Docs / The growth loop

AI Visibility: the score, the layers, and the methodology

Wavefront's Brand Visibility Audit answers one question: when someone asks an AI engine about your category, do you show up, described correctly, and recommended with confidence? It scores your workspace 0-100 across five layers and turns every problem into a finding with evidence and a fix route.

The five layers

Each layer answers a distinct diagnostic question, and the fixes belong to different teams. That's why layers have owners, not just scores.

Discoverability

AI engines can only cite what their crawlers can fetch and parse. The layer measures AI crawler access (training, search-index, and live-retrieval crawlers are three separate decisions with separate user-agent strings: declare your policy and the audit holds robots.txt to it, in either direction), content extractability (AI crawlers fetch JavaScript but don't execute it), query fan-out coverage (engines silently expand each query into sub-queries with zero measurable search volume; full audits map your top keywords automatically), freshness (statistics/news pages age out at 90 days, high-traffic or AI-cited pages at 180, evergreen at 12 months, and a lastReviewed stamp from a review-without-rewrite counts), and metadata quality.

Clarity

The layer judges everything against your canonical brand description: the one paragraph you'd sign, plus category, value proposition, target customer, and the "what you are NOT" list. Brand-prompt answers are diagnosed as wrong (third-party sources contradict you: fix the sources) versus missing (the fact isn't anywhere AI can retrieve: publish it). Your owned surfaces (homepage, LinkedIn, Crunchbase, review profiles) are scanned monthly for drift, with a corrected paragraph ready to paste. The semantic neighborhood covers the vector side: an approved list of buyer-language terms whose catalog coverage is measured deterministically; uncovered terms feed topic suggestions and add a modest bonus to keyword opportunity scores. generate_org_schema builds the Organization JSON-LD (canonical description + sameAs links to every registered profile), and the audit verifies it's installed and complete.

Authority

Roughly 84% of AI citations come from earned media, and each category has a small recurring core of sources AI actually trusts. The Citation Core is built quarterly from engine-attested citations only, with a frequency-by-reachability priority grid (1 = quick win, 4 = honestly not worth chasing). The finding that matters most: a core domain that cites your competitors but not you.

Trust

Inclusion is not endorsement. Comparison answers classify as endorsed, hedged, not recommended, or absent, with the hedging language quoted verbatim; hedge direction is the leading indicator. Every hedged answer is paired with the sources that produced it: a hedge with named sources is an outreach task, a hedge without sources is a content task, and the finding says which. Business-critical prompts alert when they stay non-endorsed for two consecutive scans (single-scan flips never alert: AI answers are volatile). Claim credibility measures substantiated statistics across your recent articles: the same structural check the draft QC enforces, so what's measured is what's produced.

Agentic readiness

Fourteen deterministic checks: HTTPS, agent-bot access, descriptive link anchors, actionable conversion points, labeled forms, policy/contact/pricing discoverability, hygiene signals like the copyright year, and your catalog's task-intent mix. If a site only informs, the agent completes the task on a site where it can act.

Methodology (the honest part)

• Scores are 0-100 per layer, equally weighted overall. Every layer score decomposes into named components with fixed weights; the breakdown is always visible.
• No data never means zero. A component without data says so, and its weight redistributes across the layer's measured components. A brand-new workspace sees "no data yet: here's what to connect", not a fake failing grade.
• Citations come only from engines that cite. The model-panel proxy has no real sources and writes nothing to the citation ledger; citation intelligence requires live-retrieval scans.
• Classification is enum-validated and degrades visibly. When the classifier returns something unusable, the result is labeled unclassified, never guessed, and unclassified scans neither fire nor suppress alerts.
• Volatility is designed for. Scores are smoothed over consecutive audits, regressions must hold across two scans, and the per-engine consensus view shows when engines simply disagree (they cite from largely disjoint corpora; divergence is normal).
• Every recommendation carries an evidence grade: best practice (grounded in retrieval mechanics and research), bet (sound logic, low downside, evidence still forming), or watch (actively discussed, too thin to act). We tell you which is which.

What happens with findings

Findings are work items, not reading. Content-fixable findings land in your Review Inbox as proposals the assistant can write; technical, surface, and outreach findings name the team that owns them (and the roadmap document phases everything 30/60/90). Re-running the audit auto-resolves anything it no longer detects, so fixes become visible without bookkeeping.

Common questions answered

Why did my score move when we changed nothing? AI answers are volatile run to run. The smoothed score damps single-audit swings, and the engine-consensus row shows whether one engine diverged. Sustained movement across two audits is signal; a single wiggle usually isn't.

Why is a layer showing "no data"? The honest state: that component's input isn't connected yet (a domain, a scan, competitors, or a canonical description). The component note says exactly what to add.

Can Wavefront fix my robots.txt or LinkedIn page? No, and we won't pretend otherwise. The audit generates the exact robots.txt snippet and the corrected surface paragraph; a human pastes them. Content fixes, though, the assistant writes end to end.

How long until fixes show up in AI answers? Technical fixes (crawler access, extractability) take effect on the next crawl: days to weeks. Clarity and trust fixes depend on third-party sources refreshing: weeks to months. The trend lines exist because this is a compounding practice, not a switch.