Every newsroom that has tried AI-assisted content tagging has had a version of the same experience.

The suggestions look reasonable in a demo. In production, they drift. An editor adds the obvious corrections, then a few more, then starts ignoring the assistant altogether. The time saved disappears into the gap between what the model thinks is relevant and what the publication actually does.

The gap is not a language problem. AI models are good at understanding what an article is about. The problem is that knowing what an article is about is not the same as knowing how a specific publication classifies it. Those are different kinds of knowledge, and only one of them lives inside the model.

What editorial conventions actually look like

Take a news story about a central bank interest rate decision.

A financial news publication tagging that article will not just add “interest rates” or “central bank.” Depending on the publication’s taxonomy, editors will also apply the specific institution (Federal Reserve, Bank of England, etc.), the relevant economy (U.S. economy, U.K. economy, etc.), the policy area (monetary policy, etc.), and possibly the broader market context (bonds, equities, etc.) if the story covers market reaction.

Those choices determine where the story surfaces across the site: in topic pages, related content modules, newsletter feeds, and search archives.

A semantic AI model can recognise the article is about interest rates. It will likely suggest the central bank by name. But without access to the publication’s taxonomy and tagging history, it will not reliably know editors at this publication consistently apply the economy tag alongside the institution, or that “monetary policy” sits in a different part of the taxonomy hierarchy from “market reaction,” or that both are expected on this article type.

The model sees related language. It does not see the classification rules the newsroom has built up over years of editorial practice.

This matters more than it might appear. A political story about a party leadership contest will, at most publications, carry tags for the party, the constituency or region, the individual politicians named, and the broader political context.

A semantic model may suggest the party and the lead politician. It will often miss the regional tag populating the local news feed or the policy area tag connecting the story to related coverage from six months ago. Those omissions are invisible in a demo and costly in production.

Your archive is the training data you already have

The fix is not a better language model. It is giving the system access to the right evidence.

Every mature newsroom has tagging conventions that were never formally documented. They live in the archive, in the habits of experienced editors, and in the corrections made silently every day.

Across thousands of articles, patterns are legible: This politician is consistently tagged with their party, their constituency, and their ministerial portfolio; this company always carries its sector tag and its parent group; breaking news articles on this beat routinely include a geography tag even when the location is not in the headline.

A multi-signal tagging system can learn those patterns from the archive and use them as evidence when ranking suggestions for new articles. The suggestions still come from the publication’s own taxonomy but they are ordered by how well they fit both the article and the way the newsroom has handled similar stories before.

Taxonomy structure adds a further layer. Hierarchical taxonomies carry relationships that matter editorially: a politician belongs to a party, a company belongs to a sector, a region sits within a country. A system using those relationships when ranking suggestions will handle the connections that semantic similarity alone tends to miss.

The results are measurable

We tested this approach against 1,000 real articles, drawn from an archive of more than 27,000. Each article was sent to the suggestion engine with its headline and body text, and the output was compared against the taxonomy terms editors had actually applied.

A semantic-only baseline found an average of 1.7 correct suggestions per article with 20% recall. The multi-signal system found 6.7, reaching 97%.

Precision was 47%, which means editors still have some suggestions to dismiss. In an assistive tagging workflow, though, the bigger cost is usually the missing tag: The editor has to notice the gap, recall the right term, search for it, and add it manually.

High recall reduces that overhead. The balanced configuration, returning the top 10 suggestions, produced the best combined score across both measures.

Why it matters beyond the tagging interface

Tagging quality has consequences extending well beyond the moment an editor hits publish.

Poor tags affect archive quality, related content modules, site search, topic page search engine optimisation (SEO), and the accuracy of newsletter personalisation. They also create a slow drain on editorial capacity: When first-pass suggestions are unreliable, editors spend time on corrections rather than coverage, consistency suffers across the team, and institutional knowledge stays locked in the heads of the people who have been there longest.

A multi-signal approach makes institutional knowledge legible to the system. The goal is not to replace editorial judgment. It is to give AI enough context to reflect how the newsroom actually works, so the suggestions editors see are grounded in their own practice rather than a model’s best guess from the open Internet.

At Glide Publishing Platform, we built this into our content management system (CMS) tagging workflow after seeing the same pattern repeatedly: publications trialling AI tagging tools, finding them promising in controlled conditions, and watching the results degrade in production. The gap was consistent. The model understood the content, but it did not understand the publication.

The newsrooms that get the most from AI-assisted tagging treat their own archive as the primary source of truth. The model’s job is to surface patterns from that archive, not to substitute for it.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .