Case Study: AI Content Engine, 100 Posts/Month Publisher

A publisher AI content engine case study: an editorial team shipping twelve posts a month manually, with declining traffic and no fact-check discipline, rebuilt the program around an eight-stage AI-orchestrated pipeline and reached one hundred and four posts a month sustained inside six months. Fact-check pass rate landed at ninety-six percent, schema compliance at ninety-one, organic traffic recovered, and LLM citation share climbed materially — without losing editorial voice.

The engagement was a publisher in a mid-sized B2B category with a mature back catalog and a clear authority position that had started to erode under volume pressure from category competitors. The brief: rebuild the engine, hit the velocity target, protect voice, and ship measurable outcomes inside two quarters. The client requested anonymisation; the playbook and the numbers are as deployed.

This case study walks the situation as found, the four approach pillars in the order they were built, the outcomes at six months, and the lessons that replicate cleanly to smaller publishers. Where a section can point to a deeper playbook, it does — the goal is a case study that is actionable, not just illustrative.

The publisher arrived with a mature back catalog of just under seven hundred posts accumulated across six years of editorial output. The catalog had earned a credible authority position in the category — the kind that historically produced enough organic traffic to support the program without paid amplification. By the time of the engagement that position had eroded for eighteen months. Velocity sat at twelve posts a month, three editors, a backlog of half-drafted posts in the production queue, and a growing concern that the category had outpaced them.

The deeper diagnosis surfaced three structural issues, not the volume issue the client initially named. First, no shared brief library — every post was scoped by the editor assigned to it, and consistency drifted post to post. Second, no fact-check chain — verification happened ad-hoc, mostly by the editor who also drafted, with predictable blind spots. Third, no schema discipline — titles ran long, descriptions were inconsistent, and structured data was partially deployed across the catalog. Volume was a symptom, not a cause.

The cultural surface the diagnosis hit was important. The team had a strong editorial identity, a recognisable voice, and a visible reluctance toward AI-augmented drafting. The fear — a reasonable one — was that introducing AI would dilute voice, erode trust with the audience, and trade authority for volume. The brief explicitly named voice protection as a non-negotiable constraint. The engine had to ship volume and protect voice; either alone was not the deliverable.

The pre-engagement audit confirmed one further structural issue. Posts referencing specific vendors, model versions, or pricing had not been refreshed against current values — and the audit found that roughly forty percent of the catalog contained at least one stat or reference that had aged out. The catalog was not just under-amplified; it was actively misleading on the posts that drew the most traffic. Refresh cadence was going to be a phase-one priority, not a phase-three one.

The first artifact the engagement shipped was a stage diagram — eight stages from research to amplification, each with a named owner, an exit criterion, and a target cycle time. The diagram was deliberately simple: a hundred-post-a-month engine cannot run on a thirty-stage pipeline. Each stage had to fit on one row of the operations dashboard, and each had to be measurable in cycle hours rather than narrative description.

The stages clustered into four functional groups. Research and briefing constrained the input; drafting and fact-checking produced the verified output; schema and staging review enforced the publishable shape; publication and amplification extracted the return. Each cluster had a lead — editorial, content engineering, schema engineer, growth — and the leads met weekly. The pipeline was treated as a production line, not a creative workflow.

The cycle-time targets were the operational backbone of the pipeline. Research and brief: four hours per post end-to-end, mostly editorial. Draft and fact-check: two hours per post for most content types, longer for deep guides. Schema and staging: thirty minutes per post combined, with the schema gate enforced in CI rather than by reviewer goodwill. Publish and amplify: an hour per post, mostly social and newsletter scheduling. The total — roughly seven-and-a-half hours per post — was what made the velocity target arithmetically possible across the existing headcount.

For the deeper architectural rationale behind staged pipelines in publisher contexts, the engagement leaned on our content engine service playbook. The publisher's eight-stage shape is a direct instantiation of that playbook with cluster-leadership adapted to their editorial structure.

