Content · FAQ · GEO — Master

2026-07-22 · strategy

Content · FAQ · GEO — Master

Everything about content quality, the FAQ overhaul, AI-citation (GEO), and editorial authority. The FAQ overhaul is the active workstream (handoff: bowtie note #181). Full 552-item FAQ inventory lives in the generated faq-master-index.md (data appendix, regenerate with scripts/faq-index.mjs).

Contents

• FAQ Authoring Rules — The Binding Standard (build against this) — active
• FAQ Build Kit — /vpn/best-free-vpn/ (reference cluster) — active
• FAQ Corpus Audit — Fit vs Redo (Fable, 2026-07-22) — active
• FAQ System Overhaul — Audit + Target Spec — active
• GEO — Getting VPN.com Cited in AI Answers — active
• Authority Roadmap — Making "Reviewed by VPN.com" Mean Something — active
• Comparison Supercluster — Provider vs Provider Matrix — parked

---

FAQ Authoring Rules — The Binding Standard (build against this)

Consolidated from faq-authoring-rules.md · status: active

FAQ Authoring Rules — The Binding Standard

The rules every FAQ set is built and reviewed against. If a question or answer fails a rule here, it does not ship. This exists to prevent the one thing we will not do: generic questions padded onto a page for SEO volume.

Read alongside: faq-master-index.md (what exists), faq-corpus-audit.md (84% keep / provenance flag), faq-system-overhaul-spec.md (render + schema plan).

---

Rule 0 — The Tailoring Law (the one that matters most)

Every question must name the page's specific subject — its use case, niche, provider, or topic. A question that would read identically on ten different pages is banned.

The subject goes in the question, not just the answer:

Repurposable slots: max 2-3 per page, and even these are customized templates, never pasted verbatim. Allowed evergreen templates (each MUST take the page's [subject]):

• Does a free VPN work for [this specific use case]?
• How much does the right VPN for [use case] actually cost?
• Is using a VPN for [this use case] legal? — only if the page's topic isn't already owned by the are-vpns-legal page (see Rule 6).

If you can't make an evergreen slot page-specific, drop it and mine a real tailored question instead. We would rather ship 13 sharp questions than 20 with 5 generic ones.

---

Rule 1 — Structure: two tiers (internal buckets + user-facing labels)

Every FAQ carries two tags:

1. Internal intent bucket (for our coverage/balance QA — never shown to users).
2. User-facing topic label (what the visitor sees as the FAQ toggle/chip — "Speed", "Pricing", "Setup").

The user-facing label is the genericness test: if a question does not fit one clear, specific topic label, it is generic — rewrite it or cut it. "Is a VPN worth it?" fits no sharp label. "Does a VPN slow 4K streaming?" fits Speed. That difference is the whole rule.

Tier 1 — internal buckets (coverage check, hidden)

• 12-20 questions per page. Below 12 = incomplete. Above 20 = split or prune.
• Four buckets, 3-5 each — enforced for balance, not display:

A page with 15 research and zero outcome questions fails. New sets over-index outcome until the site balances.

Tier 2 — user-facing labels (display + genericness gate)

• Rendered under a single FAQ heading as topic toggles/chips: FAQ · Speed · Pricing · Technical · Money-back · Setup …
• Labels come from the controlled vocabulary below (per cluster) — not free-typed, so they stay consistent and schema-clean.
• Every question maps to exactly one label. No clear label → generic → cut (this is Rule 0 made operational).
• A label group shows when it holds ≥2 questions; a lone question folds into the nearest fitting label. Order labels by what the visitor cares about most for that page (Speed/Streaming first on a streaming page; Pricing/Money-back first on a provider page).
• The internal bucket still rides along on each item (for our QA + so we can verify all four buckets are covered across whatever labels a page uses).

---

User-facing label vocabulary (controlled — pick from these)

Labels are the visible toggles. Use only these; add a new label only if a real question genuinely fits none (and add it here). Each maps to a typical internal bucket so we keep balance while displaying topics.

VPN clusters (providers, best-X, use-case, hub)

Domain broker

Cybersecurity / Hosting

• Cybersecurity: How it works · Threats · Protection · Tools · Best practices
• Hosting: Hosting types · Performance · WordPress · Management · SEO

The vocabulary is the discipline: if you're reaching to force a question into a label, the question is generic — that's the signal to cut or sharpen it.

---

Rule 2 — Real intent (the question must be actually asked)

• Every question maps to demonstrated search demand: page-filtered GSC queries (get_search, filtered to the page — NOT the domain top-50, which is ~90% brand/navigational) + People-Also-Ask + real support/sales questions.
• Invented framings are banned (audit example of a REDO: "Why does upload speed matter more on iPhone than download?" — nobody searches it). If you can't tie a question to real demand, it's filler.
• Phrase questions the way a human types them, not the way we'd title a section. Question case, natural language.

---

Rule 3 — Answer quality (useful + data-backed + front-loaded)

• Front-load the direct answer in sentence 1 (LLM-citation format). Then the why, then the proof. 40-60 words is the citation sweet spot.
• Use specific, verifiable data — counts, prices, dates, jurisdictions, device limits, audit findings, cited studies with sources. "NordVPN ranks #1 of 22 we compare, audited no-logs (Deloitte 2023), 10 devices" — not "NordVPN is the best."
• Answer the objection honestly, including when the answer is unflattering ("No — CyberGhost ranks #10 of 22"; "Windscribe's 3-day guarantee is the shortest among majors"). Honesty is why the corpus scores 84%.
• Banned answer patterns: "availability varies, check the page"; "regularly / several times per month"; any pure assertion with no number when a number exists one click away (the hub-answer failure mode).

