Contents
For one client, we built an SEO content engine that turns the content process, from keyword gap to internal linking, into a fixed workflow: research, briefing, drafting, fact-checking and link suggestions run automatically, while approval stays firmly with a human. In this post we show how the engine is built, which decisions we made along the way and why, what typically goes wrong in projects like this, and how to tell whether an engine makes sense for your business.
The short version first: a content engine is not a text generator. It is a process in which AI takes over the routine and people keep the responsibility. Flip that around and you quickly produce a lot of text, but little that ranks or gets cited.
Starting Point: Content Fails on Consistency, Not Ideas
We see the same pattern in many companies. There is a topic list, often a good one. There are experts who know a lot. And there is an editorial calendar that stalls after a few weeks as soon as day-to-day work takes over.
The reason is rarely a lack of motivation. A single good post consists of many small steps: checking search terms and intent, looking at what already ranks, building an outline, gathering expert input, writing, checking facts, adding internal links, setting metadata, entering it into the CMS, getting it approved. Each step is small on its own, but together they make half a project. And every handover between two people is a point where work gets stuck.
The result is irregular output: phases with several posts, then nothing for a long time. For SEO this is doubly unfortunate, because topical authority comes from connected, regularly extended content, not from occasional bursts. The goal of the project was therefore not “more text”, but a workflow that keeps running even when nobody has time to push it.
The Architecture Step by Step
The engine breaks the process into clearly separated stages. Each stage has a defined input, a defined output and a check before the next one begins.
// From keyword gap to approval-ready draft.
async function runContentEngine(topicCluster: Cluster) {
const gaps = await findKeywordGaps(topicCluster); // search gaps & intent
const brief = await buildBrief(gaps); // structured briefing
const draft = await writeDraft(brief); // on-brand draft
const checked = await factCheck(draft); // sources & consistency
const linked = addInternalLinks(checked); // internal linking
return queueForReview(linked); // human approval
}
1. Keyword gaps and search intent. The engine compares search demand, competitor rankings and the site's own content inventory. The last part matters most: before proposing a new topic, the engine checks whether an existing page already covers the search intent. If it does, the result is an update task instead of a new post. This prevents two of your own pages from competing with each other.
2. Clustering and prioritization. Terms with the same search intent are grouped into one topic. Prioritization is based on business relevance, existing impressions in Search Console and the format Google actually shows for the query. A guide, a comparison page and a landing page are different assignments.
3. Briefing. The brief is the engine's real product. It contains the target question, search intent, format, outline, required questions, approved sources, internal link targets, the desired next step for the reader and, crucially, a list of questions for the client's subject experts. It states which experience, which examples and which proprietary data the post needs so that it becomes more than a summary of the internet.
4. Draft. Writing happens section by section along the brief, using a style guide, a glossary and a list of claims the brand must not make. Every section starts with the answer, not with the lead-up.
5. Fact-checking. A separate step pulls every factual claim out of the draft and checks it against the sources in the brief. Anything that cannot be supported gets flagged, not silently deleted. The editor sees the list and decides.
6. Internal linking. The engine knows the entire site inventory and suggests links in both directions: from the new post to relevant existing pages, and from existing pages to the new post. It also checks that every link target resolves.
7. On-page optimization. Title, meta description, heading structure, alt texts and structured data are suggested, not forced. A title that contains the keyword but makes nobody click is not a win.
8. Human approval. The finished draft lands as a draft in the existing CMS, together with the list of flagged claims and link suggestions. It is only published after approval.
9. Feedback loop. After publication, Search Console data flows back into stage 1. Pages with many impressions and few clicks, or with falling positions, become tasks again. So the engine does not just write new content, it also maintains the existing inventory.
| Stage | What the engine does | What the human does |
|---|---|---|
| Keyword & intent | find gaps, check inventory, cluster | confirm or drop topics |
| Briefing | structure, required questions, sources, link targets | provide experience input |
| Draft | on-brand first draft | sharpen, add examples |
| Fact-checking | extract claims and check against sources | decide on flagged claims |
| Linking & on-page | suggest links, metadata, structure | accept or reject |
| Approval | provide the draft in the CMS | publish or send back |
The Decisions Behind It
Every content engine is a sequence of decisions. These five shaped the architecture.
Mandatory approval instead of full automation. Technically, publishing posts directly would have been easy. We deliberately chose not to. The reason is not distrust of AI, but accountability: a text published under a company's name needs someone who stands behind it. The price of this decision is that approval sets the pace. That is intentional.
The brief is the product, not the text. Many content automations start with a prompt and hope for a good text. We put most of the logic into the brief. A good brief leads to a usable post even with an average draft. A bad brief cannot be rescued by any draft.
Writing and checking are separate. The step that writes the draft does not check it. Fact-checking is its own step with its own task and its own standard. Whoever checks their own work misses the same mistakes twice.
Existing content before new production. Where an existing page already generates impressions, an update is almost always the faster lever than a new post. The engine therefore treats updates as equal-ranking tasks.
Integration instead of a new tool. The editorial team keeps working in its CMS. The engine delivers there. An additional tool that nobody logs into would not have made the process more consistent, just one handover longer.
