How our SEO audit
actually works — under the hood.

Data sources

• Firecrawl — full markdown + rawHtml for every page (headers/footers included so NAP and JSON-LD survive).
• DataForSEO OnPage — bulk task crawl: pages, duplicate tags, internal links, redirect chains, non-indexable.
• Parallel content-analyzer subagents — fanned out for content QA in batches of ~30 pages.

Methodology

• Bulk-only on the happy path — two async jobs do the heavy lifting; no per-page calls until exceptions surface.
• A Python "ground-truth" step picks the homepage, contact, and location pages, extracts canonical practice name/phone/address/doctors, and asks the operator to confirm before flagging anything. Wrong ground truth = false positives everywhere.
• Pre-flight pattern checks against rawHtml: heading hierarchy, Google Maps URL resolution (must land on a real GBP CID, not an address pin), iframe form privacy/terms link inspection, deep JSON-LD validation, hostname canonicalization (catches Kinsta/WP Engine leaks).
• Cross-page Python pass detects cannibalization (exact title/H1 duplicates, neighborhood-doorway templating) and internal-linking issues (orphans, missing footer links, broken links).

Rubric contribution

Not directly scored. Produces crawl_pages.json + ground_truth.json that every other sub-audit consumes. Findings feed Cat 3 (On-Page) and Cat 1 (Content) indirectly.

Data sources

• Ahrefs — site-explorer-organic-keywords top 1,000 by volume, per domain.
• DataForSEO Labs — google_domain_intersection with intersections: false flag = "rival ranks, we don't".
• Search Atlas — se_analyze_keyword_gap + se_get_keyword_gap_results (async submit-and-poll, cross-validation).

Methodology

• Pulls our top 1,000 organic keywords and each rival's top 1,000 from all three providers; subtracts our set from each rival's set — what remains is the "gap".
• Merges results across providers and dedupes. A keyword present in 2–3 sources is high-confidence; single-source gets a caveat.
• Scores each gap keyword as highest_volume × rival_position_inverse (position 1 → ×100, position 50 → ×51, position 100+ → ×0). Rewards both volume AND rival proximity to page 1.
• Buckets the top 100 by topical category (braces, invisalign, ortho_general, info) via deterministic keyword-token rules — no LLM guessing.
• Cross-references against existing internal link graph to tag each gap as coverage: partial (rewrite existing page) vs coverage: none (build new).

Rubric contribution

Cat 1.3 Cluster Content (max 50 pts). Computes a coverage_pct and maps to a delta: ≥90% coverage = 0; <30% = −20 pts off baseline.

Data sources

• crawl_pages.json from prior site-review (required, ≥5 pages).
• DataForSEO Labs — google_ranked_keywords for cannibalization confirmation (Phase 2).
• Search Atlas GSC — gsc_get_keyword_performance for canonical Google truth (Phase 3).

Methodology

• Cannibalization detection: normalize each page's title + H1, strip brand and geo tokens, cluster pages with identical phrases. Confirm via DFS (≥2 URLs ranking top 50 for same keyword) AND GSC (≥2 URLs accumulating impressions for the same query). GSC is the strongest signal — it's Google literally telling us two pages compete.
• Vendor boilerplate: hash every sentence ≥50 chars across the site. Sentences appearing on ≥2 pages indicate template reuse. Flag pages with <40% unique sentences AND >50% boilerplate score.
• Freshness: extract dates from JSON-LD first, then OpenGraph, then visible text. Bucket: fresh (≤6mo), current, stale (>12mo), untraceable.
• Pillar architecture: build URL tree, identify Mother pages matching ground_truth.services, count Sons and Grandsons. Compare to ideal (each Mother → 2–5 Sons → 1–3 Grandsons each).

Rubric contribution

Cat 1.4 Service Pages (40) + 1.5 Location Pages (40) + 1.6 Content Velocity (25) + 1.7 E-E-A-T (20) = 125 pts. Also emits cross-category cannibalization evidence to Cat 3.

Data sources

• Pure synthesizer — zero new measurements. Reads all the other audit envelopes plus ground_truth.json and crawl_pages.json.

Methodology

• Generates three candidate types: NEW_PAGE (from content-gap keywords, missing pillar Mothers/Sons, competitor top pages), UPGRADE (page-deep-dive SCHOLAR <60, cannibalization winners, buried pages with topical value), CONSOLIDATE (cannibalization losers, stale vestigial orphans).
• Each candidate gets a type-specific impact score. NEW_PAGE blends keyword volume (log-scaled), winnability (inverse of rival position), pillar gap severity, and parent Mother's PageRank percentile.
• Sequences into phases with dependencies: Phase 1 (90d) = all consolidations + top 3 upgrades + top 3 new pages. Phase 2 (180d) = remaining upgrades + Son/Grandson buildout. Phase 3 (365d) = depth content. A Son can't ship before its Mother.
• Generates a deployable brief for each item: title, slug, primary/secondary keywords, target word count, required schema types, H1/H2 skeleton, FAQ prompts, internal-link plan, E-E-A-T requirements.

