Was gibt es Neues bei Contextter
Jedes Feature, jeder Fix und jede Verbesserung — dokumentiert wie wir liefern.
Directus Production Readiness
Directus service notices now render through the public website shell, and the scheduler uses a dedicated least-privilege policy with update tracking for edited published content. Markdown draft import is now Directus-native as well. Local Blog and Glossary Markdown files can be dry-run or uploaded idempotently with a Publishing token, and the first German legacy glossary seed drafts are available for review as Directus drafts. Directus now also has an editorial checks dashboard for missing translations and publish blockers, ops events for health/revalidation/auth/rollback monitoring, a Hetzner monitor worker, a rollback restore drill, and a German Studio language setup script. Public service-state CTA links are scheme-validated before rendering. The Directus seed library now includes five German Blog drafts alongside the glossary drafts, so editors can review realistic Blog and Glossary content directly in the CMS before publishing.
Directus Release Channels
Directus Blog and Glossary content now supports production and development release channels, plus a CMS-managed service-state foundation for maintenance and coming-soon notices.
AI Runtime Integrity Hardening
The AI runtime and admin observability stack now use stronger shared contracts, exact run lifecycle telemetry, safer cache identity, and richer admin traces so Writer, Structure, Optimizer, and the AI tools dashboard stay much closer to the same runtime truth. This update also fixes known blind spots around manual style instructions, partial research sessions, voice resolution, stale pending runs, and drift detection, while adding deeper regression coverage for cache isolation, run correlation, and admin status semantics.
Writer Resilience
The Writer now uses a shared revision-gated document commit contract, which means manual saves, restores, and AI-driven apply flows can no longer silently overwrite newer content. We also added explicit Writer auth and access-loss states, browser-side recovery foundations for offline or crash scenarios, and stale-response guards so long-running AI or link-apply results do not clobber newer editor changes.
Writer Browser Editor Overhaul
The Writer browser editor was restructured around shared editor configuration, extracted surface/sidebar boundaries, and dedicated Writer state hooks so the tab now acts as an orchestrator instead of a large mixed UI/runtime component. This overhaul also adds focused regression coverage for slash-command configuration, live extension rehydration, extracted Writer layout branches, and editor recreation flows so future Writer changes are less likely to reintroduce drift or hidden browser-editor bugs.
Brain Entries Panel Grouping
The Brain knowledge page now keeps search, filters, result counts, and the entry list inside one shared entries panel. This makes it much clearer where search results appear while still keeping the page calmer by allowing the list itself to stay collapsed until people want to inspect entries. The panel also opens automatically when someone starts searching or filtering, and empty or no-results states now render inside the same result area instead of feeling disconnected from the controls above.
Brain Entries Panel Lazy Load
The Brain knowledge page now treats the entries panel as a real lazy-loaded surface instead of only a visual collapse. When the panel is closed and no search or filter is active, the paginated entry query is skipped entirely. Opening the panel starts loading entries on demand, and the infinite-scroll observer only runs while the panel is expanded. Active search and filter interactions still open the same panel automatically so results remain easy to understand.
Project Context Tab redesigned around readiness, visible context, and full localization
The Project Context Tab now starts with a readiness summary and a visible context inventory instead of a one-question-at-a-time wizard. Existing Brain and Research signals stay visible by category, multiple readiness questions can be answered inline, completion keeps the inventory in view, and the full flow now respects the active locale across UI copy, backend fallbacks, and gateway-generated readiness questions.
Payload Multilingual Grouping and Editorial Views
Payload blog posts and glossary entries now behave as one multilingual editorial document instead of feeling like loosely related locale copies. Editors can set a shared source language, track translation state per locale, and use new dashboard/list overviews for drafts, review, scheduled, and live content. We also aligned the public website with that model by deriving hreflang links and sitemap alternates from the shared Payload document identity plus localized slugs, so blog and glossary locales no longer depend on the removed `translationGroupId` assumption. At the same time, the public marketing landing pages were detached from the retired CMS adapter layer. Those routes are code-owned again, while the website keeps Payload as the source of truth for blog and glossary content plus Payload-native preview/revalidation.
