One keyword in.
One approved article out.

Every morning, Kourt pulls the last 24 hours of Search Console impressions, clicks, and average position, cross-references them against your competitor scan set, and produces a ranked queue of things worth writing. You never type a keyword into a prompt box. The queue is the prompt.

Opportunities are scored on three axes: impressions (latent demand your site already brushes against), current position (where page-two ranks are the easiest wins), and competitor gap (topics your set covers that you do not yet). A fourth axis, brand fit, rejects anything too far off your voice profile or outside your compliance pack’s scope.

What gets rejected

Not every impression is worth writing for. Triage actively rejects:

• Branded queries. People searching for you by name are already converting.
• YMYL topics without a compliance pack. Medical, legal, or financial claims without an active pack are rejected at the queue. You will see them flagged, not drafted.
• Duplicates. Semantic overlap with your existing library above 85% similarity is deduplicated back to the canonical article.
• Out-of-voice topics. A paediatric dentistry clinic does not get queued for cryptocurrency SEO just because the CTR looks good.

The queue updates hourly. Send-to-pipeline is one click. From there, the next five stages run without you.

The cheap way to generate an SEO article is to re-summarise the top result. Kourt does not. Stage 02 reads the top ten results for the chosen keyword, parses their heading structure, and returns a condensed map of what is covered, what quotes and statistics appear, which sub-intents are addressed, and most usefully what is missing.

From that map, the research agent proposes three angles. An angle is a specific framing, not just a title. It gives the draft its shape: what it argues, who it is for, what it answers that nothing in the top 10 currently does. You pick one, or let Kourt pick the highest-scoring.

What the agent actually extracts

For each of the top 10, stage 02 records H1 and H2 structure, word count, presence of tables and FAQs, cited sources, price ranges, timelines, and any quoted experts. That becomes the coverage matrix. The matrix is what lets the brief writer in stage 03 pick things to include (because they are table stakes) and things to include uniquely (because the gap is real).

You do not see the matrix directly. You see the three angles, and the reasoning for each.

A bad brief makes the whole pipeline fail. Stage 03 produces a structured brief that stage 04 cannot deviate from. It is the contract. The brief is not a paragraph of instructions; it is a typed object with an outline, a target reading grade, phrases your voice profile requires or forbids, compliance rules that apply, and a proposed internal link plan tied to pages that already exist on your site.

You see the brief before it runs. Approve, edit, or reject. If you edit, the changes become part of your voice profile for next time.

What the brief locks in

• Outline. H2 structure with one-line descriptions. The draft cannot invent new H2s; it can only fail back to stage 03 for amendment.
• Target word count. A range, not a number. Stage 05 fails the draft if it lands outside the band.
• Voice profile reference. The version of your profile in effect at the time the brief was cut. If you update your voice profile mid-draft, the draft uses the old version; the next brief uses the new one.
• Compliance pack. Named and versioned. Same principle: if AHPRA guidance updates, drafts already in-flight use the pack they were briefed under.
• Link plan. Five internal links, scored and pre-anchored. Stage 04 must place all five.

If you are on Kingdom, the link plan cross-references your full content graph, including orphan-tier detection and cluster balancing. Other tiers use a simpler proximity-only plan.

Stage 04 is the only stage that actually generates prose. Every earlier stage shaped the contract it works under: the brief says what, the voice profile says how, the compliance pack says what is off-limits. The writer agent reads all three, produces a draft, and then self-checks against the brief before handing off to stage 05.

The voice profile is enforced sentence-by-sentence. Tone triple, sentence-length range, forbidden phrases, preferred phrases, pricing references, reading grade, locale (en-AU vs en-US spelling and idiom). Your profile starts with a template and learns from every approved or rejected draft. On Kingdom, it also learns from your live site. Duchy and below work from the initial template plus your approvals.

The compliance pack runs as an inline linter. When a sentence would trigger a pack rule, the writer rewrites before the sentence ships. If a rule cannot be satisfied without changing the brief, the draft fails back to stage 03 with a note.

What the writer agent refuses to do

• Invent statistics. Numbers come from sources earmarked in the brief, or they do not go in. No "studies show" hedges.
• Paraphrase the top result. Similarity against the SERP set is checked in stage 05; drafts that score above 20% on any single result get flagged.
• Use forbidden phrases.Your voice profile carries a blocklist. "In today’s fast-paced world" (for example) is blocked by default and most users leave it blocked.
• Ship without all the links. If the writer cannot place all five internal links in context, it returns the draft to stage 03 for a link-plan revision rather than ship four.

Draft time varies with length and compliance complexity. A 1,400-word article under the AHPRA pack averages 40 to 75 seconds.