The fact-check chain was the artifact that did the most structural work in the engine. The publisher arrived with no chain at all — verification was ad-hoc by the drafting editor — and the engagement replaced that with two tiers that ran in sequence. Tier one was automated: link-rot checks on every source URL, claim extraction against the pre-loaded sources, banned-phrasing detection. Tier two was human: a verification gate run by an editor who had not drafted the post, against a ten-point checklist.

The structural lever was upstream placement. Source URLs were verified before drafting and pre-loaded into the brief, so the drafting model was constrained to sourced claims rather than free-generating claims that would need post-hoc verification. The anti-fabrication rule was explicit in every brief template — "do not invent metrics, quotes, case studies, company names, or product features; if a claim cannot be sourced to the pre-loaded URLs, omit it or flag it for editorial review" — and the rule was tested against the model output in tier one before it reached the human gate.

The chain protected against the failure mode that publishers fear most when introducing AI — fabricated metrics or invented quotes shipping into a piece that carries the publisher's editorial brand. Two examples surfaced in the phase-one weeks where tier one caught a fabricated stat that the drafting model had generated despite the pre-loaded sources covering the same topic. In both cases the source URLs were updated to include a citation for the corrected stat, and the drafting prompt was tightened. The chain was not just a filter; it was a feedback mechanism that improved the upstream brief over time.

The cost of the chain was measurable and bounded. Tier one added roughly six minutes of CI time per post. Tier two added roughly twenty minutes of editor time per post. On twenty-five posts a week, that is about ten editor hours a week dedicated to verification — substantial, but well inside the budget that the velocity gain produced. Verification was treated as a first-class stage with its own budget line, not as a slack-time activity squeezed between drafting and publish.

Schema failure is the silent killer of publisher performance. The publisher arrived with no schema validation in the pipeline — title length, description length, canonical URLs, structured data, and image alts were enforced by editor goodwill, which under volume pressure became no enforcement at all. The remediation was AST-level validation in CI, configured to block the merge on failure rather than warn. The shift from warn-only to blocking was the single highest-leverage move in the schema cluster.

The choice matrix below covers the four schema controls the engagement deployed, why each was prioritised, and what the failure mode looks like when the control is missing or warn-only. Each control was implemented in the same week of phase two; the cumulative effect on compliance was visible within the first three weeks.

The downstream effect of the schema cluster surprised the team. The hypothesis going in was that schema compliance would lift SERP performance — which it did. The unexpected second-order lift came from LLM citation share: the consistent structured data and metadata gave large-language models a cleaner surface to cite from, and the publisher's posts started appearing as citations in AI-search answers at a measurably higher rate. The team had not optimized for that surface; the schema discipline produced it as a downstream effect.

The remediation pattern the publisher converged on — AST validation in CI, blocking at merge, no warn-only — is the pattern we recommend to every publisher engagement. For the staged rollout plan that lands an engine at this level of schema compliance inside ninety days, our AI content engine launch 30/60/90 day plan covers the milestones in order — schema discipline lands in phase two for a reason.

A six-year back catalog without an active refresh cadence decays as sources move, statistics age, vendor names change, and pricing shifts. The pre-engagement audit found that roughly forty percent of the catalog contained at least one stat or reference that had aged out, and the most-trafficked posts were disproportionately represented in that forty percent — high-traffic posts attract more refresh-triggering external change than low-traffic ones. Refresh cadence was the highest-leverage stage for the back catalog and the cluster that produced the fastest recovery curve in the engagement.

The cadence ran on three triggers, all feeding one weekly queue with one named owner. Time-triggered quarterly was the default — every post in the catalog hit a refresh check at least once a quarter, regardless of any other trigger. Model-version overlay surfaced posts referencing specific model versions, vendor pricing, or feature availability when the underlying reference changed. Event overlay surfaced pillar posts referencing competitive launches or industry events. Three triggers, one queue, one owner — the structural simplicity of that arrangement was a deliberate design choice.