---

Rule 4 — Data provenance (hard gate)

We have real, citeable data — use rank, cite it honestly. The speedlab block in each src/data/providers/*.json + getSpeedLabData() gives every provider a rank out of 22 across speed/streaming/security/etc. That is the gold source for FAQ answers.

DO:

• Cite rank, pulled from the Speed Lab JSON: "NordVPN ranks #1 of 22 in our Speed Lab" / "CyberGhost ranks #10 of 22 for speed." Programmatic, true, defensible.
• Frame the Speed Lab as aggregated from independent labs, matching the published methodology.md: "across the independent labs we aggregate (Cybernews, Cloudwards, Security.org, et al.)."
• Use the other real data freely: deal counts ($75M+, 1,000+ domains), audit dates + auditor, jurisdiction, price, device counts, refund windows, cited third-party studies with the source.

DON'T (the two real liabilities the audit found):

• Never say "we tested" / "our tests" / "in our lab we measured." The methodology page explicitly states "We don't run our own speed tests — we aggregate." Those 14 answers contradict our own published methodology. Say "ranks #N in our Speed Lab (aggregated)," not "we tested at N Mbps."
• Don't cite a precise Mbps figure as a measured fact. The throughput ladder in the JSON is hand-authored to track rank (a smooth 730→320 staircase), not sourced medians — it won't survive "show me the source." Use rank + comparative language ("among the fastest we track") unless/until a number is reconciled to a named external source.

Net rule: "[Provider] ranks #[Y] of 22 for [Z], per our Speed Lab" is the sanctioned pattern — sourced from our JSON, honest about aggregation. Hard Mbps and "we tested" are out until the data is reconciled (separate task). New authoring uses rank; it does not add to the 43 Mbps / 14 "we tested" debt.

---

Rule 5 — LLM-optimized + schema (non-negotiable plumbing)

• Every FAQ set emits FAQPage JSON-LD (one emitter, Base.astro — no double emission; see the historic duplicate-schema bug).
• This extends to provider + pillar templates, which currently render FAQs but emit no schema. A page with FAQs and no schema fails.
• Answers must survive extraction: no answer that only makes sense with the surrounding prose; each Q/A is self-contained (an LLM will quote it alone).

---

Rule 6 — No duplicates (with the templating exception)

• True duplicate = banned: the same question answered on two pages splits schema/citation authority. When two pages want the same question, one page owns it (the most authoritative/specific) and the other links to it or asks a tailored variant. Known offenders to resolve: "public Wi-Fi?" (what-is-a-vpn vs most-secure-vpn), "charged during money-back guarantee?" (expressvpn vs cyberghost).
• Legit per-entity templating is allowed ("Does [Provider] keep logs?" across provider pages) — because the subject differs, so the answer differs. The dedup gate must distinguish "same question, different entity" (fine) from "same question, same scope" (dupe).
• Cluster authority: if a dedicated page owns a topic (are-vpns-legal), the hub/other pages defer to it rather than re-answering — a tailored angle or a link, not a copy.

---

Rule 7 — Policy alignment

• "Access," never "unblock/unlock" (3rd-party ToS risk — see feedback_access_not_unblock).
• Torrenting: the July 21 meeting pulled torrenting from homepage positioning. FAQ copy follows suit — no torrenting-forward framing; reframe or drop (vpnHubFAQ #5 is flagged EDIT for this).
• Recommendation #1 stays NordVPN, but the answer must earn it with a reason ("#1 of 22 for X"), never a bare "NordVPN is best."

---

Per-page-type MUST-have archetypes

Templates to customize, not paste. Each [bracket] is filled with the page's real subject. These are starting skeletons — mine real tailored questions to reach 12-20.

Provider review (e.g. nordvpn.md)

• Research: "What is [Provider]'s [signature feature] and does it matter?" (NordLynx / DAITA / Lightway)
• Objection: "Is [Provider] safe after [its specific incident]?" (Nord 2018, IPVanish 2016 logs, ExpressVPN/Kape) — or "Why does [Provider] cost more than [named competitor]?"
• Pre-purchase: "How many devices with [Provider]?" · "[Provider] refund window?" · "Does [Provider] access [the streaming service its buyers ask about]?"
• Outcome: "How do I confirm [Provider]'s kill switch is on [platform]?" · "What happens when my [Provider] plan expires?"
• Ranking anchor: "How does [Provider] rank among the VPNs we compare, and for what?" (real ranking, Rule 4)

Best-X ranking / use-case (e.g. best-vpn-for-gaming, best-vpn-for-streaming/netflix)

• Research: "Does a VPN [reduce ping / unblock the library] for [use case], and how?"
• Objection: "Won't a VPN [add lag / slow 4K] for [use case]?" (real tradeoff, honest)
• Objection (evergreen, customized): "Does a free VPN work for [use case]?"
• Pre-purchase: "Which VPN is #1 for [use case] and why?" (names it + tailored reason)
• Outcome: "How do I set up a VPN for [use case] on [platform]?" · "What if [the service/game] blocks my VPN?"

Hub / pillar (e.g. /vpn/, /cybersecurity/)

• Broader than a post, but still specific to the pillar and routing to money pages. Defer topic-owned questions to their dedicated pages (Rule 6). Pipe in the post-level data the hubs currently lack (audit: hubs are the data-poorest tier).