Stage 05 is the most opinionated part of the pipeline, and the most load-bearing. Everything before it produced a draft that could exist. Stage 05 decides whether it should. It runs three independent systems in parallel, not because one is insufficient, but because they catch different kinds of failure.

The first is deterministic rules: things that are either true or false. Does the draft use em dashes? Is the H1 hierarchy clean? Are all external links absolute URLs? These are binary pass/fail, no scoring, no judgement calls. The second is graded dimensions: things that exist on a spectrum. How well does the prose match your voice profile? How strong is the E-E-A-T signal? How effective are the CTAs? These get scored 0 to 100 with commentary. The third is similarity & originality: a statistical check against the open web and a broader human-writing corpus to catch paraphrasing and stock cadence.

Layer 1 · Deterministic rules

Ten rules, re-run on every save. Most workspaces pass 9 of 10 on a first attempt; the last one gets flagged and waits for either an auto-fix or your attention. The rules are intentionally picky: they are the kind of thing a pedantic senior editor would catch before publish.

Layer 2 · Graded dimensions

Deterministic rules catch what is wrong. Graded dimensions catch what is weak. Seven dimensions are scored on every draft, each with specific commentary. Thresholds differ by tier (Kingdom sets tighter floors than Duchy or Page) but the dimensions themselves are the same.

Layer 3 · Similarity & originality

A separate system, run against a different reference set. Similarity checks the draft against the open web and your own existing library. High numbers mean passages are echoing a source, which is a publisher risk even when it is not copied. Originality is the inverse check: how human the cadence reads against a broad corpus of human-authored writing. Both scores are independent of the graded dimensions; a draft can score well on voice match and still fail similarity if it paraphrased a SERP result too closely.

Layer 4 · The blockers list

Anything from the three layers above that the grader judges “must fix before approve” rolls up into a single blockerslist. Think of it as a senior editor’s margin notes. Specific, referenced, and opinionated, not vague.

What happens when a draft fails

If the composite score lands below the tier threshold, or any blocker fires, stage 05 does not reject the draft outright. It branches. Minor failures route back to stage 04 with the blocker list as additional constraints; the writer re-runs with those notes added to the brief. Structural failures (outline does not cover the brief, link plan cannot be satisfied) route back to stage 03 for a brief revision. You do not see any of this by default. You see the finished, clean draft in stage 06.

Each draft is allowed two re-run attempts before escalating to human review as-is. On Kingdom, the grader is tuned stricter and the re-run budget is higher. You can afford more loops because the per-article spend envelope is larger.

Stage 06 is the only stage a person must sit in. Everything before it happened while you were not looking, 60 to 120 seconds of pipeline work, often while you finished a coffee. Now the draft is in your drafts list, the composite score is calculated, the blockers are surfaced, and the grade is attached. You sit down and decide.

The four actions

When you open a draft, the side panel shows the composite score, the deterministic-rule grid, the graded dimensions, the blockers list, and four action buttons. Each does a specific thing, no overlap.

Editing the draft itself

The draft editor is a normal rich-text surface. You can rewrite, delete, restructure, and swap links freely. Most users edit very little and approve: the grader has already caught what it can catch. When you do edit, your changes train the voice profile for the next draft. The system notices which words and phrases you remove repeatedly, and pre-empts them next time.

SEO metadata is yours to own

One area the pipeline deliberately does not own end-to-end: title tag, meta description, and URL slug. Kourt generates a candidate, but the review screen gives you an inline SERP preview with character counts, and it is on you to shape the final snippet. This is the single highest-leverage editing surface on the whole page, and it lives where you can see the live Google preview as you type.

Published drafts go to your connected CMS (WordPress, Ghost, HubSpot, Notion, Shopify, Webflow, or a signed custom webhook of your own) with the approved draft, its metadata, and its internal link plan intact. DOCX download is also available on Kingdom and Empire. Nothing gets rewritten on the way out.

Most content tools treat tone as a slider and compliance as a manual review afterwards. Kourt treats both as typed objects that the writer agent has to consult on every sentence. They are how the same pipeline can produce a relaxed Brisbane car-detailing post in the morning and a peer-reviewed implant journal piece in the afternoon, without the second sounding like a Yelp review or the first sounding like a deposition.

Voice profile · what you configure

A profile is six fields plus a trained model of your edit history. You set the bones once at workspace setup and the system fills in the rest as you approve and reject drafts.

