. For the current production matrix and the model-freshness policy, see methodology.html §1 and §7.
C. What do I actually receive?
You receive three core files plus an Agent Bundle Manifest:
- A 1-page human PDF for leadership.
- An AI-readable Markdown brief for your internal AI agents or your Obsidian vault: with stable finding_id, evidence_id, and action_id identifiers, so an AI can summarize without paraphrasing.
- A raw-data index that links to every LLM response for independent verification.
- An Agent Bundle Manifest for automated workflows and agent adapters.
Every report also includes deterministic GEO scores: AI Citation Readiness and Source Trust Baseline. For repeat customers, Clarivy can load sanitized repeat-audit memory so future audits compare against prior runs without storing contact details, billing data, raw responses, or raw JSON in memory. Opt-out and deletion are available.
Every claim in the human PDF cites a raw JSON line. A sanitized, fictional sample is at /audit/sample/: the numbers in that sample are mock, the deliverable shape is real. We publish it so you can inspect the structure before requesting scope. Snapshot is one input among many; it is not a substitute for user-research interviews, conversion analytics, an SEO backlink audit, or a brand-tracker survey. See methodology.html §4 and §11.5 for the full delivery package and what "decision-grade" means in practice.
D. Is a Snapshot enough?
A Snapshot is enough if you want to risk-screen a known query, spot-check a category entry, or baseline a brand before a larger engagement. It is not enough as a category-wide audit; its cover page states "quick diagnostic, not a complete audit".
Enterprise Monitor is the primary recurring workflow when procurement, multi-market coverage, multilingual work, trend tracking, or custom evidence requirements matter. Enterprise Audit remains available when a one-time enterprise assessment is preferred. Professional is suitable when a team needs implementation-ready findings and source mapping. Enterprise and Professional are quoted after scope review. Snapshot is a fixed-scope, one-time USD 149 self-serve purchase.
The audit is one input. It is not a substitute for user-research interviews, clickstream / conversion analytics, an SEO backlink audit, or a brand-tracker survey. You decide how to combine it with the rest of your evidence. See methodology.html §5 for what Snapshot does and does not cover, and methodology.html §11.5 for what "decision-grade" means for a buyer.
E. Is the audit process professional?
Treated as a regulated analytics deliverable, not a marketing artifact. Three concrete gates:
- Analysis node is locked. methodology.html §9 lists the 6 non-negotiable rules (no "best" / "only" / "guaranteed" / "always" about Clarivy or competitors; every claim cites a raw JSON; no percentage in headlines; no invented citation URLs; no ranking of competitors; Data Provenance footer on every report). A violation forces a re-draft before delivery.
- Delivery QA is fail-closed. methodology.html §8 lists 5 gates (PDF render, AI-readable Markdown brief schema, raw-data count, evidence reference check, partial-rate gate at > 20% missing). A draft that fails any gate is held for human review; the customer is told, not silently sent a broken deliverable.
- Reproducibility is verifiable. Every report includes a delivered
raw-data-index.mdand evidence bundle for the order. You (or any auditor) can inspect the raw JSON references, replay the orchestrator inputs where permitted by scope, and diff the result.
The operating entity is documented on the Legal operator page. Refund terms are defined in the approved quote, order form, or invoice. We do not publish aggregateRating or customer-count schema on this site; we will not, until we have 50+ real customers with their own written consent to be cited.
F. After the audit, is there an action plan?
Yes. The AI-readable Markdown brief ships with prioritised actions and a 30-day implementation plan. Every action links to at least one finding_id and one evidence_id, so you can trace every recommendation back to a specific (endpoint, query) datapoint. Actions are ordered by lift-per-effort, not total lift, so a small marketing team can execute the top two in a week and see the next monthly run start to move.
The audit is one input. It is not a substitute for user-research interviews, clickstream / conversion analytics, an SEO backlink audit, or a brand-tracker survey. You decide how to combine it with the rest of your evidence.
Enterprise Monitor is the recurring version of this loop: monthly full audits by default, optional weekly pulse checks, and AI-agent files such as an action backlog, evidence map, verification plan, and memory delta. It can be prepaid or subscription-based only after the quote or order form confirms cadence, payment structure, and cancellation terms. We do not promise a specific revenue or traffic lift; the Princeton 2023 paper reports +40% visibility in a controlled setting, but individual results vary widely. The audit is an analytics instrument, not a guarantee of business outcomes. See methodology.html §11.5 for what "decision-grade" means in practice.