The refresh queue produced the fastest visible win of the engagement. By week six, the most-trafficked twenty posts had been refreshed and the organic traffic curve on those posts inflected — the high-leverage subset of the catalog stopped leaking traffic before the new-post velocity had even ramped to half its target. The lesson the publisher internalised: refresh is a first-class stage with first-class ROI, not a janitorial activity squeezed in around new-post production.

The quarterly report that the owner produced — back-catalog traffic trend, refresh volume, highest-lift post, lowest-lift post and reason — became the artifact the editorial team used to plan the next quarter's content calendar. Refresh patterns surfaced gaps in the brief library (content types that aged faster than expected), gaps in the source list (sources that link-rotted more often), and gaps in the amplification rhythm (refreshed posts under-amplified relative to new posts). The cadence was a learning loop as well as a maintenance loop.

Six months in, the engine had landed the brief. Velocity at one hundred and four posts a month sustained — eight times the manual baseline — across the existing editorial headcount with the addition of one content engineer and one schema engineer. Fact-check pass rate held at ninety-six percent at run-rate volume. Schema compliance ran at ninety-one percent in CI, with the remaining failures mostly stylistic edge cases the team chose to triage rather than auto-block. Organic traffic recovered the prior twelve-month decline inside four months and continued to climb through month six. LLM citation share — measured by sampling AI-search answers in the category — moved from an unmeasured baseline to a measurable share of cited sources.

The outcome that mattered most to the client was voice. Reader feedback through the comment system and the newsletter remained consistent with pre-engagement patterns — the engine had not produced a perceptible change in editorial tone. The brief-library voice constraints carried the house style across more than six hundred posts in the run-rate quarter without per-post tonal review. The hypothesis the engagement had set out to test — that voice could be encoded in templates rather than enforced by editor-by-editor judgment — held under volume pressure.

The numbers behind the bars matter less than the pattern across them. Every cluster the engine prioritised hit its target inside six months, and the clusters were prioritised in the right order — pipeline first, fact-check second, schema third, refresh fourth. Reordering would have produced different numbers. Skipping the fact-check chain to chase velocity would have shipped volume without verification discipline; skipping schema for amplification would have left silent SERP failures in place; skipping refresh would have continued the catalog decay even as new posts shipped.

The engagement also surfaced two outcomes the brief had not asked for. First, content engineer satisfaction — the editor reports were universally positive on the brief library and the fact-check chain, both of which removed the most frustrating parts of the prior workflow. Second, cost-per-post — measured stage-by-stage, the average dropped by roughly sixty percent from the manual baseline, not because any single stage got dramatically cheaper but because the cumulative effect of brief reuse, CI-enforced gates, and standardized amplification compounded across stages.

The case study sits at the upper end of the publisher size range — a six-year catalog, three editors pre-engagement, authority position in the category. Most publisher engagements are smaller, often a single editor with a shorter catalog and no schema engineer to hand. The patterns that replicate cleanly to smaller publishers are not the headline numbers — they are the structural decisions underneath. Five lessons, each of which holds at smaller scale, frame the replication shape.

The pattern across the five lessons is consistent. The scale-dependent moves in the engagement — model routing across three frontier models, two-engineer staffing, weekly refresh queue at twenty-five posts a week — do not replicate to a single-editor publisher. The structural moves underneath them — staged pipeline, upstream fact-check, CI-enforced schema, voice in templates, quarterly refresh — do, and they are what produced the outcomes. A publisher shipping twenty posts a month can run the same five disciplines and land at a smaller version of the same compounding curve.

For publishers ready to translate the lessons into a staged ninety-day launch sequence, the companion playbook is the AI content engine launch 30/60/90 day plan. For publishers ready to audit an in-flight pipeline against the granular gates the engagement enforced, the companion checklist is the AI pipeline quality audit eighty-point checklist. The case study is the proof point; the playbook and the checklist are the implementation artifacts.