Publish Article

Publish Article

Writing the article is the easy half. This skill covers the half that decides whether the published page earns traffic, citations, or link value: what it is titled for, who it is attributed to, what image it carries, what it links out to, and whether the thing you intended actually rendered.

Every rule below has a check you can run before publishing. A rule without a check gets skipped, so run the checks.

Modes

• Content mode (default) — publish an original article. No backlink involved.
• Backlink mode — the same, plus one planned keyword link to a target page.
• Supplied-copy mode — the user hands you finished text to publish. See "Supplied copy is immutable" below; it overrides the writing guidance, never the verification.

Gate 1 — measure the title keyword BEFORE writing

This is the decision that most determines whether the page is worth publishing, and it is the one usually made by instinct while drafting. Print this table before the first word:

Volume is measured, never assumed. A phrase invented from the article's own finding ("the 1.3% narration gap", "the delivery problem") has zero volume by construction — nobody searches a coinage they have never read. Clever titles are the default failure mode here, not an occasional slip.

The winning phrase must belong to the topic family you are actually competing in. This is the subtler failure and it survives the volume check cleanly. Linking to a slide-deck product, ai training video generator (250/mo) beats ai training slides (0/mo) on volume and is still the wrong answer: it ranks the article for a video need while the target sells slides, weakening the topical signal at exactly the moment the choice looks optimised. Take the best-supported phrase from the right family over a bigger number in a neighbouring one.

Keyword tools badly understate AI and long-tail volume — do not read a 0 as "no demand". A tool can report 0 for a phrase your own Search Console shows taking clicks at position 1. Where search-console data and a keyword tool disagree, the console wins; the tool is only there to rank candidates against each other.

Fixing this after publishing is cheap on any CMS that keeps a redirect from the old slug — do it rather than living with it. The window closes once the URL is indexed or cited.

Gate 2 — plan the anchor and the topic (backlink mode, before writing)

A backlink's value is carried by its anchor text and by whether the surrounding article is about the same thing as the target page. Both are decided up front and stated, never improvised while drafting:

Read the anchors already pointing at that target page first. New anchors are chosen against the existing distribution, not in isolation.

• Every anchor contains the target keyword or an obvious variant. A branded or generic anchor (here, this tool, the bare brand name) passes almost nothing for the keyword being fought for. Spend one only after the keyword-bearing anchors are in place.
• Vary the surface form — exact, plural, branded+exact, partial. One identical anchor repeated across every placement is the clearest possible footprint of bought links.
• The article's topic must match the target page's topic. A link to an educational-video page from an e-bike article is worth close to nothing however good the anchor is.
• One link per article to the target. A second adds no value and looks manufactured.
• Apply Gate 1's two rules to the anchor as well. AI training presentation maker reads like a keyword and is one nobody types; AI training presentation is the phrase with real demand.

Original research (original copy only)

A press release, company blog post, or wire story is a lead, not an article. Add something a reader cannot get by opening the source:

• Read the originating source plus at least two independent primary sources contributing different facts. Reprints and syndicated copies count as one source.
• Produce at least one original finding by calculating, comparing, ranking, or tracing change over time. State the denominator, period, and method beside the result so it can be reproduced.
• For a data-bearing story include one table and one chart: the table carries exact values, the chart reveals the pattern. Bar chart for comparable categories, line chart only for a real time series. Never stretch two data points into a decorative chart or invent numbers to satisfy this rule. Where a topic genuinely has no comparable data, ship a structured comparison table and say why a chart would mislead.
• Charts are supplemental editorial media, embedded in the body with descriptive alt text and a visible source/method line — never the cover image.

Before publishing, print: sources used, the original finding, table row count, chart type, word count. If the draft is mainly a rephrased announcement, research further instead of publishing.

Supplied copy is immutable by default

When the user provides an article, draft, title, or passage and asks to publish it, that text is approved copy, not a brief. Never rewrite it without explicit permission.

• Preserve the supplied title, thesis, framing, examples, conclusions, tone, ordering, and level of certainty. Do not replace the user's argument with a safer or more conventional one.
• An editing instruction authorises only the direction requested. "Make it more affirmative" means strengthen the supplied thesis; it does not authorise changing that thesis.
• Factual verification still applies. If verification finds a problem whose correction would materially change a supplied claim, show the exact issue and the proposed change and get permission before publishing. Never silently substitute a different claim.
• Without rewrite permission, limit changes to publication mechanics that preserve the words: Markdown formatting, escaping, platform syntax, placement of requested images.
• Before publishing, diff the final copy against the supplied copy. Every substantive change must map to an explicit instruction in the current request. If it does not, stop and ask.

The quality bar — every published article MUST have

1. Title

• ~60 characters max so it fits a SERP <title>. No multi-clause headline with an em-dash subtitle; trim to the single core hook.
• Put the searched entity in it — proper nouns beat clever framing. People search names, products, and companies, not concepts. Demis Hassabis steps down, Jeff Dean leaves Google outranks Who runs Google AI now? because both names and the company are literal query terms.
• The test is "would someone type this", not "does it contain a proper noun". The wrong proper noun is as dead as none. Write out the 2-3 searches a reader would actually run, then check the title contains those words. Two ways this fails in practice:
  • Naming your source instead of the subject. The Information Drew Dario Amodei as a Prophet has two proper nouns and misses the query, which is Anthropic IPO. Who reported it belongs in the byline, never the title.
  • Spending the character budget on tone. A wry kicker that costs you the word statue, cast, review, price, or specs costs you the highest-volume word available. Cut the kicker, keep the noun.
• Exactly one H1. The title is the page H1, so the body must not open with its own # heading or repeat the title (demote body headings to H2+).

2. Author

A real author is attributed — never the CMS default, "Staff", or a null fallback. The author record has a name and a bio, avatar preferred. Create one if needed rather than shipping the fallback. Never fabricate credentials or expertise.

3. Images

• Source order: public-domain photo first, AI-generated second. Search Wikimedia Commons for the actual subject before generating anything, and take the file only if its licence is Public domain or CC0 — never CC BY, CC BY-SA, or anything else carrying someone's terms, even though attribution would technically satisfy them. Use the Commons API (list=search&srnamespace=6, then prop=imageinfo&iiprop=extmetadata, filter on LicenseShortName) and send a real User-Agent or it 403s. Credit a public-domain cover in an italic line at the foot of the article.
• Prefer a real photo when the subject is a real, specific object or place. An AI render of a thing that exists invents its appearance, which is a fabrication with a picture around it. When no free photo of the subject exists, a public-domain photo of the setting beats an invented photo of the subject.
• Landscape, and no text baked in — roughly 1200x630 for OG. The title lives in the page, not the picture. Crop or regenerate if the ratio is wrong.
• Realistic, not "AI-looking". Editorial photographic look; avoid the glossy over-rendered aesthetic and metaphor-soup (one object morphing into another). Reserve conceptual imagery for pure-science topics; business, policy, and product stories get realistic scenes.
• The OG image URL resolves — curl -I returns 200, absolute https://, not a placeholder.
• Distinct per article. In a batch, assert distinct(images) == count(articles). Omitting the field and letting the platform generate one beats shipping a duplicate.

4. Outbound source links — at least two, in the body

• Naming a source in prose does not count. Link it. A reader checking the claim has to get there in one click, and an unlinked source earns nothing.
• Link the primary document where one exists (filing, dataset, release, docket, court order) and the report that carried it otherwise. Descriptive anchor, never "here".
• Assert before publishing: grep -c '](http' <body> is ≥ 2. On one audit of ten freshly published articles, six cited their sources in prose and linked none of them — the crawler correctly reported zero external links.