• Tone triple.Three adjectives that name the register. “Direct, warm, expert” reads very differently from “measured, evidence-led, neutral” even though both claim to be professional.
• Sentence-length range. A min/max in words. Short bands (8–18) read snappy; long bands (15–32) read considered.
• Reading grade. Target Flesch-Kincaid, between 6 and 14. Stage 05 fails the draft if it lands two grades off.
• Forbidden phrases.A blocklist that the writer agent treats as physically uncallable. Defaults catch the obvious slop (“in today’s fast-paced world”, “dive deep”, “leverage”); you add your own as you notice them.
• Preferred phrases.The opposite list. Anchor phrases the writer should reach for when natural. Useful for brand-specific framing (“before-and-after consult”, “on-record price guide”).
• Locale, allowlist, pricing references. en-AU vs en-US spelling, the citation domains the grader accepts as authority, the currency and price format the writer is allowed to publish.

On Kingdom, the profile also learns from your live site. Kourt ingests your existing pages, builds a vector index of your real cadence and vocabulary, and the writer consults it on every paragraph. Duchy and below work from the initial template plus your approvals; the learning loop is slower but the same shape.

Same keyword, two profiles

Below: stage 04 output for the keyword “why your booking lead time matters”, drafted twice from the same brief, against two voice profiles that share nothing.

Compliance packs · what ships with Kourt

A compliance pack is a typed bundle: forbidden claims, required disclaimers, attribution rules, and a versioned history so audits have something to point at. The writer agent treats pack rules as non-negotiable. If a sentence cannot satisfy the active pack, the draft fails back to stage 03 rather than ship.

What an AHPRA pack looks like inside

The pack is human-readable when you open it, and the writer agent is told which subsections apply on the current brief. The structure below is the actual shape of the version 2.3 spec the workspace ships with.

Packs evolve when the regulator updates. We track the change and ship a versioned pack within ten business days; your existing briefs stay pinned to the version they were cut under, and your next brief picks up the new one automatically.

Stage 03 produces a link plan for every brief. The link map is what that plan is built against. Every published article on your site becomes a node, every internal link an edge, and the brief writer reads the graph in full before it proposes the next five links. Pages without inbound or outbound links land in an orphans list the day they appear, not the quarter you remember to audit.

Open /graph in your workspace and the map renders live. Hubs (your highest-linked clusters) sit at the centre; orphans glow at the edges. Clicking a node reveals the article, its inbound links, and its outbound links. Clicking an edge shows which brief proposed it and which draft stage placed it.

What the brief writer reads from it

For every new brief, Stage 03 queries the graph for candidate internal links: pages on the same topical cluster, pages with anchor-text proximity to the target keyword, and pages currently sitting in the orphans list. The five top-scored candidates become the brief’s link plan, pre-anchored, before the draft writer ever reads a sentence.

Cluster balancing runs on every publish. If three new articles land in the same cluster within a week, the map flags it before it becomes a thin cluster the next algorithmic update penalises.

On Kingdom, the full live graph is yours to pan and zoom. Duchy gets a view-only snapshot updated daily. Lower tiers see orphan flags on the dashboard but not the map itself.

Answer engines are now the second referrer of buyer-intent research for most consumer brands and the first for some. They do not show up in GSC. They never will. AI Visibility is how you measure them on a schedule rather than ad-hoc by typing questions into a chat window once a quarter.

You define a prompt list, the way you would define a target keyword list. Each prompt runs weekly across three answer engines. The response is logged, a detector decides whether your brand was mentioned, where in the ranking, and which competitors landed alongside.

What you set up

Prompts in groups (e.g. “discovery prompts”, “comparison prompts”, “objection prompts”). Competitors you want tracked by name, so the detector can score competitor density even when you are not mentioned. The brand profile, which the detector treats as a synonym group covering alternate spellings, abbreviations, and the parent company.

What the dashboard tells you

Mention rate over time, weekly. Top competitors cited alongside you, ranked by frequency. Cited source domains, because some answer engines link the pages they synthesised from and that link becomes its own SEO signal worth chasing. Per-prompt drill-in, so the prompts that move the most are the ones you act on first.

On Kingdom, you get 25 tracked prompts and full history. On Empire, 100 prompts and multi-brand support. Lower tiers see the AI Visibility report card on the dashboard with an upgrade pitch instead of the live data.

Most weeks the picking decision is mechanical. You scan the queue, pick three keywords with the best mix of intent, brand fit, and ranking effort, and click “send to pipeline” for each. Autopilot is the option to skip that. Tell it the rules once, and on a schedule it picks the top N candidates and pushes them into the pipeline for you.

Two modes

Dry-run picks the keywords and drafts the briefs, but stops at human review. Drafts arrive in your queue with their briefs attached and wait for you before stage 04 runs. Use this for the first month, until the picks earn your trust.