Research now distinguishes sparse-but-usable sessions from truly blocked runs
Research sessions now handle thin DataForSEO and SERP edge cases more honestly. Empty upstream responses are no longer treated like provider outages by default, sparse sessions with usable secondary signals can render in a limited state, and the diagnostics layer now separates real provider failures from cases where the provider simply returned too little data to support a full report.
Writer Hardening
The Writer now uses one shared AI edit path for slash commands and the bubble menu, which means editor rewrites keep the selected text intact until the AI request succeeds and always send the same project, context, and scoring payload to `/api/generate`. We also consolidated Writer and Structure setup behind a shared generative context helper, introduced an explicit embedding gateway facade for Writer-adjacent retrieval, aligned Writer route tests with the real action contract, and finished the remaining i18n cleanup in the issues widget.
Payload CMS Cutover Hardening
Payload is now the explicit editorial source of truth for blog posts, glossary entries, authors, media, taxonomies, and site settings. The public website no longer reads landing-page section or SEO data from the old Convex page CMS adapter, so landing pages remain code-owned while editorial content flows through Payload. We also removed deprecated editorial CMS links from the app admin navigation, redirected old typed CMS editor URLs to Payload or analytics surfaces, and documented the remaining legacy CMS surfaces so follow-up removals can happen safely in smaller slices.
Billing Enterprise Hardening
Billing now consistently follows the active organization as its security and data boundary, including checkout, top-ups, invoices, customer-portal access, transaction history, and usage analytics. We also hardened Polar refund handling for partial refunds and repeated webhook deliveries, removed raw model exposure from billing views, and made billing history, activity, and invoice actions more reliable for production use.
Organization Membership Enterprise Hardening
Organization membership is now hardened for enterprise rollout: invites expire after 7 days, only the invited email can accept them, Clerk no longer overrides Contextter's granular org roles, and clients stay assignment-only even in open workspaces. We also tightened member/invite visibility and extended the same tenant checks into research-session background-job wrappers so org-scoped access cannot drift around the main permission system.
Review System Hardening
Reviews now run through a shared cycle-based workflow for both content and research, with stronger reviewer validation, safer multi-reviewer concurrency, durable notification delivery, hardened external review sessions, and centralized lifecycle cleanup for archive, delete, and member-removal flows.
Enterprise Mobile Hardening for the Public Website
The public Contextter website now behaves like a true mobile-first marketing surface on real device widths instead of relying on desktop-scale fallback behavior. The shared shell now ships an explicit device viewport contract, a hardened floating mobile navigation overlay, safer safe-area handling, stronger touch targets, cleaner footer/CTA wrapping, and route-family responsive rules that keep homepage, feature, solution, compare, glossary, changelog, tool, and legal/support pages readable and stable on phones. This pass also removes unintended horizontal overflow across representative public routes, improves mobile handling for hero layouts and comparison/table-like sections, and adds Playwright regression coverage for the mobile navigation flow plus narrow-viewport smoke checks across the major page families.
Content Creation Flow Realignment
Content creation now follows one consistent project briefing from setup through writing and review. Projects can carry a real `Content Brief`, explicitly choose a `Primary Research Session`, and show a truthful `Project Briefing` overview instead of random placeholder SEO metrics. The Writer, Score, Precision Editor, Optimizer, and link suggestions now react to the same briefing contract. Live score and precision checks use the current editor snapshot, meta and internal-link findings are evidence-backed, and Structure/Writer now respect source toggles so disabled Research or Brain context no longer slips into the downstream pipeline.