Domain broker (/domain-broker/*)

• Objection-led (already the site's best set). Back every objection answer with the data that exists in the posts ($75M+, 97% success, 30-100% offer uplift, proofSaved) — the hub versions currently assert without it.

---

Definition of Done (per page — the review checklist)

• 12-20 questions, 3-5 in each of the 4 internal buckets, outcome bucket present
• Every question maps to exactly one user-facing label from the vocabulary — no clear label = generic = cut
• Every question names the page's subject (Rule 0); ≤3 evergreen slots, each customized
• Every question tied to real demand (page-filtered GSC / PAA / support) — no invented framings
• Every answer front-loads the direct answer, uses verifiable data, handles the objection
• No banned answer patterns; no un-sourced performance numbers / "we tested" (Rule 4)
• No true duplicates vs other pages; per-entity templating only where the subject differs
• "Access" not "unblock"; no torrenting-forward framing
• FAQPage JSON-LD emits (incl. provider/pillar); each Q/A self-contained for extraction
• Label-grouped accordion render (FAQ · [labels]); label group only shown at ≥2 questions

---

Decisions — all resolved (2026-07-22)

1. Labels: user-facing topic labels from the controlled vocabulary (Speed / Pricing / Technical / Money-back …) under a single FAQ heading. Internal buckets stay hidden.
2. Evergreen allowance: max 2-3 per page, each customized to the page's subject.
3. Provenance: cite rank from the Speed Lab JSON (getSpeedLabData), framed as aggregated per methodology.md. No "we tested," no hard Mbps as measured fact (see Rule 4). Reconciling the existing 43 Mbps / 14 "we tested" answers is a separate cleanup task.

Rules are locked — ready to build.

Bottom line

The standard is tailored or it doesn't ship. Structure (12-20, four buckets, outcome-weighted), real demand (page-filtered GSC/PAA), data-backed honest answers (no un-sourced metrics), no true dupes, schema everywhere. Generic questions — even "does a free VPN work" in its bare form — are banned; the page's subject lives in the question. Build every set against the Definition of Done.

---

FAQ Build Kit — /vpn/best-free-vpn/ (reference cluster)

Consolidated from faq-buildkit-best-free-vpn.md · status: active

FAQ Build Kit — /vpn/best-free-vpn/ (the reference set)

Turnkey spec for the build worker. This is the reference implementation every other page copies. Build it against faq-authoring-rules.md + its Definition of Done. Questions are grounded in real demand (GSC page: 14,031 impressions/30d; PAA + Reddit/comparison SERPs for "best free vpn", "free vpn netflix", "proton vs windscribe free", "free vpn no data cap", "free vpn data limit").

Target: 16 questions (inside the 12-20 band). Current page has 5 → keep 4, edit 1, add 11.

Data sources the answers pull from (all verifiable — Rule 4 compliant)

• Speed Lab RANK (getSpeedLabData() / src/data/providers/*.json): hide.me #6 of 22 · Proton #12 · Privado #13 · Windscribe #18. Cite rank, framed as aggregated per methodology.md. No Mbps, no "we tested."
• Free-tier facts (providers' own public specs): Proton — unlimited data, open-source, independently audited, 5 countries, 1 device; Windscribe — 10 GB/mo, unlimited devices, unblocks Netflix US/UK/CA; Privado — 10 GB/mo, strongest streaming of the free tiers; TunnelBear — 2 GB/mo.
• The 72%-trackers stat (2024 study, already cited on best-vpn-for-android.md) — reuse with the same source.

The 16 questions

Legend: Label = visitor-facing toggle · Bucket = hidden coverage tag · Status = keep / edit / new.

Coverage check (Definition of Done)

• Buckets: Research 2 (#4,#6) · Objection 6 (#1,#2,#7,#8,#11,#16) · Pre-purchase 4 (#5,#9,#10,#13) · Outcome 4 (#3,#12,#14,#15). ✅ Outcome well-represented (fixes the site-wide ~10% gap).
• Labels used: Safety & Privacy (3) · Free vs Paid (4) · Streaming (2) · Speed (1) · Devices (1) · Setup (2) · Troubleshooting (2) · How it works (1). Each group ≥1; Speed/Devices/How-it-works are singletons — either fold into the nearest group at render, or add a 2nd question to promote them to their own toggle (author's call, both valid).
• Evergreen: 1 (#11), customized. Within the ≤2-3 cap.
• Every question names the subject (free VPN / a named free provider / a use case). No generic survivors.

Dedup guards specific to this page

• #11 must stay free-specific — do not reuse "Can a VPN protect me on public Wi-Fi?" (owned by what-is-a-vpn, and a true-dupe on most-secure-vpn per the audit).
• Streaming questions (#8, #9) are free-tier-scoped — the best-vpn-for-streaming cluster owns general streaming; keep these about the free constraint.
• "Are free VPNs safe" family recurs across provider pages — this hub owns the category answer; provider pages answer per-provider ("Is [Provider]'s free tier safe").

Render + schema (Phase 0 must be done first, or do inline here)

• Route through FAQ.astro accordion, grouped by the 8 labels under one FAQ heading.
• Emit FAQPage JSON-LD (this is a ranking template → already gets schema via [...slug].astro; verify each Q/A is self-contained after the extractFaqItems flattening fix).
• All 16 answers: front-load the direct answer sentence 1, 40-60 words, honest, data-backed per the sources above.

