The self-driving content machine

At a glance

• Starting point: Long-tail SEO in a niche needs many well-researched articles in several languages, complete with imagery and linking, at a volume that can't be produced manually.
• Solution: We built a modular content machine, orchestrated on n8n, that generates articles, illustrates them, translates them and even turns them into a podcast, running unattended round the clock.
• Outcome: A consistent, continuously growing magazine with several thousand articles in up to six languages, with no team growing to match.

Part 1: blog, SEO and automatic illustration. Part 2: from article to podcast.

• Part 1, text and image: generates, refines, illustrates, links and translates blog articles largely without human intervention. The technical centrepiece is a multimodal image vector database that classifies every photo with AI and later places the right image at the right spot in the article, automatically.
• Part 2, audio: turns the same articles automatically into a two-voice dialogue podcast, spoken by custom-trained, recognisable brand voices and backed by multi-stage AI quality assurance that checks factual accuracy for an expert audience.

Both run autonomously round the clock, not despite several built-in checking and correction stages, but because of them, stages the system earned over hundreds of iterations.

Stack: n8n, Anthropic Claude (Opus/Sonnet/Haiku), OpenAI GPT-4o and GPT-5, Google Gemini, Perplexity, Pinecone, OpenAI embeddings, Imagga, ElevenLabs, Auphonic, Airtable, Cloudinary.

Part 1: text, SEO and illustration

Starting point

Long-tail SEO in a niche needs many well-researched articles; manual editing is expensive, slow and doesn't scale. The same quality across six languages multiplies the effort. Illustration, finding the right images, cropping them, naming them for SEO, inserting them, is extremely time-consuming manual work. And clean internal linking plus FAQ sections, now also relevant for AI answer systems, is error-prone grunt work.

The pipeline at a glance

Built modularly on n8n, orchestrated through a central Airtable database that acts as the production's memory. Filter views serve as work queues, every article runs through specialised stages, and a central index synchroniser keeps status consistent.

Stage 1: generation, longform instead of building blocks

The article takes shape through a multi-stage chain (Claude Opus, partly with extended thinking): title, then an outline with main chapters and sub-points, LSI keywords, recurring terms and sources, then section-by-section writing with context memory (each part knows what's already been written and avoids repeating it), then tagging from a curated catalogue and SEO metadata. A hard-wired blocklist of phrases bans typical AI clichés, so the result doesn't read like a machine wrote it. Target length: several thousand words per article.

Persona engine. At the start of the pipeline sits a persona engine that feeds the machine topics and finished articles: automatic topic discovery from a shop category, complete with scoring and prioritisation, then an AI classifier picks the right writer persona depending on the topic, each with its own tone, its own style and its own reference examples. Not one style for everything, but several curated author voices, chosen automatically by topic. That lifts the copy noticeably above generic AI mass-production.

Stage 2: enrichment, from text to asset

Specialised steps turn the raw text into a complete SEO asset: the right shop category gets assigned, multilingual tags get set, internal text links go to matching categories, and an FAQ section gets generated (20 questions through a three-stage AI chain: SEO analysis, question design, answers). The FAQ answers are deliberately phrased so that AI answer systems like Google AI Overviews or ChatGPT learn to treat the shop as a competent source in its niche. SEO for the world after classic search.

Stage 3: the trick, image vector database and automatic illustration

The goal: articles illustrate themselves, with genuinely fitting images from our own library, in the right spot.

• Image preparation. Smart cropping to consistent formats via a smart-crop API (Imagga), keeping the subject centred; SEO naming via vision AI (GPT-4o looks at the image and generates a filename from specialist keywords, verifiable facts only); a logged import into the media library.
• Classification and vectorisation. Every image gets understood multiple ways: keywords, a detailed vision description, dominant colours, and stored as a semantic vector in Pinecone, in separate namespaces for product and mood images.
• Automatic illustration, in two stages. A planner model (extended thinking) breaks the article into meaningful sections, decides where an image makes sense, and writes a precise image description for each spot; the number of images scales with the text length. Each description becomes a search vector, matched against the image database with a minimum similarity threshold, a preference bonus and deduplication. A second model picks the best, globally duplicate-free image for each spot. Then insertion, publishing, status feedback.

No "grab any image with the right keyword" logic, but genuine semantic image-to-text matching with a division of roles (one model plans, another decides), quality thresholds and deduplication. The result: articles that look editorially illustrated, without an editor.

Stage 4: multilingual output

A dedicated translation workflow carries each article across, section by section and preserving HTML, into up to six languages, with metadata and FAQ handled separately. A German article becomes an international set with no extra effort.

Part 2: from article to podcast

Same data foundation, next stage: every magazine article can become a podcast episode fully automatically, as a two-voice dialogue that sounds like a hosted conversation. Neither voice is a generic off-the-shelf TTS voice; both are custom-trained, recognisable voices built for the brand (ElevenLabs). That makes the podcast sound like it comes from the house itself. The real engineering achievement, beyond that, is the content quality assurance for an expert audience.

Stage 1: script generation

Claude Opus turns the article into a conversation structure with segments, time targets, speaker allocation and an emotional arc, including a built-in myth- and fact-check system that actively spots the niche's typical clichés. Each segment is then scripted individually, with live fact-checking via Perplexity and memory to keep segments coherent with each other.