Rubric contribution

Evidence only — informational under Cat 1. The output is the deliverable; it doesn't change the score.

Data sources

• DataForSEO OnPage task crawl — task post, summary, pages, duplicate tags, redirect chains, non-indexable, links.
• DataForSEO Lighthouse on top 10 pages — Performance, SEO, Accessibility, Best Practices + Core Web Vitals.
• Ahrefs Site Audit — independent error/warning verification.
• Firecrawl JS-render diff on 3 pages — detects hydration issues where content only appears post-JS.

Methodology

• One task crawl, then ~5 GET calls retrieve the whole site (versus 150 sequential page hits).
• Per-page checks.* boolean flags are the authoritative issue source — no inference, no LLM guessing.
• Severity is rule-based: Critical = 404/5xx, missing title, missing H1, HTTP on HTTPS site. Warning = duplicate tags, thin content (<300 words), missing alt, canonical mismatch, orphan. Info = LCP/INP, render-blocking, no schema.
• v3.13.0+ meta-conventions pass validates every title (50–60 char band, no YMYL superlatives, no "near me", no year), every meta description (120–160 chars, CTA before mobile cutoff, no phone numbers), and Title↔H1 token similarity ≥80%. State-board awareness layer flags FTC/state-dental-regulator risk for YMYL violations.

Rubric contribution

Full Cat 5 Technical (75 pts): 5.1 Schema (25), 5.2 CWV (25), 5.3 Crawlability (15), 5.4 Security/HTTPS (10). Also feeds Cat 3 sub-scores 3.1 (25) + 3.2 (20) + 3.4 (25) = 70 pts via title/meta/alt-text data.

Data sources

• crawl_pages.json from site-review (no re-crawl).
• DataForSEO — on_page_instant_pages in parallel batches of 5; on_page_lighthouse on top 5.
• Firecrawl as JSON-LD fallback when rawHtml is missing.

SCHOLAR framework — 7 letters × 0–10 each, scaled to 0–100

• S — Specific: title / H1 / URL contain the primary service from ground truth + a tracked city.
• C — Conversational: ≥3 question patterns, FAQ schema present, 2nd-person voice ≥5 occurrences, dl/details blocks.
• H — Helpful: ≥2 CTAs ("book", "schedule", "consult", "call"), price or before/after content, descriptive alt text, clear next-step section.
• O — Original: practice-specific signals (doctor names, local landmarks), low vendor-boilerplate ratio (<0.6 of phrases like "state-of-the-art technology"), internal link to a related Mother page.
• L — Long: word-count tiers (≥400 / ≥800 / ≥1500) plus well-formed H2/H3 hierarchy.
• A — Authoritative: doctor credentials (DDS, DMD, AAO), Author/Person schema, outbound link to .gov/.edu, Reviewed/Updated date <12 months.
• R — Relevant: title keyword matches a tracked rank-tracker keyword, no sibling cannibalization, semantically-aligned inbound anchors, schema type matches content.

Interpretation: 80–100 pillar-ready · 60–79 needs work · 40–59 rewrite candidate · <40 failing.

Rubric contribution

Cat 1.3 Content Quality (30) + 1.4 Content Depth (30) · Cat 3.5 Internal Linking (25) + 3.6 Schema Coverage (25) = ~110 pts.

Data sources

• crawl_pages.json rawHtml — regex-extract every <script type="application/ld+json"> block.
• DataForSEO — onpage_pages.checks.has_json_ld / has_microdata for the "which pages have schema" baseline.
• Firecrawl fallback when rawHtml is missing.
• Ground truth NAP + GBP data from audit_map-pack.json for cross-validation.

Methodology

• Parse every JSON-LD block; validate JSON syntax, required fields per @type, and correct subtype (Dentist/Orthodontist array form, not bare LocalBusiness).
• Determine expected schema by page_type: homepage → LocalBusiness + Organization; service → MedicalProcedure; provider → Person/Physician; location → LocalBusiness per office; blog → Article + Person author; FAQ section detected → expect FAQPage.
• Cross-reference NAP three ways: schema-vs-visible content, schema-vs-GBP, visible-vs-GBP. Mismatches go to the action list.
• Stale/vendor schema detection: out-of-date phone, wrong address, generic vendor strings injected by old plugins.
• @id cross-reference validation — every {"@id": "..."} must resolve to an entity in the @graph; broken refs are removed, never emitted.

