How Woord works

Eight stages.
Two approval gates.
One publish-ready article.

Most AI writers are one-shot prompts: keyword in, slop out. Woord splits the job into specialists — researcher, strategist, writer, editor — and gives each one a contract from the previous stage.

The pipeline at a glance

Minutes per article, not hours
Specialist agents
Each stage has one job. No single LLM call holds research, structure, voice, and SEO at once.
Contract-driven
Each stage hands the next a contract. The contract is what stops downstream drift.
Two human gates
Approve the outline. Approve the draft. You steer the article without throwing it out.
STAGE 01

Research

Before Woord writes a single word, it builds a 360° picture of the keyword: what's ranking, who's ranking, what entities matter, what facts are verifiable, and what your own site is already doing for you.

SERP analysis

Search volume, KD, CPC, opportunity score, search intent, SERP features (featured snippet, PAA, knowledge graph) and seasonality — pulled live from DataForSEO.

Competitor analysis

Top 10 SERP results pulled with title, snippet, position, domain. Pages can be scraped and analyzed by Claude for entity coverage and structure.

Entity extraction

Claude — not spaCy or Google NLP — pulls core entities (must-mention) and supporting entities. Placed structurally by the outliner, not sprinkled randomly.

Primary sources

For comparison and YMYL keywords, Perplexity pulls verified product facts, pricing, regulatory and clinical sources from official pages — with URLs. This becomes the stat bank.

YMYL detection

The researcher scans the SERP for health, finance, legal, and news domains and classifies the topic. YMYL gets stricter sourcing downstream.

Your GSC data

Striking-distance keywords (positions 11–20), cannibalization risks, refresh candidates, and competitor wins surfaced from your real Search Console data.

What this means for you

  • No hallucinated stats — if a number can't be sourced, it doesn't appear.
  • Articles compete with what's actually ranking, not the LLM's hazy memory.
  • Your own data shapes what's worth writing.
  • Entity coverage is enforced structurally, not by keyword density.

DataForSEO + Perplexity + Claude. YMYL classified across health, finance, legal, news, product, general.

STAGE 02

Brief

Before drafting, Woord writes a contract. The contract locks word count, format, required elements, sourced facts, and exact section structure. Drafting then has only one job: execute the contract.

Content contract

Locked fields: keyword (year-updated for listicles), volume/KD/CPC, target word count by format, format recommendation, stat bank, topics to emphasize/avoid, custom instructions.

Format-aware word count

Answer: 400–1,200. Listicle: 800–2,500. Guide: 1,200–2,500. Comparison: 1,000–2,200. Pillar: 2,000–4,000. No one-size-fits-all guess.

Stat bank

Verified statistics with source URLs. The writer's system prompt enforces: "You may ONLY state as fact what appears in CONTENT_CONTRACT.facts_allowed."

Decision support

Commercial content gets first-class structures: disqualifier blocks, self-assessment checklists, comparison tables, decision summary boxes.

What this means for you

  • No drift — the writer can't add an irrelevant section, the contract doesn't allow it.
  • Format hygiene: a comparison actually has a comparison table; a guide actually has numbered steps.
  • Every statistic traces back to a URL. Receipts in the brief.

Anti-hallucination is enforced at the system-prompt level, not as a soft suggestion.

STAGE 03

Outline

The outline turns the brief into an exact section-by-section plan — required elements per section, extractable ledes for AI Overviews, and one designated differentiator section per article.

Format-required sections

Comparison: intro → criteria → evaluation → table → verdict. Listicle: intro → numbered items → conclusion. Guide: intro → numbered steps → conclusion. Answer: ≤60-word answer first, then context.

Per-section spec

Heading, section type, target word count, key points, required elements (table, list, checklist, examples, pros/cons, decision blocks), and SEO terms assigned per section.

Extractable lede

A single declarative sentence at the start of each section, designed to be pulled into AI Overviews, featured snippets, and ChatGPT/Perplexity citations.

Differentiation marker

Exactly one section per article is marked as "the section that makes this article worth reading" — your angle competitors don't have.

What this means for you

  • Approve once, draft locked. The writer can't drift away from the structure.
  • AI Overviews ready by design — extractable ledes are how you get cited, not a happy accident.
  • Entity placement is structural — entities land in the sections that need them.

First human approval gate sits here. You approve the outline, drafting begins. Up to 3 manual regenerations per stage.

STAGE 04

Draft

Drafting is the easy part — because everything has already been decided. The writer takes the contract, the outline, the stat bank, and your voice profile, and writes the article. ~2 minutes for 2,000 words.

Default model

The latest Claude Sonnet (claude-sonnet-4-6). Workspaces can override to Claude Opus (1M-token context for very long content) or Claude Haiku (cost-optimized).

Voice profile

Not a tone slider. During onboarding, Woord deep-crawls your site and extracts your tone, vocabulary, sentence patterns, point-of-view, and hedging frequency. Injected as STANCE EXECUTION rules.

Custom instructions

Workspace-level or per-article. "Never recommend competitor X." Products that must appear. Topics to avoid. CTA targeting. All marked HIGH PRIORITY in the writer's prompt.

Required elements

Every draft enforces: meta description in YAML frontmatter, TL;DR after intro (3–5 bullets, ≤20 words each), keyword in H1/first 100 words/conclusion, no duplicate summaries.

YMYL handling

On health/finance/legal/news topics, drafts cite NIH, Mayo Clinic, CDC, government sources directly via Perplexity — not random blogs.

