The Engine Handbook

Point Bold Pilot at a URL and it drafts the business context itself — name, description, audiences, competitors — for you to correct rather than compose. One primary language per site (publishing in several splits a young site’s authority), Search Console connects by OAuth, and the first keyword fill runs during setup so the calendar is never born empty.

The engine continuously measures candidate keywords against their live results page and approves only what this site can win — intent-classed, difficulty-capped, cannibalisation-checked (details: Report 2). The Content Planner turns that bank into a month: strongest first, spread across subjects, at most 15% definitional, every day’s pick re-verified at handout. Move, remove, or add your own topics any time. Topping the bank up is deliberately in your hands, not on a timer: when unused topics fall below a third of a month’s runway, the dashboard says so and a one-click scan refills it.

Each article is planned against its real SERP (length budgeted from the winners’ median), written in parallel parts, and forced through an editorial gate: banned-pattern rules, unsourced-statistic deletion, citation verification, sentence-rhythm floors (details: Report 3). Content types include guides, listicles/round-ups, and copy-ready template pages with a download button (held to roughly one template per five articles). A per-site founder note— first-hand experience only the owner can write — is woven into at most one section per article, verbatim facts only; competitor-named pages carry hard tone limits (honest strengths, linked claims, never “X is dead”). Covers and first-party data charts are code-drawn and deterministic; the three or four mid-article photos are real photographs matched to each section, from a library licensed for free commercial use.

Hosted blog articles go live instantly; customer sites receive a signed webhook (markdown + HTML + cover/chart PNGs), with WordPress, Ghost, Shopify, Webflow, Wix, Notion and Feather connectors and dev.to/Hashnode syndication. “Published” is only stamped when a destination actually accepted the article, and Google is notified three ways: sitemap submission, the Indexing API, and IndexNow.

Search Console data rendered where decisions happen: clicks, impressions, CTR and average position with 28-day deltas, a daily traffic line, top queries per article, Core Web Vitals, and per-article gate diagnostics in the library. The keyword bank’s health (approved supply, source quality) reads from the same screen.

Every Monday, four maintenance questions per published article — under-clicked, sliding, invisible-with-a-sibling, never-indexed — produce evidence-backed proposals. A human accepts or rejects (with a reason); accepted items get an automatic 28-day outcome reading, and a retitle that made things measurably worse is flagged for reversal by a pre-registered harm limit. Queues earn automation only past 50 decisions at 90% acceptance; merging never automates.

One plan with a monthly article quota, reserved atomically per article and refunded on failure. A public demo, free SEO tools, and competitor comparison pages ride the same site. Everything the engines measure is stored as diagnostics — quality trends are read from records, never re-counted by hand.

Everything starts from one frozen profile per site: core topic, niche, audiences, 60 seed phrases, competitor domains. At least a quarter of the seeds must be decision seeds (“X alternative”, “X pricing”, “X vs”) — the supply line for buying-intent traffic. Half the list is the audience’s problems in the audience’s words; industry jargon is banned.

Before spending, the engine estimates how many judgements the topic gap needs and plans rounds. Budgets are per provider: row-priced keyword data and cheap SERP calls carry separate ceilings. Judge waves are sized to what the request’s model budget can still afford — a wave the ceiling cannot pay for is never launched — and a wave cut mid-flight keeps its finished verdicts: they are banked, the run stops cleanly, and the next call continues instead of re-buying the same answers.

• Seed expansion: each seed opens into real queries containing that phrase; seeds are validated against autocomplete first so dead phrases cost nothing.
• Search Console channel: content gaps — queries whose impressions land on a generic page with no dedicated article. These carry a measured-demand marker.
• Competitor channel: keywords the rivals rank for; dig history expires per domain, marketplaces are excluded, and a channel that has never produced gets promoted.
• Difficulty relaxation: when supply dries up, the difficulty ceiling loosens in controlled steps; over-ceiling rows are recorded so they are not re-read every round. What comes back is a measured dose — the best-ranked slice the target arithmetic says it needs, never the whole set-aside pool in one billed breath.

Every candidate is reduced to a root (inflections and word order folded); the bank is unique per (user, site, root).

Candidates go to the model in waves. The judge answers two things: keep/drop (should this site write about it) and, for keeps, the funnel class — transactional / commercial / definitional. Relevance asks “who reads this”; class asks “why they searched”. The judge prompt is fingerprint-guarded: it cannot change silently.

• Dictionary check: a page one full of dictionaries means a definition lookup — reject.
• Winnability: giant seats are counted. The giant list is static + per-site dynamic (domains that keep appearing in this site’s SERPs join automatically) + a monopoly penalty when one domain holds 3+ seats (capped; forums exempt).
• Accessible slot: is there at least one page-one domain with authority ≤300? Median authority misleads; one enterable seat decides. Worth +6.
• Composite score: relevance 0.35 · winnability 0.30 · funnel class 0.20 (TX 100 / CM 60 / DEF 10) · volume 0.15, plus measured-demand +8 and slot +6.

Difficulty and volume are bought in bulk for keywords on the approval path. A late difficulty read never blocks approval (deliberate fail-open), but if it comes back above 35 the handout gate holds the row. High-winnability shortlists get AI Overview observations; a verdict needs two consecutive agreeing looks.

Spellings that one article can serve together are grouped; the strongest member becomes the primary target, the rest ride as secondaries — the article covers the topic, not one phrasing.

candidate → judged → approved / rejected (with a reason: irrelevant, unwinnable, reference-query, quarantine kinds) → used once written. Reject reasons stay separated because “why did we come up short” is answered there — folding dictionary lookups and low winnability into one bucket once inflated a floor-relaxation analysis.

