Operational Document

Methodology

How the weekly digest is collected, ranked, and published

Product: A weekly curated intelligence digest covering AI & Strategy, Ecommerce & Retail Tech, Luxury & Consumer, and Jewellery Industry. Each issue presents seven ranked articles in each of four categories (28 in total), delivered simultaneously to the web, by email, and as a podcast briefing.
Sources: ~53 RSS feeds across six source tiers — global newswires, retail trade press, fashion and luxury publications, jewellery specialist titles, specialist technology feeds, and business commentary. Two static page scrapers supplement the feed list.
A Tavily web-discovery layer runs in parallel each cycle, using LLM-generated search queries per category to surface high-quality articles that fall outside the curated feed list.
Ingestion: Articles are ingested on an automated schedule into an append-only store. Duplicate detection runs on URL and title similarity — each article appears at most once regardless of how many feeds publish it.
Weekly window: Monday 00:00 through Sunday 23:59, Europe/Copenhagen timezone (CET/CEST). Each digest covers only articles with a publication date that falls within that window.
Classification: Articles are assigned to one of four topic categories using a deterministic keyword and source heuristic. Source-level overrides take priority (e.g. a jewellery trade publication is always assigned to Jewellery Industry regardless of article content). Classification uses no LLM inference and is fully auditable.
Ranking: A deterministic pre-score is computed for each article: recency weight (constant baseline), source tier weight (0–0.15), and keyword relevance boost (0.05 per match). Low-signal markers carry a −0.20 penalty.
The top 18 candidates per category (at most 4 from any one source) are then passed to OpenAI’s gpt-6.1-sol (reasoning model), which applies editorial criteria with hard news first — deals, results, leadership changes, strategy and launches ahead of analysis and opinion — to select the final 7 articles per category. A hard diversity guard caps any single source at 3 articles per category.
AI — what it does: +Reranks pre-scored article candidates using editorial criteria, hard news first (gpt-6.1-sol)
+Generates concise web summaries from title, source, date, and snippet only — no full article required (gpt-6.1-sol)
+Generates email digest bullets (3 per article) from up to 8,000 characters of the full fetched article text (gpt-6.1-sol)
+Produces the weekly pull-quote in two stages: one pass writes four candidates from different angles (prediction, contradiction, myth-buster, hidden trend); a second pass judges them and picks the strongest (both gpt-6.1-sol)
+Tags the week’s key themes and translates the summaries into Danish and Spanish (gpt-6-luna)
+Labels brand news on the Markets pages as company moves or not, and writes their weekly summary (gpt-6.1-sol)
+Scripts the podcast briefing (roughly 13–18 min) from the top 8 articles (gpt-6.1-sol) and voices it (gpt-realtime-2.1-mini)
+Generates the weekly cover image: a SceneDirector step writes the scene (gpt-6.1-sol) and an image model draws it (gpt-image-2.5-sunburst)
AI — what it doesn't: —Accessing paywalled or subscription-gated content — articles that cannot be fetched are skipped entirely
—Human curation or manual editorial override of ranked results
—Rewriting or editorialising source material — email bullets reflect the article as published
—Generating content not grounded in a published source article
Cadence: Published weekly. The pipeline runs automatically each week; new issues appear in the archive as soon as they are built. Subscribers receive the email digest on the same schedule.
Transparency: Article selection is AI-augmented, explainable-first. Every ranking signal is deterministic and logged; the LLM reranker sees only pre-scored candidates and applies editorial criteria, not preference.
We are not affiliated with any publisher. All links go to original source articles. Summaries are AI-generated and may miss nuances — always read the full piece for complete information.
Questions or source suggestions? Share your feedback →

Pipeline last run: 19 October 2026 · CET · 43 issues archived