What this means for you

  • It sounds like you. Voice profile from your real site, not generic SEO-speak.
  • You can steer it. Custom instructions are HIGH PRIORITY, not buried.
  • Formatting is guaranteed by the system prompt, not left to chance.

Latest Claude Sonnet by default. Long-form override: Claude Opus 1M context. Cost mode: Claude Haiku.

STAGE 05

Edit

After the draft, an auditor + editor pass identifies weak sections and rewrites them with targeted prompts. A deterministic shape verifier confirms required elements (tables, numbered steps, decision blocks) are actually present.

Auditor agent

Reads the draft against the contract and flags structural and tonal weaknesses — sections that wandered, ledes that aren't extractable, voice drift.

Editor agent

Rewrites the flagged sections with focused prompts that target the specific issue — not a full-article regeneration that risks regressing strong sections.

Shape verification

Rule-based check: did the comparison actually produce a verdict? Did the guide produce numbered steps? Did the listicle items each have Best for / Pros / Cons / pricing? Catches structure misses before scoring.

What this means for you

  • Targeted rewrites, not full regenerations — the strong sections stay strong.
  • Format compliance is checked deterministically, before any LLM scoring.

Editor + shape verifier. The improvement loop runs again automatically if quality scores miss in the next stage.

STAGE 06

Quality Gate

Every article is graded on four axes — Readability, Humanity, SEO, AEO — before it's allowed near "Ready". If any score falls short, the editor rewrites the weak sections automatically. This is the layer that stops AI slop from reaching production.

Readability (≥80)

Flesch-Kincaid grade level, sentence-length variance, passive voice percentage, adverb density. Google demoted dense content years ago.

Humanity (≥80)

AI-writing tells: hedging patterns, banned phrase list (delve, unlock, tapestry, navigate the…), first-person rule violations. Banned word list is workspace-customizable.

SEO (≥80)

Primary keyword density in title/H1/H2s/first 100 words/conclusion, H2 term coverage, body term placement, meta description, internal/external link distribution.

AEO (65–80 by format)

Extraction surface density (tables, lists, blockquotes, definition blocks, how-to steps), extractable lede presence. The score nobody else has — measures readiness for ChatGPT, Perplexity, and Google AI Overviews.

What this means for you

  • No bad articles ship. If quality is below bar, the system fixes it before you see it.
  • Honest scoring — Readability 92, Humanity 81, SEO 88, AEO 76 — not a vague thumbs-up.
  • All scorers are rule-based. No per-article LLM cost on quality checks.

AEO is unique to Woord. No competitor scores explicitly for AI Overviews extractability.

STAGE 07

Compliance

Beyond the four numeric scores, an LLM contract auditor reads the article against the brief and grades issues high (blocks publish) or medium (should fix). Eight distinct violation types are caught.

Contract violations

Breaks the declared stance, intent, differentiation, or warnings from the brief.

Structure violations

Does not follow outline headings, order, or scope.

Fact risk

States or implies a fact not present in facts_allowed. The hard line on hallucination.

Avoid violations

Covers, implies, or recommends something on the topics_to_avoid list.

Tone drift

Tone does not match the declared stance from the voice profile.

Decision non-divergence

Comparison or listicle picks no winner — "both have their merits" is forbidden by regex pattern.

Format adherence

Comparison must pick a winner. Answer must open in ≤60 words. Guide must have numbered steps.

Commercial completeness

Listicles and comparisons must have Best for, Pros (≥2), Cons (≥1), pricing line, summary table, methodology, and a verdict — for every item.

What this means for you

  • High-severity issues block publishing. You don't have to catch them.
  • Fact-risk auditing catches unsourced statistics before they ship.
  • Hedging that hides a recommendation gets flagged automatically.

Second human approval gate sits after compliance. Approve the draft → ready for publish.

STAGE 08

Images & Publishing

A finished article isn't just text. Woord generates schema markup, hero images, infographics, and internal link suggestions, then publishes directly to one of 12 destinations.

Schema markup

JSON-LD generated automatically — Article/BlogPosting, BreadcrumbList, Organization, FAQ (when an FAQPage section exists), HowTo (for guides), ItemList (for listicles). No LLM cost.

Image generation

Google Gemini — Nano Banana Pro for text-heavy images, Flash for conceptual. Hero, inline, and infographics. 15 style presets from minimal to editorial to photorealistic.

Internal linking

Strategist agent suggests up to 15 internal links per article using a 3-tier algorithm: cluster-based → intent-based → keyword overlap. You review and accept what fits.

Publish anywhere

12 destinations with real adapters: WordPress (Plugin, REST API, .com), Shopify (App, REST), Notion, Ghost, Webflow, Wix, Framer, Feather, custom webhook. One endpoint, twelve targets.

What this means for you

  • No SEO plugin needed — schema, meta description, internal links all generated.
  • Hero images included — no stock photo subscription, no design team.
  • Approve in Woord, live on your site in seconds. No copy-paste.

12 CMS destinations, all with real adapters (200–1,000 lines per integration, not stubs).

Most “AI writers” are one-shot prompts. They produce slop because no single prompt can hold research, structure, voice, fact discipline, and SEO compliance at once.

Woord splits the job into specialists — exactly the way a real content team does — and gives each one a contract from the previous stage.

See it run on your keyword.

Closed beta. We're working with a small group of teams while we sharpen the system.