Quality, E-E-A-T and GEO: Why the Content Stays Citable
Google has stated its position clearly: what matters is whether content is helpful, not how it was created. At the same time, since 2024 Google treats mass-produced content without added value as spam, regardless of whether humans or machines wrote it. An engine that only produces volume runs straight into that policy.
For AI assistants such as ChatGPT, Perplexity or Google AI Overviews, the same logic applies even more strictly. An assistant cites a source when it contributes something other sources lack, and when a passage can be cleanly lifted out. We described how classic SEO shifts as a result in Generative Engine Optimization (GEO). For the engine, we derived fixed rules from this:
- Experience is a required field. Every brief includes questions for the subject experts. Without their answers, a post does not go to drafting. Experience, own examples and own data are what AI systems cannot find elsewhere.
- Answer first. The core question is answered in the first sentences. Every section makes sense without the rest, so it works as a citation.
- Verifiable, not promotional. Concrete conditions, limitations and sources instead of marketing language. What cannot be verified gets picked up less often and hurts credibility.
- Real author, real date. Posts appear under a person with subject expertise, with a visible publication and update date.
- Technically readable. A clean heading structure, structured data and a website that AI crawlers are actually allowed to read.
Whether this works cannot be answered with rankings alone. It takes a fixed set of customer questions asked repeatedly across several assistants. How such a measurement setup is built is described in Measuring AI visibility. And because assistants build their recommendations largely from third-party sources, your own content alone is not enough. We cover the other signals that count in Get recommended by AI assistants.
The Results
- Consistent, predictable output instead of content bursts with long gaps.
- Faster production per post, because research and first drafts run automatically.
- Topic coherence and internal linking that build organic visibility over time.
- The editorial team stays in control of tone, facts and publication.
What We Learned
The following points are typical pitfalls with content engines. We built the architecture to avoid them, and we recommend that anyone planning a similar system account for them from the start.
Without expert input, everything sounds the same. A model writes fluently about almost any topic, but it writes what the average of all sources says. The difference between an interchangeable post and a citable one is almost always a sentence that only someone inside the company can write.
Sources belong in the brief, not in the post-check. If you only look for evidence after writing, you are looking for evidence for a finished text. If the sources are fixed beforehand, the text grows out of them.
“Gaps” are often not gaps. An automated keyword analysis likes to surface topics the site has long covered, just with different wording. Without checking the inventory, an engine produces cannibalization at scale.
Style cannot be described with adjectives. “Professional but approachable” is in every style guide and achieves nothing. What works are examples: good paragraphs, bad paragraphs, banned phrases, the company glossary.
Approval has to be easy, or it becomes the bottleneck. Someone reviewing a long draft without pointers reviews superficially. An approval view with flagged claims, sources and link suggestions turns review into a decision rather than a search.
Automatic links need a limit and a purpose. An engine that links every topical overlap makes texts unreadable. The rule: a link has to help the reader at that point, or it goes.
The engine needs maintenance. Models change, search results change, the company's offering changes. Brief templates, style guide and checking rules are not a one-off setup but ongoing work.
Is a Content Engine Right for You?
An engine is not worthwhile for every company. These questions help you decide:
| Question | Speaks for an engine | Speaks against |
|---|---|---|
| How regularly do you need content? | continuously, across several topic clusters | a few posts a year |
| Is your topic area clearly defined? | yes, with visible search demand | broad, shifting, no focus |
| Is there internal expertise that can contribute? | yes, and it is reachable | no, or never available |
| Who approves? | a named person with time for it | unclear, or “everyone” |
| Do you have access to Search Console and analytics? | yes | no, and nobody can set it up |
| Do you use a CMS with an API? | yes, drafts can be created | manual editing only |
| Do you need several languages? | yes, with the same topic structure | no (not a dealbreaker, but less leverage) |
If you hesitate on the approval question, that is the most important signal. An engine without clear approval ownership produces either a backlog or texts nobody stands behind.
What Drives the Effort
Concrete prices cannot be stated seriously without knowing the starting point. The drivers, however, are easy to anticipate:
- Data access and data quality: Search Console, analytics, a keyword tool and a complete crawl of the existing content.
- Number of topic clusters and formats: guides, landing pages and comparison pages each need their own brief templates.
- Languages: each language needs its own search terms, its own style guide and its own review. A translation is not a localized post.
- CMS connection: the more open the API, the fewer manual handovers.
- Style guide and glossary: built once with examples, then maintained.
- Time from subject experts: the most important and most often underestimated factor. Without their input, content stays generic.
- Operations: approval, monitoring, adapting the engine to new models and search results.
In Short
A content engine turns content into a process rather than a project. It takes over research, structure, first drafts, fact matching and link suggestions, and it keeps the workflow running when nobody has time in the day-to-day. Judgment, experience and responsibility stay with people.
How this approach carries over to other workflows is shown in Process automation for Swiss SMEs and AI agents in the enterprise. How weekly content work, measurement and prioritization fit together in one project is shown in the solar provider case study. And how content work feeds into our ongoing service is laid out on the SEO agency service page.
Want to make your content predictable without giving up quality? Book an introductory call.