Handoff

Logged in bowtie note #181. Worker builds the answer prose + wiring; this kit fixes the questions, labels, buckets, data, and dedup so no invention is needed.

---

FAQ Corpus Audit — Fit vs Redo (Fable, 2026-07-22)

Consolidated from faq-corpus-audit.md · status: active

FAQ Corpus Audit — Verdict: FOUNDATION, not teardown

Adversarial quality audit of the 552-item English FAQ corpus (see faq-master-index.md) against Ben's rubric: real question + useful, data-backed answer + objection handling; no SEO-filler; no vague answers.

Headline (552 items)

Corroborated by the May 2026 site-wide rewrite (memory project_site_history: 1-3/10 → 7-9/10) which already did the quality pass. Answers front-load the direct answer, use numbers, handle objections honestly.

⚠️ The one hard flag — data provenance (cuts across KEEP)

43 FAQ answers cite Mbps figures and 14 claim "our tests / we tested" — and Speed Lab data is static/fake (memory project_speedlab_research). "CyberGhost ranks #10 of 22 VPNs we tested at 612 Mbps" is exactly the data-driven format the rubric wants, built on numbers we never measured. Structurally KEEP, ethically a liability, fatal if an LLM-citation strategy invites scrutiny. Resolve before the overhaul: get real numbers, attribute to external benchmarks, or strip the "we tested" claims. The 5 speed-test pages (16 items) are scored EDIT entirely for this reason.

Intent-bucket distribution (134 hand-scored items)

Outcome items that exist are strong (Mint "verify VPN is working" with real curl ifconfig.me; "why does my VPN disconnect on mobile data" with reconnect times) but concentrated on Linux/free-trial pages. Streaming, provider, hub pages have almost none.

Methodology

• All 24 hub FAQs read + scored individually.
• System A: 110 of 528 items (21%) read in full, stratified across every template family; per-family rates projected to family sizes.
• Site-wide duplicate scan (539 questions normalized, exact + provider-templated).
• GSC (askbowtie, 90d) used, but domain top-50 is ~90% navigational/brand (traffic reality), so demand validated at cluster level. Step 2 question-mining must use page-filtered GSC + PAA, not the domain top-50.

Concrete examples

KEEP — best-vpn-for-android "Are free VPNs safe on Android?" (2024 study, 72% had trackers); netflix "How does Netflix detect VPNs?" (names IP reputation/DNS mismatch/WebRTC + counters); nordvpn "safe after 2018 breach?" (objection met head-on); windscribe "why only 3-day guarantee?" (honest); sell-premium-domain "What if I already received an offer?" (floor-not-ceiling, 30-100% uplift).

EDIT — nordvpn-pricing "Accept gift cards?" ("availability varies, check the page" = non-answer); domainBrokerFAQ (all 5 — great questions, pure-assertion answers while $75M+/proofSaved data sits one click away); 5 speed-test pages (fabricated provenance); perksFAQ "how often new deals?" (vague); vpnHubFAQ "VPN for torrenting?" (off-policy per access-not-unblock + the homepage torrenting removal).

REDO — avast-secureline "What's the best VPN for Netflix?" (off-topic funnel, dupes streaming cluster — textbook filler, but only ONE found in 134 read); most-secure-vpn "public Wi-Fi?" (verbatim dupe of what-is-a-vpn); best-vpn-for-iphone "why upload speed matters more" (invented framing; salvage the data under a real question).

Model rewrites (vague → data-driven):

1. Broker hub "How do I know you'll create value?" → "You pay $0 unless the deal closes. Across $75M+ closed we've moved first offers 30-100%, incl. saving {{broker.proofSaved}} on our own $1M purchase." (data already in buy-premium-domain.md)
2. Hub "Can I use a free VPN?" → name the 3 trustworthy free tiers (Proton unlimited/audited, Windscribe 10-15GB, hide.me) + the 72%-trackers stat.
3. nordvpn-pricing gift cards → verify which countries/retailers/crypto and state it, or cut. "Availability varies" must never ship.

Duplicates

• True cross-page dupes: "public Wi-Fi?" (what-is-a-vpn + most-secure-vpn); "charged during money-back guarantee?" (expressvpn + cyberghost, verbatim); "charged after free trial?" (surfshark + protonvpn).
• Templated per-entity families (legit, but colliding phrasing the dedup gate must distinguish): does X keep logs ×7, is X worth it ×6, student discount ×6, is X trustworthy ×6, X+Netflix ×8, X in China ×4, etc.
• Cross-cluster near-dupes: "free VPN on/for [X]" ×~13; hub "Is a VPN legal?" vs the dedicated are-vpns-legal page (authority split); cyber hub "VPN vs antivirus" vs best-vpn-for-windows equivalent.

Bottom line

Build on it, don't tear it down. ~84% already meets the bar. Real deficits, in priority: (1) volume — every page under the 12-20 target, an expansion job; the questions to add are known (outcome bucket + GSC/PAA-mined); (2) data provenance — the 43 Mbps / 14 "we tested" claims from static Speed Lab, resolve before any LLM-citation push; (3) hub answers are data-poor vs the posts (broker hub worst); (4) ~83 one-sentence EDIT fixes. ~90% of overhaul spend belongs in net-new questions + data verification, not rewriting what exists.