Privacy-Safe Marketing Analytics Expansion
The public website and help center now send richer privacy-safe Umami analytics without any per-page manual setup. New marketing and docs routes keep automatic page-level tracking by URL path, Web Vitals are enabled for both surfaces, and the shared analytics layer now adds low-cardinality engagement signals like scroll depth, engaged visits, `page_context_view`, and normalized CTA metadata so landing pages, blog posts, glossary terms, changelog entries, docs routes, and free-tool pages are easier to compare in Umami. This rollout also adds bounded `section_view` telemetry for major landing/content blocks, richer glossary and blog interaction signals, pricing-card impressions, cookie-consent and breadcrumb tracking, plus non-PII tool engagement markers such as first input focus and submit-readiness buckets.
Research Production Hardening
Research sessions now enforce assignment visibility more consistently in restricted and private workspaces, including project-linked lookups, retry actions, and session-scoped artifact reads. We also hardened the background-job wrappers so tenant-scoped research reads and writes must match the owning organization and workspace instead of trusting only the shared job secret.
Content Brief Tracer Bullet
Projects can now carry an explicit `Content Brief` through setup, settings, Structure, and Writer. The first tracer-bullet release adds page-goal, audience, must-answer question, conversion goal, and notes fields so teams can steer output toward sales pages, answer pages, and templates instead of falling back to generic blog-style drafts.
Glossary Localization and Editorial Layout
Glossary term pages now use a wider editorial layout with cover images, a definition sidebar, and improved related-term navigation. The first localized glossary proof point ships with `alt-text` in English, German, French, and Spanish, and glossary metadata now builds canonical and hreflang links from real published translation records.
Sitemap and Canonical Host Hardening
The marketing website sitemap now ships as styled XML with preserved `hreflang` alternates, image entries, and the full public landing/blog/glossary surface. We also tightened the web middleware so `www.contextter.com` requests that reach the app redirect back to the canonical apex domain, keeping sitemap and metadata output aligned with `https://contextter.com`.
Sitewide Marketing SEO Hardening
Localized marketing pages now build cleaner metadata with consistent canonicals, alternate-language links, and non-duplicated brand titles across the homepage, pricing, contact, docs, feature, solution, and compare routes. The docs route is healthy again for sitemap inclusion, German/French/Spanish marketing pages now ship localized visible copy instead of English fallbacks, and localized breadcrumb trails now stay translated in both the UI and breadcrumb JSON-LD.
MVP backend workload hardening
Background jobs now emit structured workload telemetry and Inngest registration can be split into light and heavy function groups. The default behavior is unchanged, while heavy crawl, research, and keyword jobs are ready to move to the existing Hetzner server when production metrics justify it. Crawl starts are also safer locally: duplicate active crawls are rejected atomically before dispatch, failed dispatches no longer leave blocking queued jobs behind, Crawl V2/V3 share the same worker concurrency cap, and crawl history reads use bounded limits plus an indexed daily count path.
On-Page SEO audit foundation
Contextter now ships the first On-Page SEO workspace surface with source health, URL-level audit detail, Google Search Console connection controls, and resilient fallback handling across crawler, Google, CrUX, Lighthouse, and DataForSEO inputs.
Research data plane hardening
Research Intelligence v5 now handles sparse and provider-impaired runs more honestly. Related-search and keyword-difficulty fallbacks are more resilient, admin diagnostics explain provider cascades more clearly, and the unified Gaps card now fills the available width when fewer sections are visible. We also hardened the failure path for fatal DataForSEO incidents: when critical keyword or SERP payloads are unavailable because of billing, quota, or provider failures, the research run now stops early and shows a direct session error instead of persisting an empty customer report.
Marketing analytics hardening
The public website and help center now use stricter Umami host scoping, key CTA events, and updated privacy text to reflect the self-hosted no-cookie analytics setup on marketing surfaces.
Simplified Brain workspace
The Brain workspace now opens directly into the knowledge view and uses a smaller navigation model: Knowledge, Sources, and Analytics. Uploads, crawled URLs, and document processing have moved into Sources, while the knowledge view now shows compact Brain health cards backed by lightweight server summaries instead of multiple dashboard panels.