• Today’s-rules re-check: definitional shape, winnability under 42, difficulty over 35 — an old approval is not enough.
• Family gate vs written: a candidate sharing ≥4 page-one URLs with an already-written article is refused (threshold calibrated on 24,829 real pairs).
• In-delivery family gate: the picks are also checked against each other — one of a family stays, the rest drop.
• Subject spread: at most 3 per reader-subject; the quota relaxes before the delivery shortens.
• Funnel quota: at most 15% of a delivery may be definitional.
• Competitor dose: topics leading with a rival’s brand name are capped at ~20% of a delivery — wanted pages, but a site must not read as a comparison farm. This cap never relaxes.
• Delivery memo: a shortlist that already passed the final relevance pass and the set-audit clean is fingerprinted (set + rules epoch + judge); an unchanged set skips both on later rounds instead of re-buying the same judgement.

An external scheduler pokes the pipeline every 15 minutes. The endpoint is now a dispatcher: it claims due rows and hands each article its own process — its own time budget, its own memory, its own failure record.

• Duplicate guard: if an article already covers this root, generation refuses before money is spent.
• Quota reservation: the monthly article right is reserved atomically and returned on failure.
• Destination-ready: no connected channel → the article is not generated; the date holds until a destination exists. Writing into the void burns a credit.

Three memories load before writing: business context; the skeleton summary of the last 8 articles (opening/closing styles, stats-block usage — the variety rotation’s memory); and internal-link targets from this site’s published articles. Pages inside a live experiment lock cannot become link targets — you don’t pump authority into a measurement.

• The top 6 pages of the keyword’s SERP are actually fetched; sentences carrying figures are extracted (max 10, 3 per page, noise filtered).
• Every item carries its source URL and a detected publication year — a study older than 3 years must be dated in the article’s own sentence.
• On our own sites the pack gains first-party platform aggregates (≥5-site privacy floor).
• Fetched pages’ word counts feed the length budget.

• Title/meta rules: don’t copy the SERP’s dominant title pattern; the answer lands in the meta’s first ~120 characters; a year appears only when freshness is truly part of the promise.
• Length budget derives from the SERP: winners’ median ×0.8–1.2, as a ceiling-guard, not a quota — stopping under it when coverage is done is correct. Template pages are iron-banded at 400–800.
• A gap question — something the ranking pages fail to answer — is chosen and assigned to a section.
• Link and evidence assignments happen in the plan, because parallel writers cannot see each other.
• Skeleton rotation picks the stats-block and closing move against the last-8 summary.

Intro, sections, and closing+FAQ are written in parallel calls sharing one cached system prompt: the format spec (editorial / round-up / template), voice rules (deliberate unevenness; fragment-punch beats capped at 2 per article; bold capped at 1 per paragraph), evidence rules, and banned patterns.

Every part passes these checks; violations return in one combined revision call:

• Banned patterns (~25 rules) — survivors of revision get a deterministic sweep for the deletable class.
• Unsourced statistics (institution + number, no link; prices/years/acronyms exempt) — survivors are deleted sentence by sentence, on record.
• Citation verification: a linked claim must match its assigned source and figures; mismatches go to a human review queue.
• Sentence rhythm (editorial): every 3+-sentence paragraph needs one sentence under 8 words and one over 25; a rhythm-only second pass runs when needed and is rejected if it regresses the floors.
• Bold overuse, duplicate source+figure, unlinked prices (round-ups).

All counts are stored on the article — the gate cannot fail silently.

Parts join; product links are repaired; punctuation debris from revision seams is cleaned deterministically; cross-section duplicate citations are diagnosed; the skeleton is measured for the next article’s rotation.

• Cover: deterministic, code-drawn SVG (five rotating palettes) with a background-stroke halo behind the headline — and on long titles the artwork steps back (3 lines → 55%, 4+ → 35% opacity) so ink never fights ink; customer copies get a PNG from a raster route.
• Section photos: three or four per article (two or three beside a data chart, which counts as one), each searched from a literal scene the planner writes for that section and picked deterministically per slug — a regenerated article keeps its pictures, and no photo repeats within one. The library's own description of each photo travels as its alt text. Articles from the 30 Aug – 8 Sep 2026 window keep their code-drawn, topic-aware vignettes.
• Video: embedded only when one truly fits; on our pages the player loads on click (facade), external bodies get a link.
• Data visual: on our own sites, first-party aggregates render as a branded bar chart — SVG on the page, PNG for customer copies.

The article saves with full diagnostics: gate counts before/after, skeleton signature, gap question, chart data, assembly fixes, target vs actual words. The weekly quality cassette and the rotation read these back.

Hosted articles go live directly (page, index, sitemap refreshed); customer sites receive a signed webhook (markdown + HTML, cover PNG, chart PNG, video link). “Published” is stamped only when something actually arrived somewhere. Then Google is told: sitemap + Indexing API + IndexNow.

• Quality cassette: recent articles are scanned weekly against the audit baseline.
• LLM judge (rubric v2): scores what code cannot see — does the opening fully answer, does the body keep the title’s promise, how many fragment-punch beats.
• Improvement queues: every Monday, four questions per published article against Search Console — under-clicked (retitle), sliding (refresh), invisible for 90 days (merge), never seen (index check). All proposals: a human decides, accepted items get a 28-day follow-up reading, and no queue automates before proving 50 decisions at 90% acceptance. The merge queue never automates. Retitles also carry a pre-registered harm limit, written before any data existed: normalized CTR down 20%+ on 200+ accrued impressions, or the position 3+ slots worse, marks the item reverted — a decision that counts against the queue’s automation record and asks for the old title back.