---

FAQ System Overhaul — Audit + Target Spec

Consolidated from faq-system-overhaul-spec.md · status: active

FAQ System Overhaul — Audit + Target Spec

Meeting mandate (July 21): every cluster gets a real FAQ block — 12-20 questions, split across four intent buckets (research / objection / pre-purchase / outcome), no duplicates, LLM-optimized, schema, real human-intent questions as the staple. Plus a machine-page version (faq + citations + takeaway + metrics), a master FAQ page, a master index page, and an authority / citation-token evaluation.

Deliverable mode is spec/doc. This is the target design + phased plan for the implementing worker. Nothing has been changed.

---

1. Current state — the core problem is TWO FAQ systems

Consequences to fix:

1. Fragmentation — two authoring paths, two render styles, inconsistent schema. A reader/crawler sees accordions on hubs, flat text on posts.
2. extractFaqItems() flattens answers ([...slug].astro:175-191): answer = first paragraph only, multi-paragraph collapsed, markdown links passed as raw text into schema. LLM-quality answers need this fixed.
3. Provider (22) + Pillar pages have ## FAQ markdown that renders but emits NO schema — faqItems isn't passed to those layouts. Pure loss.
4. Hubs have only 4-5 FAQs — half the 12-20 target, no intent structure.

Existing infra to reuse (don't rebuild):

• rehype-citations.mjs — auto-harvests authoritative links → [n] superscripts + Sources section + citation JSON-LD. This is the citation engine for the machine page.
• public/llms.txt (hand) + public/llms-full.txt (generated, scripts/generate-llms-full.mjs) — the machine-index scaffolding for a "master index" already exists.
• "Bottom Line" blockquote convention (75 EN posts) — the "takeaway" the meeting wants, already a pattern; rehype-strip-bottomline.mjs strips it on ranking pages.
• src/data/providers/*.json → /data/providers/{slug}.json — machine-readable data precedent.
• Strategy already written: src/content/ops/geo-ai-citations.md (front-load answers, stat-rich, benchmark data cited 2.8× more), authority-roadmap.md.

---

2. Target model — one FAQ contract

2a. Unify authoring onto markdown + a structured convention

Keep FAQs in the content markdown (translatable through the existing pipeline), but make the intent buckets explicit so dedup, ordering, and the machine page can key off them. Two options — recommend Option A:

• Option A (recommended): intent-tagged ### questions. Keep ## Frequently Asked Questions + ### Q structure (zero new authoring syntax, existing translation flow untouched), and group under intent sub-headings the parser understands but that render cleanly:

  ## Frequently Asked Questions
  
  ### [research] How does a VPN actually hide my traffic?
  Answer…
  
  ### [objection] Isn't a free VPN good enough?
  Answer…

  Parser strips the [bucket] tag into a category field on each FAQ item; the visible question omits it. Fully backward-compatible — untagged ### questions default to research.
• Option B: frontmatter FAQ array. Structured faq: [{q, a, intent}] in frontmatter. Cleaner data, but breaks the current markdown-body translation path (frontmatter isn't translated the same way) and forces re-authoring 97 posts. Higher risk — only if we want FAQs fully decoupled from body prose.

Migrate the 5 hub FAQs (System B) OUT of translations.ts into the same markdown convention on their hub content, OR into a shared TS shape that the same renderer + schema path consumes. Goal: one render component, one schema emitter, one dedup check — not two.

translations.ts is a translation SSOT with contamination landmines (see CLAUDE.md § Translation SSOT). Moving hub FAQ objects touches it — pull up LayerView first (lv describe file:src/i18n/translations.ts vpn-astro), and run node scripts/i18n-contamination-check.mjs --gate after.

2b. Render everything through the accordion

Route the parsed faqItems through FAQ.astro (<details> toggles) on all templates — posts, provider, pillar, hub. Kills the flat-prose inconsistency. Group visually by intent bucket (optional sub-labels).

2c. Schema everywhere

Extend faqItems to the Provider and Pillar layouts (currently schema-less). Fix extractFaqItems() so answers keep full text (all paragraphs) and render markdown links as text without dropping them. One FAQPage JSON-LD emitter in Base.astro (already there) — just feed it from every template.

Guard the known duplicate-schema bug (domain-broker-recovery.md:23): only ONE emitter (Base JSON-LD). FAQ.astro must not re-emit microdata. Verify in GSC Rich Results after rollout.

---

3. The 12-20 / four-intent content model

Every cluster's FAQ set = 12-20 questions, roughly balanced:

Rules (enforce, don't just recommend):

• Real human-intent questions are the staple — mine from GSC (mcp__askbowtie__get_search top queries per cluster) + People-Also-Ask, not invented. This is the highest-leverage step; the questions ARE the moat.
• No duplicate FAQs across pages. Same question answered on two pages splits schema authority and confuses LLM citation. Needs a dedup gate (see §5).
• LLM-optimized answers: front-load the direct answer in sentence 1 (the geo-ai-citations.md finding), then the why, then a stat/citation. 40-60 words is the citation sweet spot per that doc.
• Per-cluster FAQ authoring is a content task per cluster — sequence by traffic (vpn hub, best-free-vpn, streaming, gaming first).

---

4. Machine-page version (faq + citations + takeaway + metrics)

The meeting wants a machine-readable page variant carrying: faq, citations, takeaway, metrics. All four inputs already exist in some form:

• faq → the parsed faqItems (intent-tagged).
• citations → rehype-citations.mjs output (the citation JSON-LD + Sources).
• takeaway → the > **Bottom Line:** blockquote.
• metrics → provider JSON fields / Trust Score / Speed Lab (⚠ currently static-"fake" per geo-ai-citations.md — flag which metrics are real before publishing them as cited facts; do not emit fabricated metrics into a machine page).