Billing Control Plane
Admin Billing Control Center with feature-specific pricing, fail-closed external API billing for Research providers, and a profit ledger for provider cost, charged amount, platform absorption, tax estimate, and operating-cost estimate.
Dashboard density refresh
The workspace dashboard now uses a tighter overview layout, removes the quick action strip, and surfaces compact project, keyword, research, and knowledge signals.
Contextter Help Center refresh
The public documentation has been reshaped into a customer-facing Contextter v5 Help Center focused on onboarding, workflows, product usage, team process, billing, and troubleshooting.
Feature Control Plane
Feature access now uses a canonical Convex-backed control plane with server-side evaluation for plan, organization, workspace, project mode, dependencies, and global kill switches.
Logging hardening
Contextter now uses stricter structured logging in server, AI, and Convex runtimes. Logs redact more secret and PII-shaped fields, strip URL query strings, carry organization context, and avoid raw runtime `console.*` logging outside explicit migration/seed surfaces.
PostHog feature-control semantics aligned
PostHog feature flag diagnostics now respect Convex as the production Feature Control Plane. Convex access decisions are authoritative, while PostHog remains available for analytics, experiments, and rollout diagnostics.
Research Data Plane
Research Intelligence v5 now has a dedicated data plane for DataForSEO, Autocomplete, Keyword Database handoff, and module diagnostics. Admins can see which provider call ran, what normalized successfully, which capabilities each module required, and whether a missing report section is caused by billing, provider downtime, empty SERP data, rejected normalization, or a true module input gap.
CSV / XLSX Bulk-Upload für Keywords
Bis zu 10.000 Keywords pro CSV- oder XLSX-Datei in einem Schritt importieren — mit Spalten-Mapping, Vorschau, Duplikat-Erkennung und Live-Guthabenprüfung vor dem Commit. Optionale Spalten für Tags und Listen erstellen fehlende Einträge automatisch.
SERP Features On-Demand
Anreichern Sie SERP-Features (AI Overview, Featured Snippet, People Also Ask, Local Pack, etc.) für jedes Keyword in Ihrer Datenbank — auf Knopfdruck. Pro Keyword nur ~€0,003 mit klarer Kostenkontrolle vor jedem Bulk-Job.
SERP Analysis rebuild
The SEO tab now opens with a full-width "SERP-Analyse" view: a slot timeline on the left, a faithful Google rebuild in the centre, and six intelligence cards (Ranking-Chance, Scorecard, Klick-Verteilung, AI Overview Dekonstruktion, SERP-Struktur, Content-Brief) on the right. Replaces the standalone S8/S13/S15 cards with a single coherent dashboard. Flag-gated per workspace; legacy layout remains the default until rollout.
Keyword detail sidebar redesigned around three purposeful tabs
The keyword DetailSidebar now condenses all DataForSEO metrics into three focused tabs — Overview (KPIs, signals, ad potential, competitive context, taxonomy), Historie (12-month volume chart, traffic-by-position curve, trend deltas), and Verwandte (session-cohort siblings + on-demand Google Suggest). SERP and Content-Ideen placeholders are gone; tab choice persists across keyword switches.
Research Discovery → Keyword Database
Research sessions no longer render a separate Discovery tab. Every completed session now auto-creates a keyword list in the Keyword Database named `"${keyword} — YYYY-MM-DD"` and bulk-loads every research-discovered keyword (primary + autocomplete expansion) as a member, so there is exactly one home for keyword work across the app. The auto-list is idempotent on `(workspaceId, sessionId)` — Inngest retries never produce duplicates — and fails open: if list creation hits an error the research session still completes and a warning is logged. Four components (`KeywordDiscoveryPanel`, `filter-sidebar`, `keyword-discovery/index`, `GapDiscovery`) and the legacy `keywordEnrichments.backfillFromSessionReport` mutation are removed.