G. Why contact us for an audit scope?
Snapshot Baseline is fixed at USD 149 per one-time audit. Enterprise Monitor and Professional remain quote-based because markets, prompt depth, languages, cadence, delivery windows, AI-agent workflow needs, and procurement requirements vary. Their quote defines the prompt set, deliverables, invoice terms, delivery window, and refund boundary.
| Audit scope | State | Math | Datapoints | Delivery | Commercial path |
|---|---|---|---|---|---|
| Enterprise Monitor | Quoted after review | Monthly full audit; optional weekly pulse | Confirmed in quote | Human report + AI-readable brief + raw-data index + agent action backlog + verification plan as scoped | Contact us; prepaid or subscription; invoice and bank transfer supported |
| Enterprise Audit | Quoted after review | Custom one-time prompt set | Confirmed in quote | Human report + AI-readable brief + raw-data index + agent-ready files as scoped | Contact us; invoice and bank transfer supported |
| Professional | Quoted after review | Structured prompt matrix | Confirmed in quote | Human report + AI-readable brief + raw-data index | Contact us |
| Snapshot | USD 149, one time | 5 queries × 8 endpoints | 40 | Human report + AI-readable Markdown brief + raw-data index | Self-serve Stripe Checkout |
Enterprise Monitor does not create a subscription by default. A recurring record or recurring billing starts only after the customer approves cadence, payment structure, and cancellation terms in the quote or order form.
We do not put sales pitches inside the report. We do not advertise a "starting from" price. We do not advertise "12+" endpoints. We do not advertise aggregateRating or customer-count schema markup. We do not advertise SOC 2 or ISO 27001 until we have them; we say "not yet, target H2 2026 / Q4 2026" instead. See methodology.html §1 (the scope table) and methodology.html §5 (Snapshot's scope).
Related pages
- /audit/methodology.html: the technical contract: 8-endpoint matrix, core delivery package plus Agent Bundle Manifest, Snapshot vs Professional/Enterprise, API vs product surface, Model freshness policy, Delivery QA gate, the 6 anti-hallucination rules, §11 Confidence / accuracy / limitations.
- /audit/self-audit-01.html: Day-1 self-audit (historical dogfooding artifact; current production matrix is different). Kept for transparency.
- /legal/: Legal and trust pages for documents, compliance status, and public proof links.
- /enterprise/: Enterprise Monitor product page and sample package.
- /pricing/: Enterprise Monitor, Professional, and Snapshot baseline scope paths.
- /zh/audit/faq.html: the Traditional Chinese mirror of this page.
Changelog
- v1.0 (2026-06-15): initial release. 7 buyer questions (accuracy, decision-grade, model endpoint coverage, model freshness, professionalism, post-audit improvement, price).
- v1.1 (2026-06-16): historical P0 product-status reset. Superseded by v2.6.
- v2.6 (2026-06-22): Enterprise-first commercial reset. Removed public fixed Snapshot pricing, replaced Standard naming with Professional, and added waitlist-based Snapshot access while checkout is prepared.
- v2.0 (2026-06-16): P1 rewrite. Cut from the "we commit to" voice to plain customer-first language. Reframed the 3-file delivery as human PDF + AI-readable Markdown brief + raw-data index (the brief's full name replaces the earlier internal shorthand in the customer surface). Renamed "is the result safe to use as a commercial decision input?" to "is a Snapshot enough?" so the buyer-objection map is led by buyer questions, not by our engineering decisions.
- v2.1 (2026-06-16): P2 split. English-only content on this page; Traditional Chinese mirror moved to /zh/audit/faq.html. No copy change beyond the language separation.
- v2.2 (2026-06-17): Reframed coverage from language-labelled coverage wording to production model endpoints. Clarified that all tested models are multilingual.
- v2.3 (2026-06-18): Historical changelog: endpoint grouping was then framed as open-weight/source-available versus closed proprietary API model surfaces; v3.3 later tightened this to vendor-hosted API endpoints with separate public-weight availability notes.
- v2.4 (2026-06-19): Added deterministic GEO scores, sanitized repeat-audit memory, and recurring-monitor continuity language to match Privacy Policy v1.1, DPA v1.1, Terms v1.1, and Methodology v2.9.
- v2.5 (2026-06-21): Clarified the Agent Bundle Manifest delivery entry point, repeat-audit memory opt-out/deletion, and explicit approval requirements for recurring monitoring subscriptions.