Rubric contribution

Cat 1.5 Schema Completeness (20) + Cat 2.4 GBP-Website Alignment / Schema NAP (25) + Cat 5.1 Schema Markup (25) = ~70 pts.

Data sources

• Search Atlas Local SEO Heatmaps (primary) — list_businesses, list_businesses_heatmaps, get_heatmap_details, single_competitor_versus_report.
• DataForSEO — serp_organic_live_advanced with location_coordinate for cross-validation at the address center.

Methodology

• Match each practice location to a SA heatmap business by address/lat-lng (within ~100m).
• For each (location × tracked keyword), pull the existing grid — every cell has a GPS coordinate and the practice's rank at that point. Typically 39 cells per grid.
• Compute four KPIs per grid: average rank, top-3 share %, the largest concentric radius where median rank ≤3 (proximity radius), and dead-zone clusters (≥3 contiguous cells where rank >20).
• Cross-validate with DataForSEO at the address center — if the practice doesn't appear at its own front door, SA data is suspect.
• Optional rival comparison runs single_competitor_versus_report against the primary rival domain.

Rubric contribution

Cat 2.2 GBP Heatmap Performance (75 pts): 2.2.1 avg rank (25), 2.2.2 top-3 share (25), 2.2.3 proximity radius (15), 2.2.4 dead-zone coverage (10). Returns null if no heatmap project exists — "data unavailable" is not "failed."

Data sources

• Search Atlas — gbp_list_citation_submissions + gbp_get_aggregator_details (covers ~30–50 directories via 5 aggregators: Data Axle, Localeze, Foursquare, Factual, Acxiom).
• DataForSEO — business_data_business_listings_search, google_my_business_info, google_reviews.

Methodology

• Path A (preferred): read prior SA citation submissions if they exist.
• Path B (fallback): initialize a citation draft populated from ground_truth.json and submit a fresh scan; poll up to 5 minutes.
• Normalize every found listing: phone to E.164, address with abbreviation collapse, name with legal-suffix stripping, website with host-only.
• Compute Levenshtein distance per field vs canonical NAP — matches must be exact or ≤2 character edits (single-typo tolerance for name/address; exact match for phone/website).
• Bucket each mismatching listing into name / phone / address / website / hours discrepancies.

Rubric contribution

Cat 8 (25 pts total): 8.1 aggregator presence (10), 8.2 NAP consistency % (10), 8.3 niche directories — AAO / ADA / Healthgrades / Zocdoc (5).

Data sources

• Search Atlas GSC integration — 9 tools including gsc_get_site_property_performance, gsc_get_keyword_performance, gsc_get_page_performance, gsc_compare_performance.
• Search Atlas KRT (Keyword Rank Tracker) — daily SERP-position history beyond GSC's 16-month window.
• DataForSEO Labs — google_ranked_keywords for cross-validation of estimate accuracy vs GSC truth.

Methodology

• Hard gate: gsc_get_sites must return the practice domain. If GSC isn't connected, the audit returns status: error — it deliberately does NOT fall back to estimates.
• Pulls two 30-day windows in parallel (current and prior) for period-over-period delta.
• Computes branded vs non-branded split by matching keywords against practice name + doctor surnames from ground truth.
• Identifies "winnable" keywords: position 4–10 with ≥50 impressions — small CTR/meta tuning lifts them to page 1.
• Computes mean absolute position delta between DFS estimates and GSC reality for top 50 keywords — feeds confidence intervals on the other audits.

Rubric contribution

Does not own a category. Emits a cross_category block that the composite scorer overlays onto Cat 1.7, Cat 3.6, and Cat 6.3 — plus the SA Holistic Technical pillar feeds Cat 3 and Cat 5.

Data sources

• Search Atlas — gbp_list_locations (enumerate verified GBP profiles) + gbp_get_review_stats per location (total reviews, weighted rating, reply count, rating distribution).
• Optional: gbp_list_reviews for individual review pulls when negative-handling drill-down is needed.

Methodology

• List all GBP locations under the practice's SA account; flag any unverified locations as warnings.
• Per location, pull review stats and compute reply rate (replied / total).
• Weights aggregate rating by review volume across locations — a 5.00 location with 118 reviews shouldn't dominate a 4.90 location with 405.
• Flag any location with reply rate <90% as a "reply gap"; flag any location with ≥5 one-star reviews as needing sentiment investigation.