Recommended shape: a per-page .json (or .md) machine variant, e.g. GET /{permalink}/index.json, emitting { url, title, takeaway, faq: [{q, a, intent}], citations: [{n, source, url}], metrics: {...} }. Reuse the src/pages/*.json.ts route pattern (languages.json.ts is the precedent). Keep it English-first (matches llms-full.txt being EN-only).

Decision to confirm: machine variant per-page vs. one big feed. Given llms-full.txt already aggregates, the per-page JSON + a link from llms.txt is the lower-lift, more-citeable route.

---

5. Master FAQ page + master index page + dedup

• Master FAQ page (/faq/ or /vpn-faq/): aggregates every cluster's FAQs, grouped by cluster + intent. This is a strong GEO/AI-citation asset (one authoritative Q&A hub). Schema caution: a page with 100+ FAQPage items can trip Google's rich-result limits and dilute per-page schema — decide whether the master page carries full FAQPage schema or links to the per-cluster canonical FAQ. Recommend: master page = navigation + on-page answers, but each Q's canonical schema stays on its cluster page (avoids the duplicate-schema authority split).
• Master index page: a human sitemap of clusters → pages → their FAQ/machine variants. Pairs with llms.txt (the machine index) as the human-facing counterpart.
• Dedup gate (build script): a scripts/faq-dedup-check.mjs that collects every FAQ question site-wide, normalizes (lowercase, strip punctuation), and fails the build on near-duplicate questions across pages (fuzzy match, e.g. Levenshtein/token overlap). This is what enforces "no duplicate FAQs" — otherwise it drifts back immediately. Wire into deploy.sh gates alongside the existing i18n/link checks.

---

6. Authority / citation-token evaluation

Meeting item: "evaluate authority, token use for citations." Concretely:

• Audit rehype-citations.mjs output — how many citations per page, are the harvested sources actually authoritative (gov/edu/major-press vs. random blogs), and what's the token cost of the appended Sources block + [n] markup on every page (it ships in HTML on ~all posts).
• Deliverable: a short report — citation count distribution, source-authority tiering, and a token/byte cost per page — to decide whether citations stay auto-harvested, get curated, or get capped. geo-ai-citations.md already argues cited benchmark data is worth it; this validates the implementation is pulling good sources, not padding.

---

7. Phased plan

Phase 0 — foundation (unblocks everything, low content lift):

1. Fix extractFaqItems() answer flattening + link handling.
2. Extend faqItems → Provider + Pillar layouts (instant schema win on 22 provider pages).
3. Route all FAQs through FAQ.astro accordion (kill flat prose).
4. Add the [intent] parser tag (Option A) — backward-compatible.
5. Build scripts/faq-dedup-check.mjs, wire as a build gate.

Phase 1 — content, per cluster by traffic: 6. Mine GSC/PAA questions per cluster; author 12-20 across the four buckets; migrate hub FAQs off translations.ts into the unified path.

Phase 2 — machine + master: 7. Per-page machine JSON variant (faq/citations/takeaway/metrics), flag fake metrics. 8. Master FAQ page + master index page (schema-canonical decision per §5).

Phase 3 — evaluation: 9. Authority / citation-token report (§6); adjust citation harvesting.

---

Open decisions to confirm with Ben

1. Authoring convention: Option A (intent-tagged ###, recommended) vs. Option B (frontmatter array). Gates all downstream work.
2. Machine page format: per-page .json route (recommended) vs. aggregated feed.
3. Master FAQ schema: answers-on-page-but-schema-stays-canonical (recommended, avoids authority split) vs. full FAQPage on the master page.
4. Metrics in machine page: which metrics are real (Speed Lab / Trust Score are currently static) — do NOT publish fabricated metrics as cited facts.
5. i18n scope: machine page + master pages English-first, matching llms-full.txt?

Bottom line

The blocker isn't missing FAQs — it's two fragmented systems and schema-less provider/pillar pages. Phase 0 (unify render + schema + fix the parser + dedup gate) is pure infra and unlocks the meeting's asks cheaply. Then the real work is content: mine real human questions per cluster (GSC/PAA, not invented), 12-20 across four intents, deduped, LLM-front-loaded. The machine page, master pages, and citation eval all reuse infra that already exists (rehype-citations, llms.txt, Bottom-Line, provider JSON) rather than net-new systems.

---

GEO — Getting VPN.com Cited in AI Answers

Consolidated from geo-ai-citations.md · status: active

GEO — Getting Cited in AI Answers

Researched 2026-06-06 (web-validated). The question: is it worth optimizing for AI citations (ChatGPT, Perplexity, Google AI Overviews, Gemini), and what does it actually take? Companion to [authority-roadmap] (the moat/Speed Lab plan) and the comparison + authority moves in [next-50-days].

Verdict

Yes — but it's an authority + original-data play, not a schema/llms.txt play. And it's not a new initiative: the comparison supercluster and the Speed Lab/methodology work are the GEO engine. The single decision that unlocks it: make the Speed Lab data real, or stop leaning on it.

Why it's worth it (validated)

• AI referral traffic converts 4–10x better than organic: ChatGPT ~14–16%, Perplexity ~10.5%, Claude ~16.8%, vs Google organic ~1.76%.
• Growing fast: AI traffic ~165x faster than organic; +393% to US retailers in Q1 2026. Still ~1% of total traffic.
• Our AI traffic is 0.2% today — tiny in volume, but high-value per visit and compounding. This is a plant-early bet on the trajectory.