Mock Research — Full 27-Module Coverage
The admin "Start with Mock Data" flow now produces a fully populated research session end-to-end. A single click generates all 27 v5 intelligence modules (11 writing + 16 SEO), 16 canonical adapter payloads, 60+ keywords, and session/workspace clusters — all deterministic from the seed keyword and 100% zero-cost (no OpenRouter, no DataForSEO, no balance deduction). The `MockResearchModal` gained advanced toggles (`hasAiOverview`, `hasFeaturedSnippet`, `ymylCategory`, `competitorStrength`, `completenessScore`) so admins can force edge-case UI states for QA and screenshots. Sessions are flagged `isMock: true` and the full intelligence report is persisted via a new admin-gated `updateMockReport` mutation.
Smart Tags v1
Smart Tags turn the Keyword Workbench's tag layer into a cross-cutting attribute system: two German system auto-tags (Frage, Kaufabsicht) are seeded per workspace and auto-applied on ingestion, FilterPopover gains a multi-select tag chip filter with live member counts, and BulkActionBar / InlineActionRow / DetailSidebar support inline tag create and edit (label + colour swatch) in one click — no separate management page.
Cluster View Redesign
The Keyword Database cluster tab is now a full two-surface workbench: Cluster-Table (Ahrefs-style flat rows with inline sortable sub-keyword tables) and Cluster-Workbench (Semrush-style master-detail with preset chips — All, Easy Wins, Quick Conversions, High Potential — a per-cluster KPI strip, and an action bar for rename/merge/split/export/handoff). A dismissible "Uncertain" banner replaces the old bucket card and surfaces mixed-availability clusters with a one-click enrich CTA; the legacy `ClusterTreeView` / `ClusterTreeNode` / `UncertainBucketCard` components have been deleted. Empty-state gaps are closed: `createMockSession` now seeds `workspaceKeywordClusters` so fresh workspaces immediately show clusters, and a scope-aware `ReclusterButton` dispatches against the correct endpoint for workspace vs session scope with 429 toast surfacing. All 4 locales (en/de/es/fr) translated in a single pass.
Keyword Database v2 Workbench
The Keyword Database tab gains a full multi-view workbench: switch between Flat, Cluster, Matrix, and Raw views to triage keywords at any stage of enrichment. You can now tag keywords inline, save filter + column presets as reusable views, bulk-tag / delete / export via shift-click selection, and manually add up to 100 keywords per batch from a single dialog. A new Raw view surfaces keywords that have not been enriched yet so nothing gets lost between research sessions, and every adapter failure is now captured live in the admin debug panel with a one-click "mark resolved" action.
Workspace Cluster Tree & Hand-Off
The Keyword Database now grows a full cluster tree view: parent clusters and their sub-clusters render as nested cards with metrics, member chips, and expand/collapse controls across depth 0/1/2. A dedicated amber-bordered "Uncertain" bucket surfaces mixed-availability keywords with an inline DataForSEO cost preview and one-click enrichment dispatch. Root cards gain a hand-off dropdown to push a cluster's keywords to the Writer (pending brief row), Brain (inline `brain.add` fan-out), or a CSV download — wired to a workspace-scoped handoff route that anchors on the workspace's most recent research session. Shipped behind the `research_workspace_clusters_v1` feature flag (OFF by default, admin bypass) with a new `workspaceClusterRuns` telemetry tile in the admin debug panel tracking success rate, duration, cluster counts, and embedding-cache hit rate.
Keyword Database & Free Discovery
Every keyword your research sessions harvest now lives in a workspace-level Keyword Database, independent of individual session lifecycles — delete a session and the keywords stay. The per-session Discovery tab also drops its cost floor: it now always lists every deduped autocomplete keyword for free, with volume/KD/CPC columns degrading gracefully to `—` until an admin explicitly runs enrichment. Two-tier deletion, 30-day tombstones, pipeline funnel telemetry in the admin debug panel, and a brand-exclusion list complete the ship.