Rubric contribution

Full Cat 7 Reviews (50 pts): 7.1 review volume (20), 7.2 avg rating (20), 7.3 review management/response rate (10). Cross-category contribution to Cat 2.5 GBP Review Velocity.

Data sources

• Ahrefs Site Explorer (primary) — referring-domains history, all-backlinks top-500, anchors, DR history, URL Rating history, broken backlinks.
• DataForSEO Backlinks API (cross-validator) — backlinks_summary, backlinks_anchors, backlinks_timeseries_new_lost_summary, backlinks_bulk_spam_score.

Methodology

• Pulls 12 months of referring-domain history from both sources; computes new / lost / net velocity at 30 / 90 / 365 day windows.
• Classifies every anchor (weighted by referring domain, not raw link count) into five buckets: brand, exact-match, partial-match, naked URL, generic.
• Flags a referring domain "toxic" if DFS spam score ≥30, OR if DR ≤5 with exact-match anchor (anchor-bait pattern), OR if it's on a low-trust TLD with DR ≤10.
• Reconciles Ahrefs vs DFS — if they disagree by >50% on net velocity or >15pp on any anchor bucket, logs a warning and surfaces both numbers.

Rubric contribution

Owns sub-criteria 4.1–4.5 in Cat 4 (90 of 150 pts): Domain Power proxy (20), ref-domain count (20), velocity (20), anchor diversity (15), link quality/toxicity (15).

Data sources

• Ahrefs batch-analysis (primary) — one call returns ref-domain lists for practice + up to 3 rivals, aligned for diffing.
• DataForSEO — backlinks_bulk_referring_domains for cross-validation.

Methodology

• Resolves rivals from ground_truth.json or, failing that, from DFS google_competitors_domain. Zero rivals → fail fast.
• Normalizes every ref-domain, builds sets per target, computes gap = union(rivals) – practice.
• Scores each gap domain: acquisition_value = DR × rival_count × topical_relevance. Topical relevance: 2.0 for medical/dental, 1.7 for local-news (*-times.com, *.patch.com), 1.5 for chamber/rotary/community/.org, 1.0 otherwise.
• Sorts, keeps top 30 for the deliverable. Buckets targets into medical / local-news / community / general.

Rubric contribution

Cat 4.6 Authority gap closure pipeline (60 pts): pipeline depth (30), topical relevance ratio (15), tier mix of DR50+ targets (15).

Data sources

• Search Atlas Site Explorer v2 (primary) — holistic SEO scores, brand signals, organic keywords, organic pages, position distribution, indexed pages.
• Ahrefs — top-pages, DR history, traffic history.
• DataForSEO Labs — google_competitors_domain for rival selection cross-validation.
• FireCrawl — light crawl of rival homepage + top-5 pages for architecture / schema / blog signals.

Methodology

• Auto-detects the rival if not specified: top DFS competitor with >50 keyword intersections and avg position <30 (skipping aggregators like Healthgrades, WebMD).
• Pulls rival's SA holistic scores; pulls the same for the practice; computes deltas across topical authority, brand signal, top-10 page traffic, indexed-page count, blog cadence.
• Extracts cities-served from each domain's top-200 organic keywords — surfaces "cities the rival ranks for that we don't" for market-expansion strategy.

Rubric contribution

Diagnostic only — does not own a sub-criterion. Emits rubric_contributions.evidence_only.rival_diagnostics which the composite scorer overlays onto Cat 1 and Cat 4 evidence. The skill explicitly will not invent a category score; it informs, doesn't grade.

Data sources

• None external. All math runs locally in Python on crawl_pages.json. Sub-30 seconds on small sites, 1–3 minutes on a 300-page site.

The 5 algorithms — in plain English

• Internal PageRank. Every page starts equal. On each iteration, a page hands a share of its authority to every page it links to. After ~100 passes the numbers stabilize — you have a ranking of which internal pages are most "trusted" by the site's own link structure. Top of the list should be homepage + Mother pillars. If not, the link graph is misallocating equity.
• Click-depth. Walk outward from the homepage in breadth-first order: depth 1 = one click away, depth 2 = two clicks away. Anything at depth ≥4 is buried — Google re-crawls those less often and weights them lower.
• Anchor distribution. For every target page, classify each incoming anchor (brand / exact-match / partial / generic / naked URL). Targets where exact-match >50% are over-optimization risks (Penguin territory). Where generic ("click here") >60% the anchor carries no semantic signal to Google.
• Cluster cohesion. For each Mother service, find the Mother page and its Son pages. Ask: of all possible Son-to-Son pairs, what fraction actually cross-link? Score 0.0 (isolated silos) to 1.0 (every sister links to every sister). Healthy ortho clusters: 0.4–0.7.
• PR concentration. What % of total PageRank do the top 5 pages hold? 40–60% is healthy. >70% means the site is so homepage-heavy nothing else can rank. <30% means equity is so diffuse no page is strong enough to win competitive keywords.