Stage 2: the quality backbone, a multi-LLM review panel

This is where the real engineering work sits. Before anything gets voiced, the script runs through a multi-stage checking and correction chain (roughly 60 processing steps, more than 13 AI agents):

1. Parallel fact-checking by four models at once: Claude Opus (comprehensive fact-check, extended thinking), GPT-5 (scene-aware), Gemini 2.5 Pro (second fact-check) and Claude Haiku as a subject-matter guardrail for correctly distinguishing easily confused technical terms.
2. Consensus building: the findings get merged; each statement gets an agreement counter, and matches flagged by several models are prioritised.
3. Live verification: uncertain facts get checked online (Perplexity) before anything is corrected.
4. Correction and enhancement: a correction agent fixes issues, an enhancement agent increases the content's substance.
5. Narrative consistency: a further cascade of agents checks whether the corrections disrupt the narrative flow, and optimises the audio tags for speech synthesis.

This is LLM-as-a-jury: several independent models check each other instead of relying on just one. That's exactly what makes it defensible that no human has to proofread at the end.

Stage 3: speech synthesis and QA

A two-voice dialogue with our own brand voices (ElevenLabs Text-to-Dialogue) produces natural-sounding conversation instead of interchangeable standard TTS. Per-section audio QA (Gemini) checks the finished audio to confirm two speakers are actually audible, with an automatic retry on failure. A small but effective control loop that catches typical synthesis glitches.

Stage 4: mastering and publishing

Auphonic handles merging the audio sections, loudness normalisation and levelling, automatically inserts a mid-roll spot, the cover art gets AI-upscaled, the finished episode gets published and the status written back. A translation branch also turns the German episode into international versions.

Does this run unsupervised? No, it runs autonomously, because the quality earns it

Both pipelines run unsupervised, round the clock, and that isn't a risk, it's the result of consistent quality assurance. Autonomy wasn't the starting point, it was the earned end state after intensive iteration; the core workflows each carry several hundred version states. What supports it: multi-stage generation instead of a black-box one-shot, built-in correction and checking loops, schema-validated AI output (verified JSON instead of free text), genuine vision verification of images, and idempotent control via filter views.

Only once these loops had delivered stable green results over many runs did the manual approval step become unnecessary. That's the real achievement: not "an AI makes content", but a system good enough to be left to run on its own.

Roadmap: the self-optimising SEO layer (in progress)

So far, the pipeline ends with "published". The layer we're currently building closes the loop: articles get monitored continuously after publication and sharpened up in a targeted way. Real search data from Google Search Console per article, detection of quick wins (keywords just short of page one), AI-assisted sharpening of exactly those articles instead of rewriting them from scratch, internal linking through the same vector database that already feeds the illustration, and re-indexing for freshness. Under evaluation: contextual embeddings (Voyage AI), where every text section retains the context of the whole article. That turns a "publish and forget" machine into a content library that keeps improving itself.

Bonus: content that sells too

Where the Part 1 pipeline enriches every article with thematically matching images, a second, completely independent mechanism goes a step further. It uses the shop system's native infrastructure (Shopware) and inserts real, matching products from the catalogue directly into the article at runtime, with no manual placement per post.

The module reads the article's related products as stored in the shop, automatically picks the right product image via the shop's image convention and the correctly sized thumbnail through Shopware's own media service, and distributes the product blocks at spots that make sense content-wise: spread out ahead of subheadings, not just dumped at the end of the article. Each product also generates Schema.org JSON-LD with price, availability, brand and GTIN, so machine-readable product data sits right in the post. Quality thresholds and a per-product cache keep this clean and fast in continuous operation.

That closes the gap between editorial content and product catalogue, and gives the content a second job. What the pipeline produces for visibility and reach becomes a conversion channel at the same time: in the middle of the post, at a thematically fitting moment, the reader encounters real, purchasable products instead of only at the very end of the page. Every article carries traffic, and now also drives it towards a purchase, automatically, across the entire, growing content library. Content becomes conversion, built not as a parallel system but as a lean layer on top of what the shop can already do.

Outcome and impact

• Scale: a consistent magazine with several thousand articles, growing continuously, with no team growing to match.
• Multilingual with no extra effort: every article and podcast in up to six languages.
• Consistent SEO hygiene: linking, tags, FAQ, image-optimised filenames, across the board.
• Editorial look without an editor thanks to our own image vector database.
• Credible audio content thanks to multi-LLM fact-checking, important for an expert audience.
• Continuous operation: both pipelines work through their backlog autonomously.

What this means for you

This architecture isn't a one-off, it's a transferable pattern: a modular, AI-powered content engine, adaptable to your range, tone, languages and channels, from text generation through semantic illustration to multilingual publishing and voicing. The machine takes on the effort; control over quality, tone and brand image stays with you.

It's the same engineering discipline we use to digitise highly regulated logistics: build something completely new out of data that already exists, and secure it so it runs without supervision. One way of working, two worlds.

The building blocks of this machine are also available individually: RAG content from real product data and workflow automation on self-hosted n8n.