Keyword Discovery Quality & Clustering
The autocomplete keyword discovery pipeline now ships with five new quality stages: workspace brand-exclusion lists, per-result language detection with length and token sanity gates, semantic dedup that merges near-duplicate phrases via embeddings, a DataForSEO bulk enrichment worker that fans 700-keyword chunks into one batched call per session, and agglomerative topic clustering that surfaces the highest-opportunity groups using a volume × (1 − competition) × ((100 − KD) / 100) × intent-weight score. Each stage is admin-toggleable and surfaces live stats in the research debug panel.
Async-enrichment tier for research
Slow research sources (fact-check, academic, autocomplete v2) now run in the background after the initial report is ready. Sections stream in as each source finishes, with per-section retry if anything fails and a 15-minute watchdog that ensures stuck sessions never hang.
Autocomplete keyword expansion
Research now fans each seed keyword across up to 100 Google Suggest variants — alphabet, questions, modifiers, plus optional cities, competitor brands, and content themes configured per workspace. Seven-day cache, HTTP-first transport with headless-browser fallback, and zero extra credit cost.
Keyword Opportunity View
Research Intelligence now fetches a full metrics payload for every keyword candidate (volume, KD, CPC, competition bid range, SERP saturation, ad economics, 12-month trend) in one batched call per session. A new shared Opportunity View replaces the vertical stacked lists in Subtopics, Related Keywords, Keyword Clusters, and Subtopic SEO — scatter chart, compact list, and hover card with sparkline and full metrics surface strategic quadrants (Easy Wins / Contested / Hidden Gems / Skip) at a glance.
Smarter SERP Intent Classification
The Research Intelligence intent classifier is now noticeably more accurate on service-provider SERPs and much more stable run-to-run. Local-service results (cleaning, repair, agencies, trades) now get a dedicated `service` pageType instead of being miscategorised as FAQ or product pages, and dedicated pricing pages (`/preise/`, `/tarifs/`, `/pricing/`) are recognised as their own bucket. The classifier also became smarter about admitting uncertainty: when the top two intents are within a narrow margin and SERP coverage is limited, it now reports `uncertain` with an explicit explanation that the classification may flip between runs, instead of silently toggling. Under the hood we raised the confidence bar for the seven-language lexicon detector, added a ccTLD fallback so German product listings no longer get mistagged as English, tightened the commercial-compare phrase matcher to require ≥2 distinct triggers, and humanised two new evidence labels (`service provider`, `booking cta`) in the per-URL table.
Topical Fit v2: 4-dimension diagnosis
Research Intelligence's "Topical Fit" card now shows a 4-dimension diagnosis (topic, intent/format, depth, entity coverage) plus a short LLM-generated recommendation — instead of a single cosine score. Research sessions now tell you exactly *why* a keyword is a fit or a stretch against your brain and what's missing to close the gap.
Content-Based Intent Detection v2
Research v5 intent classification now reads the actual page content of the top 10 competitors — schema.org types, price/currency density, checkout forms, listicle and FAQ patterns, plus 7-language keyword lexicons — instead of relying on domain and URL patterns alone. Zero new crawls, zero LLM cost.
Research Intelligence v5
Research now runs a 27-module intelligence pipeline (11 writing + 16 SEO) that feeds a professional two-column dashboard with verdict, key metrics, SERP visualisation, and per-module cards. Multilingual YMYL detection uses embedding prototypes so it works across EN/DE/FR/ES without keyword lists. Opt-in per workspace behind the `research_intelligence_v5` feature flag.
Research Admin Debug Panel
Admins can now see a detailed pipeline debug panel at the top of every Research Session — showing which adapters succeeded or failed, phase timings, and extracted data counts.
Link Intelligence System
7-dimension link quality scoring (semantic relevance, anchor descriptiveness, density, position, hallucination detection) integrated into the editor link panel, optimizer prompt, and precision editor. External link suggestions now filter out competitor domains and SERP competitors automatically.
