Editorial Guidelines

Most articles on TechFeedWatch are drafted by an automated system, as the About page explains. That makes the rules more important, not less: a person can use judgement, a system only does what it is told. This page lists the rules it is told to follow. Each one is checked in code before an article goes live, not just written down here.

Two kinds of content

Articles answer one question people search for. The question is researched with Google Search, the article is built on the web pages Google cites, and one video from a creator or outlet who covers the subject adds a section of its own. They are drafted by a language model and published automatically. Articles written before October 2026 were drafted from the video’s transcript alone; they are being rewritten with the current method, a few a day. Images marked “AI-generated” are made by an AI model.

Guides, the glossary and the free tools are written with AI, and planned, checked and maintained by me, Jacob S. Olsen. Unlike the articles, they are not published automatically.

How a question is chosen

Questions come from Google’s search suggestions for the subjects the site covers (AI, coding, cybersecurity, fintech, crypto, SEO, quantum computing, AR/VR and a few more). A model then keeps only clear, plain-English questions that a factual article can answer.
No duplicates. If the site already has an article with the same search intent (for example “what is an AI chipset” and “what are AI chipsets”), no new article is written. Near matches are judged a second time before anything is published.
It has to be answerable. The question is researched with Google Search before any video is looked at. If fewer than four facts backed by at least two different web pages come back, no article is written. Sub-questions that no fact covers are dropped, and if fewer than three are left, the question is skipped.

Which sources count

Facts come from the pages Google cites, and every page is judged before the article is written. Strong sources stay: governments and regulators, standards bodies, universities and peer-reviewed research, established news organisations, research firms and companies publishing their own data, and official product documentation. Weak sources are removed: statistics roundups and listicles, pages that mainly repeat figures someone else measured, SEO and affiliate sites, and personal blogs.
Some sites never count as a source: forums and Q&A sites such as Reddit and Quora, social media, YouTube itself, and document-sharing sites.

How a video is chosen

It has to cover the question. The video’s transcript is checked against the question before any article is drafted. If the video only touches the subject, it is skipped.
It has to have substance. Short videos, videos without English subtitles and videos with too few concrete points are skipped.
No sales videos. Three separate checks score whether a video is mainly selling a product, service or course. If the middle score is 7 out of 10 or higher, the video is skipped, however relevant it is.
A video that has been removed from the site is never picked again.

What gets an article rejected

A draft that breaks one of these rules gets one revision. If it still breaks it, the article is not published.

Numbers must have a source. A figure that appears neither in the researched facts nor in the video is not allowed, and more than one sentence with a figure but no named source stops the article.
It must be built on the research. At least two of the researched web pages must be named in the text.
The video stays a supplement. Exactly one section draws on the video, with the channel named in its heading, and the video may supply at most 30% of the article.
It must not copy the video. If more than 15% of the sentences share eight words in a row with the video, the article is not published.
It must make sense to a reader. A second model reads the draft as someone who has not seen the video; if three or more sentences still trip that reader after revision, the article is not published.
It must be concrete. A draft with too few specific facts is rejected.
It must read as an article, not a transcript. Spoken leftovers such as “I”, “let’s take a look” or “hey” are not allowed, and neither are diagrams drawn with text characters.
No keyword stuffing. The search question may not be pasted into a sentence where it does not belong.

What every article is checked for

These are checked on every draft and sent back for revision when they fail:

Name the source. Claims are attributed to someone specific: the creator of the video, a named company or a named study, not “experts say” or “studies show”.
Easy to scan. At least two lists per article.

And these are fixed automatically, without changing a word:

Answer first. An opening paragraph longer than 60 words or three sentences is split, so the answer stands on its own.
Short paragraphs. Paragraphs longer than three sentences are split between sentences.

Every article credits the creator by name and channel and embeds the video, so you can watch the original. The checks themselves are tested automatically before every run, so a broken check stops the system instead of letting articles through.

Links

Internal links point only to articles that exist on this site and match the subject.
Sources are listed at the end of every article written with the current method: the web pages the text actually uses, plus the video. The list is built by code from what the article cites; the language model never writes it.
Links in the text to outside sites are added by the system, at most five per article and one per website. Up to three come from the video’s own description, so they are sources the creator chose. Up to two point to the official site of something the article names, and only after the address has been checked: the domain must contain the name, the page must load, and its title must contain the name.
Shortened links, affiliate links, discount codes and sponsored links in the video description are never carried over.

How much is published

At most three new articles a day: two or three on weekdays, one to three at weekends, and none on a day when nothing passes the checks. On top of that, a few older articles a day are rewritten with the current method. Publishing less is always preferred to publishing something that fails.

Money and independence

No creator, company or advertiser pays to be covered, and none is told about an article in advance. The site may show advertising; advertisers have no say in what is written. If an article turns out to promote a product rather than explain a subject, it is corrected or removed.

Reviews of the archive

I review published articles in batches and fix the system when a pattern shows up. In October 2026, for example, 56 unattributed claims in 37 older articles were given named sources or rewritten, advertising sentences were removed from four articles, and one article that mainly promoted a product was taken down.

Corrections and removals

Mistakes are handled as described in the corrections policy. If you made a video covered here and want the article changed or removed, email [email protected] and it is done.

What this site is not

It is not a news agency. It does not do original reporting or interviews, and its articles do not replace the source video. Where the site publishes its own numbers, such as on Trends, the method and its limits are on the same page.

These guidelines are maintained by Jacob S. Olsen, Denmark. Last reviewed October 2026.