The honest ceiling (what GEO-agency blogs skip)

• For commercial "best VPN" queries, AI pulls mostly from independent sources, not affiliate sites: Reddit (~40% of citations), Wikipedia (~26%), YouTube/TechRadar. We will not displace Reddit/Wikipedia for generic "best VPN."
• Citations are gated more by off-site earned authority (mentions across independent domains, being in others' "best" lists, DA60+ backlinks, multi-source corroboration) than by on-page schema. Schema/llms.txt alone barely moves it.
• Where we can win: provider-specific, comparison ("X vs Y"), how-to, and branded/credibility queries ("is NordVPN legit", "NordVPN vs ExpressVPN for streaming") — where our pages get pulled into the source stack.

What actually moves the needle → maps to the plan

1. On-page (cheap, real but bounded ~40% lift — Princeton-proven): front-load the answer in the first ~200 words; add statistics, cite sources, add quotations; claim-rich intros. 44% of LLM citations come from the first 30% of the page. Bake into all content.
2. Original/benchmark data = #1 on-page driver: pages with benchmark data cited 2.8x more. → the comparison supercluster (data tables, winners) and Speed Lab (dated benchmarks). These two moves are the GEO play.
3. Off-page earned authority (the real ceiling, slow): independent mentions, press from dated Speed Lab reports, verification-badge backlinks, a methodology page good enough to be a citable source. The flywheel: publish real data → independent sources cite it → multi-source corroboration → AI confidence. You climb the affiliate ceiling by getting others to cite your data, not by publishing more of your own.

The linchpin / the fork

The highest-leverage asset is original data others can cite — exactly the Speed Lab. But ours is static/fake. Publishing fake benchmarks that AI cites is a reputational + legal landmine — one Reddit thread or journalist catch torches the authority you're building.

Decision: make the top-5 Speed Lab genuinely real + dated (citable, press-able, backlink-bait via the verification badge) — or stop relying on it. No safe "fake-but-cited" middle. Everything in the authority half of the plan hinges on this call.

Measurement

AI citations don't show cleanly in GA4. To know if this works, track answer-share / citation frequency for our target queries across ChatGPT/Perplexity/Google AIO (an AI-visibility monitor), plus AI referral sessions + conversion in itbroke. Baseline before, measure after.

Recommended sequence

1. On-page tactics into the comparison + provider content (cheap, do alongside Move 2). Front-load, stats, cite, benchmark tables.
2. Make the call on real Speed Lab data (the fork). If yes → top-5 real + dated + methodology = the citable core.
3. Earned authority (slow, ongoing): press the dated reports, pursue verification-badge backlinks, build genuine (not spammy) community presence.

Sources

• Conversion/growth: authoritytech (AI converts 4–23x), tryanalyze (AI ~1% of traffic), position.digital (AI SEO stats).
• GEO drivers + Princeton study: aithinkerlab (Princeton-backed), Enrich Labs, Frase playbook.
• Citation sources / affiliate ceiling: Search Engine Land (Reddit/YouTube/LinkedIn study), SolCrys (17,551-citation study: Wikipedia/TechRadar/Reddit), Dataslayer (AI prefers user content over brand sites).
• Benchmark data 2.8x: ConvertMate GEO Benchmark 2026.

---

Authority Roadmap — Making "Reviewed by VPN.com" Mean Something

Consolidated from authority-roadmap.md · status: active

Authority Roadmap

Goal: VPN.com becomes the cited source for VPN data, not a site that references other sources.

Success: AI assistants cite VPN.com Speed Lab data. Providers display "VPN.com Verified" badges. Journalists reference the Trust Score. The methodology page is linked from Wikipedia.

---

Current State (April 2026)

Have: 120-page static site, Speed Lab (10 providers), 100-point rating system, JSON-LD, llms.txt, /methodology/ page.

Missing:

• Methodology page is thin marketing copy — no verifiable test details
• No dated test reports — just static numbers
• No verification program or badge
• No public API
• No original data unavailable elsewhere

---

Phase 1: Credibility Foundation

1.1 Methodology Page Overhaul

Priority: Critical — everything else rests on this.

Rewrite /methodology/ to read like a research paper. Must cover:

• Test environment: hardware specs, OS, network baseline (ISP, bandwidth without VPN)
• Test protocol: servers tested, run count, peak vs off-peak, duration
• Speed testing: tools (Ookla, iPerf3), how throughput/latency/loss are measured
• Scoring formula: exact point allocation, what earns 10/10 vs 5/10 per criterion
• Disqualification criteria: what causes a provider to lose points or be removed
• Update frequency: how often tests re-run
• Conflict of interest: affiliate relationship disclosed; how it doesn't affect scoring
• Limitations: what the tests don't cover, known biases

Tone: technical, transparent, boring. Written for peer review, not for customers. Reference: AV-TEST methodology

1.2 Brand the Scoring System

Name it. Always display as "94/100 VPN.com Trust Score." Reference by name in every provider review.

Score ranges on the methodology page:

1.3 Affiliate Disclosure

Strengthen /disclosures/ to explicitly state: VPN.com earns commissions; affiliate relationships never influence scores; methodology is public and applied uniformly; providers cannot pay for higher scores; editorial team doesn't know commission rates during testing.