Auto runs the full pipeline through stage 06. Drafts arrive ready to read, with the same composite score and blockers list a manual run produces. Approval still requires a click; publishing is your call regardless of mode.

Where picks come from

Picks come from your existing surfaces, not from anywhere new. Search Console opportunities your queue has scored. Manual ideas you have filed but not pushed. Topic Opportunities runs that scan your brand profile, published content, GSC data, and AI Visibility gaps for new angles. You toggle which origins are in scope; Autopilot ranks across them and picks the top N.

Pause and audit

A pause-now button is on the Autopilot page for any owner or admin. The next scheduled run is skipped and a banner appears on the dashboard until you resume. Every run logs to an audit table: which keywords were considered, why these N got picked, what happened to each draft. If a draft fails stage 05 twice, Autopilot escalates to your queue rather than rerouting silently.

On Kingdom and Empire only. Lower tiers run the same pipeline manually from the queue. Default picks-per-run is 3 on Kingdom and 5 on Empire; you can dial either down.

Reports turn the workspace’s data into something you can read in three minutes. Some run on a schedule, some run on demand. They land in /reports as a feed and surface in the dashboard panel until you mark them read.

Topic Opportunities · the loop closer

The most-used report in most workspaces. Topic Opportunities scans your brand profile, published library, Search Console data, and AI Visibility gaps in a single pass and returns six to ten new topic angles. Each angle gets a brand-fit score, a one-line rationale tying it to which input fired, and a target cluster. The whole batch gets pushed straight to /ideas, ready to send to the pipeline.

Daily research brief

The 30-second morning read. Your priority pick for the day with a ready-to-write angle and suggested H2s, a 24-hour pipeline pulse (what shipped, what is in review, what needs you), and five supporting starters in case the priority pick is wrong. Lands at 06:00 in your workspace timezone.

Weekly gap analysis

Monday morning, anchored on Search Console. Quick wins (positions 8 to 20 with real volume), declining keywords (dropped five-plus ranks week-over-week), and CTR opportunities (high impressions, low click-through-rate for the page’s current position). Each finding links to the underlying GSC row.

AI Visibility report

Weekly roll-up of the data on the AI Visibility dashboard, formatted as a readable report rather than a chart drill-in. Brand mention-rate trend, top competitors gaining share, the cited source domains the answer engines link to, and the prompts worth retiring because nothing has moved on them in eight runs.

On Kingdom and Empire, all four reports are live. Topic Opportunities is also available on Duchy at one run per week. Lower tiers see report previews on the dashboard with an upgrade pitch instead of live data.

A content pipeline that will do anything you ask of it is worth exactly what “anything you ask of it” is worth. Kourt has opinions. Some of them rule out legitimate use cases; that is the cost of shipping a system that does not embarrass you the week after you connect your CMS.

• No auto-publish. Ever. Approve and Publish are separate buttons, separate clicks, separate timestamps. A draft can sit in your approved-but-unpublished queue indefinitely, and often should.
• Will not draft without a voice profile. New workspaces set up a profile in the first session. You can start with a template, but the brief writer refuses to run until a profile exists. Drafts written from a blank voice are the single biggest cause of generic-sounding output.
• Will not skip the grader to hit a cap. On tight-budget tiers, pipeline work that would blow the monthly ceiling pauses rather than short-circuit stage 05. We would rather pause than publish ungraded.
• Will not write YMYL topics without a pack. Medical, legal, financial, and insurance topics are refused at stage 01 if no compliance pack is active on the workspace. You will see them flagged in the queue, not drafted.
• Will not paraphrase the top result. Similarity above 20% against any single SERP source at stage 05 routes the draft back to stage 04 with instructions to diversify. Drafts that cannot clear the threshold twice are rejected, not smoothed over.
• Will not branch a draft without telling you, on Duchy or above. Kingdom and Duchy workspaces get an audit trail of every branch stage 05 triggered. Page workspaces run quietly.

Pipeline time is bounded and predictable. The numbers below are illustrative: a representative run on a long article with a heavy compliance pack, where total automated time stays under two minutes. Stage 06 is where the calendar takes over: some teams approve in five minutes, some sit on a draft for a week thinking about it. Both are fine.

On Kingdom, stage 04 runs against a larger context window and stage 05 grades against tighter floors. Draft time stretches to roughly 60 seconds, and stage 05 to 15. The total is still under two minutes; the floor on quality is higher.

Page-tier workspaces get one finished article every sixteen days. Paid tiers run on demand: the pipeline kicks off the moment you send a keyword to it. There is no batch window.