Default Voice created automatically on every crawl
The Brain now extracts writing style inline during page extraction — no separate job needed. Every completed crawl automatically produces a Default Voice profile that can be assigned to projects for style-consistent content generation.
Settings Inheritance & Default Author
Workspaces can now set a **Default Target Score** and **Default Author Profile** that new projects inherit automatically. The project creation form shows a streamlined 4-tile Smart Defaults grid with Language, Author Profile, Target Score, and Word Count. Style and tone are now controlled through Author Profiles instead of manual dropdowns, giving agencies consistent voice across all content.
Production Monitoring & Backup Upgrade
Added professional monitoring stack with Grafana dashboards (CPU, RAM, disk, per-container metrics), centralized log aggregation via Loki, and Uptime Kuma for real-time service availability alerts. Upgraded backup system to BorgBackup with encryption, deduplication, and automated retention policies — reducing storage usage by 85%.
Crawler v3 & Benchmark System
Completely overhauled the website crawling engine powering the Digital Brain. The new Crawler v3 brings smarter content extraction, automatic error recovery, and a professional benchmarking system for continuous quality improvement. **Smarter Crawling:** Four crawler backends (HTTP/Cheerio, Browserless, Spider Cloud, Jina AI) now work together in an intelligent cascade. If one crawler fails or gets rate-limited, the system automatically tries the next one instead of marking pages as failed. Content extraction uses Defuddle-based boilerplate removal to strip navigation, footers, and sidebars before analysis. **Quality Scoring:** Every crawled page receives a computational quality score based on six metrics: content density, semantic richness, readability, freshness, structure quality, and scanability. These scores are visible per page in the Brain knowledge base and aggregated in the Analytics dashboard. **Re-Crawl Intelligence:** When re-crawling pages, the system now compares new content against existing Brain entries using similarity matching. Similar entries get updated in place instead of creating duplicates, keeping your knowledge base clean and accurate. **Benchmark System:** Admins can benchmark all four crawlers against a configurable set of test URLs, compare quality metrics side-by-side, track performance over time with trend charts and regression detection, and export detailed reports for analysis. The system supports multiple domains with different CMS architectures to ensure crawlers work well across diverse websites. **Improved Progress Tracking:** The crawl progress view now shows a real-time pipeline visualization with live counters for discovered, crawled, extracted, and completed pages, along with estimated time remaining.
EUR Balance Billing
Transparent pay-per-use billing in EUR. Real provider costs with clear markup — no abstract credit system. Reserve-Execute-Settle pattern ensures you never overspend.
Organization & Team System
Multi-tenant workspace management with team roles, member invitations, and organization-level billing. Manage multiple brands from a single dashboard.
Self-Hosted Infrastructure
Migrated to dedicated Hetzner infrastructure for better performance, data sovereignty, and cost control. Full Docker Compose stack with automated backups.
Brain Evolution v3
HyDE search, agentic retrieval, authority scoring, contextual embeddings, entity relationships, and research-to-brain smart feed. 12 feature flags for gradual rollout.
Score-Driven Optimizer v4
Complete optimizer rewrite with SSE streaming, evidence-based section mapping, parallel optimization, and coherence pass. Every section goes through LLM — system never skips.
20-Metric Scoring System
Three-layer scoring with 9 LLM-powered deep audit metrics, 6 real-time computational signals, and 5 technical SEO fundamentals. Trust Gate caps scores for factual inaccuracy.
Research Tool v2
16 source adapters scanning in parallel — SERP, Scholar, Reddit, YouTube, Wikipedia, and 11 more. 14 enrichment modules for gap analysis, authority scoring, and opportunity detection.
Design System v2
Complete visual overhaul with carbon neutral scale, 8px grid system, warm monochrome palette, and 63+ reusable components. Dark mode first, accessibility built in.