The audit then proposes the top 15 high-leverage link moves — specific source page → target page → suggested anchor — ranked by source_PR × (1/target_PR) × topical_relevance × priority_bonus. "This page has lots of authority to spare, this other page is starved despite being important, and they're topically related — add the link with this anchor."

Rubric contribution

Cat 3.5 Internal Linking (50 pts): orphan rate (12), click-depth (12), anchor distribution (8), cluster cohesion (10), PR concentration (8). Replaces the simpler "inbound link count" metric the older page-deep-dive skill used.

Data sources (triangulated)

• Search Atlas Brand Radar (primary, 8 endpoints) — mentions overview + history, share-of-voice overview + history, cited-pages, cited-domains, impressions overview + history.
• Ahrefs Brand Radar (cross-validation, 7 endpoints) — AI-response samples, mentions, SoV, cited-pages/domains.
• DataForSEO LLM Mentions (third triangulation) — aggregate metrics, searchable mention list, top cited pages/domains.
• DataForSEO SERP Live — AI Overview parsing on 5–10 focal queries (e.g., "best orthodontist in Albany", "Invisalign Albany").

Methodology

• Pulls a 30-day window from all three providers in parallel (90 days if a recent rebrand).
• Triangulates mention counts; if sources disagree by >50%, takes the median (no single-source scores).
• Detects share-of-voice by comparing the practice to a resolved rival domain (from ground_truth.json or auto-detected via SA's organic competitors, filtering out aggregators).
• Parses each focal SERP's ai_overview.references[] to count how many AI Overviews actually cite the practice's domain.
• Classifies every cited URL by path (homepage vs service vs location vs about) so the rubric can reward depth, not just homepage hits.
• Computes 30-day mention slope (last-7-day avg vs first-7-day avg) for trajectory.

Key concepts in plain English

• Mention share = what % of LLM responses for queries in your space include your brand name at all. 50% means an LLM mentions you in half the answers it generates for in-category prompts.
• Share of voice = of all brand mentions in those responses, what % are yours vs competitors'.
• AI Overview citations = Google's AI Overview shows ~3–8 source links per answer; this counts how many of your focal queries cite your domain as a source. This is the AI-era equivalent of position 1.

Rubric contribution

Drives Cat 6 GEO in full — 50 pts: 6.1 Mention Presence (15), 6.2 SoV vs Rival (10), 6.3 AI Overview Citations (10), 6.4 Cited-Pages Quality (10), 6.5 Sentiment & Trajectory (5).

Data sources

• Search Atlas only (brand-routed) — KRT projects + keyword details, Local SEO heatmap businesses + heatmaps, LLM Visibility projects + queries/topics, GBP citation submission history, OTTO engagement, GBP location detail.

Methodology

• Detects the practice's vertical (orthodontics, dermatology, periodontics, chiropractic, etc.) and generates a candidate keyword set from the per-vertical YAML template — what should be tracked.
• Pulls actual tracked state from Search Atlas across KRT, heatmaps, LLMV, citations, OTTO, GBP.
• Diffs candidate vs actual: flags gaps (recommend add), excess (consider remove), matches (good).
• Runs 10 anti-pattern checks against the STANCE rules (word-order over-tracking, near-me in KRT, mobile-desktop doubling, cross-city tracking, missing brand, high NR rate, radius-market mismatch, multi-specialty in one project, stale heatmaps, no citation history).
• Builds a credit-costed action list (rough $/month per add or remove) so operators can weigh tradeoffs.

What "STANCE" is

Not an acronym — shorthand for the agency's source-of-truth doc "Agency Tracking-Setup Standards — Canonical Stance" (setup/research/STANCE.md), which codifies 10 numbered rules compiled from Whitespark, Sterling Sky, BrightLocal, Local Falcon, SearchAtlas docs, and a 41-project NEON/HIP audit. Headline rules: canonical word-order {service} in {city} {state}, near-me → heatmaps not KRT, one KRT project per location, mobile is source of truth, heatmap radius bounded by ~20-min drive-time.

Rubric contribution

Evidence-only — does NOT move the V5 score. STANCE compliance is methodology hygiene. Writes evidence rows into Cat 2 / Cat 6 / Cat 8 for internal review only. Never appears in client-facing strategy presentations.