---

Phase 2: Original Data

2.1 Monthly Speed Lab Reports

Monthly blog posts with dated test results. Format: date range, comparison table, changes from prior month, raw data (throughput, latency, loss, stability per location).

Why it matters: creates freshness signals, each report is a citable dated data point, providers link to good results (free backlinks), builds a historical archive.

Technical: Data already exists in provider JSONs. Add lastTested date field and build a report template.

2.2 Provider Verification Program

"VPN.com Verified" badge — earned, not bought.

• 12/12: VPN.com Verified
• 10-11/12: Verified with noted exceptions
• Below 10: Not verified

Providers who pass can use the badge on their own sites — every badge is a backlink. Full transparency page at /vpn/verified/.

2.3 Static Data API

Endpoint: /api/speedlab.json — generated at build time.

{
  "lastUpdated": "2026-04-29",
  "testEnvironment": "1 Gbps fiber, US East Coast",
  "providers": [
    {
      "name": "NordVPN",
      "rank": 1,
      "trustScore": 94,
      "speed": { "download": "730 Mbps", "latency": "18 ms" },
      "verified": true,
      "reviewUrl": "https://www.vpn.com/vpn/nordvpn/"
    }
  ]
}

Open data gets cited. Researchers, journalists, and AI models consume structured JSON.

---

Phase 3: AI Citation Strategy

Already done: llms.txt + llms-full.txt, JSON-LD, clean semantic HTML.

Still needed: Every factual claim should be self-contained and dated. Instead of "NordVPN is fast," write "NordVPN achieved 730 Mbps in VPN.com's April 2026 Speed Lab test, ranking #1 of 10 providers tested."

Content only VPN.com can produce:

• Head-to-head results on the same hardware, same day (most review sites test separately)
• Longitudinal data ("NordVPN's speed improved 12% over 6 months")
• The Trust Score itself — proprietary and citable
• Verification status — only VPN.com can confirm "VPN.com Verified"

Structure answers to match AI queries: claim + evidence + source attribution + date.

---

Phase 4: Distribution

• Wikipedia: The VPN comparison article links to methodology sources. A rigorous methodology page qualifies as a reliable source — highest-value backlink in the VPN space.
• Provider sites: Verified badge → contextual backlinks from high-authority domains.
• AI training data: Structured, dated, factual content gets ingested. llms.txt signals what matters.
• Press: Monthly Speed Lab reports are press-releasable. This is how AV-TEST built their brand.

Don't: buy links, publish thin "best VPN for [city]" farms, inflate scores for affiliates, or claim "industry-leading" without proof.

---

Implementation Priority

Start with 1–3. Pure content, no code, and the foundation for everything else.

---

The Principle

Authority is earned by doing work nobody else will do, publishing it transparently, and being right consistently over time.

The Speed Lab is the moat. The methodology is the credibility. The monthly reports are the proof. The verification program is the brand. The open data is the distribution.

---

Comparison Supercluster — Provider vs Provider Matrix

Consolidated from comparison-supercluster.md · status: parked

Problem

VPN.com has 22 individual provider reviews and 13 "best-for" comparison pages. But no provider vs provider comparison content.

When a user asks "NordVPN vs ExpressVPN", there's no page to serve. The user has to read two separate reviews and do the comparison themselves. AI models can't cite a comparison because one doesn't exist.

This is the single biggest content gap identified by ChatGPT's site audit:

"You need provider vs provider matrix, protocol comparisons, use-case comparisons, pricing comparisons. Comparison intent is extremely commercial, extremely AI-native, extremely agent-friendly."

Opportunity

Comparison queries are:

• High intent — the user is choosing between two options, ready to buy
• AI-native — AI systems love structured comparisons (tables, scores, winners)
• High volume — "NordVPN vs ExpressVPN" has significant search volume
• Currently unserved — we send users to two separate review pages instead of one definitive comparison

Plan

Phase 1: Top 10 provider-vs-provider pages

Generate the highest-demand comparisons from our top providers:

Phase 2: Comparison hub page

/vpn/compare/ — a hub listing all available comparisons with a provider picker (select two providers → see comparison).

Phase 3: Data-driven comparison component

A reusable Astro component that takes two provider slugs and renders a structured comparison:

• Side-by-side data table (pulled from provider JSON)
• Winner per category
• Overall winner
• Prose analysis per category

This component makes creating new comparisons trivial — just set the two slugs and write the analysis.

Architecture

src/content/posts/nordvpn-vs-expressvpn.md
  → category: vpn
  → template: comparison (new)
  → frontmatter: compareProviders: ["nordvpn", "expressvpn"]
  → component auto-generates data table
  → prose analysis written per page

Content approach

Each comparison page should have:

1. Quick verdict — who wins and why (above the fold)
2. Data table — auto-generated from provider JSON (speed, servers, price, rating)
3. Category-by-category analysis — 5-6 sections (speed, privacy, streaming, price, ease of use, verdict)
4. Winner per category — explicit, not vague
5. FAQ — "Is NordVPN better than ExpressVPN for streaming?" etc.

What NOT to do

• Don't template these with generic copy. Each comparison needs unique analysis.
• Don't compare providers we don't have strong data on.
• Don't create a comparison for every possible pair (22 × 21 = 462 combinations). Start with the top 10.

Dependencies

• Provider JSON data (already available)
• Comparison component (needs building)
• Comparison hub page (needs building)

Status

• Design comparison component
• Build 10 comparison pages
• Build /vpn/compare/ hub
• Add to llms.txt as canonical comparison section
• Add to sitemap

---