Is the article dead?

No.

(Thanks, Betteridge’s Law!)

But the real question isn’t whether the article survives. It’s whether it remains the primary way people encounter journalism.

The written article — linear, static, and designed for a broad audience — was optimized for a world of scarce publishing space and shared attention. That world is disappearing. Referral traffic is collapsing. Homepages matter less than they once did. Increasingly, people encounter journalism through feeds, summaries, notifications, and recommendations shaped to their individual context.

That doesn’t mean the article is obsolete. But it does mean we shouldn’t assume it will stay the same forever.

Although the article has been with us for thousands of years, publishing is continually evolving. Just as the invention of the web changed who could publish and find an audience (from a small group of people privileged by an even smaller group of publishers to, well, everyone with an internet connection), the software and services available today can enable new features like personalization and new kinds of interactivity.

This isn’t technological determinism. If these features succeed, it won’t be because of the technology itself, but because that technology served an existing need that was previously harder to meet. If new tools help us to create a new iteration of the article that resonates with readers, it won’t be because they themselves bring success; it will be because they helped us create something to serve readers in a way that was previously unreachable.

Personalization is one clear example: it’s a feature newsrooms are beginning to talk more about, particularly in the context of LLMs and how they might make new capabilities possible. As Gina Chua, executive editor at Semafor, framed the need for it in this year’s NiemanLabs Predictions for Journalism:

A person reading about a tax proposal in Washington wants to know how it’ll affect their small business in Ohio, and less about political infighting in D.C. or what it’ll mean for the midterms; a young parent in Florida is more interested in the impact of a school board vote on their child’s education than on debates over diversity.

This framing deserves scrutiny — not least the claim that a young parent in Florida isn’t or shouldn’t be interested in debates over diversity — but it’s also important to highlight that Gina thinks this isn’t appropriate for every kind of article:

[…] To be sure, there will always be a need for high-end, high-quality investigations, analysis, and commentary that isn’t personalized; whose value is in fact in the universality of their narratives and ideas. But most news isn’t that. It’s the day-to-day coverage of school board meetings, fires and floods, city hall votes, and business announcements.

This suggests that personalization matters most for what we might think of as “information as infrastructure”. These are articles that provide actionable answers to practical questions: What did the planning commission decide about that development proposal? When does early voting start? What routes are closed for road work? With that caveat in mind, I find it easier to accept that personalization is desirable: I’m interested in whether my city council member voted to collude with ICE, or when the dangerous intersection on the way to daycare is finally getting a stop sign, or which candidates my precinct will see on the ballot, in a way that isn’t necessarily one size fits all.

If we accept that this is desirable, we need to ask: who is personalizing the article, and how?

It’s worth coming back to Gina’s generalization about a young parent in Florida. We may hold that opinion about that broad category of person, but it doesn’t necessarily match reality; everyone is different. Real personalization means going beyond abstracted personas and ensuring that the information any given individual is receiving is tailored to their needs.

That, in turn, means we need to know what their needs are. We have some choices here, and not all of them are compatible with the values of a service-oriented newsroom.

• We can infer their needs through third-party tracking, borrowing techniques from targeted advertising networks
• We can infer their needs through their reading activity on our sites, building an understanding through accumulated reading patterns
• We can let them consensually tell us what they care about

Each of these approaches carries very different implications for privacy, agency, and trust. 

The first option violates fundamental privacy principles and should be ruled out for public service journalism. Determining a person’s needs and interests via third-party trackers outsources the engine of personalization to opaque surveillance networks that violate reader privacy and likely have their own business priorities.

Inferring needs through reader activity has clearer ethical boundaries, but practical challenges remain. For one thing, there may not be enough signal to build an accurate profile. For another, there’s a catch-22 cold-start problem: personalization requires reading activity, but reading activity may never materialize if the content isn’t compelling because it hasn’t been personalized yet. Finally, we’re still profiling the user without their express consent, and those profiles could be revealed via subpoena, putting readers at risk.

The third option aligns with journalistic values but creates friction. Letting them tell us what they care about, consensually and overtly, is ideologically clean but in practice may be awkward and interruptive and risks user abandonment.

Both viable approaches share a fundamental challenge: they require newsrooms to build, secure, and maintain a detailed, persistent profile for each reader as a part of the publishing stack. That’s not something that most CMS platforms offer today, and is a lot to ask of smaller newsrooms in particular. It’s expensive and technically demanding, and the technology will continue to evolve: LLMs as we know them today are not AI’s final form, and new changes will require ongoing development resources. And even beyond the technical considerations, there are legal requirements that may create significant overhead.

For these reasons, I don’t think newsrooms should attempt to build or host personalization engines.

But that doesn’t mean personalization won’t happen and isn’t valuable. The problems I’ve identified — cost, technical change, privacy implications — shift considerably if personalization moves from publisher-controlled systems into reader-controlled tools like their browsers and RSS readers. In this model, the article remains universal, but the experience of encountering it becomes personal.

If personalization happens in the browser, a few important things become possible:

• A profile can be built out of a person’s reading habits across all the sites they read, with the data stored in a place under the user’s control
• They can consensually adjust their profile in their browser’s settings and through UI affordances like “more like this” and “less like this”
• The entire profile never needs to be shared with a third party
• They can choose not to be profiled at all

We can also solve an implied problem with publisher agency. The way we often talk about personalization today, for example in the context of AI agents and chatbots that rewrite or summarize content, requires that publishers lose control over the presentation of their work, and that their relationship with their readers is intermediated. But if we built it as an open standard — with transparent algorithms, user control, and no central data collection — the publisher remains fully in control.

Publishers could expose lightweight semantic signals, using HTML tags, that describe what different sections of an article are about: for example, whether a passage concerns local governance, public safety, elections, or education. These signals wouldn’t change the article itself; they would simply make its structure and subject matter legible to readers’ tools.

A browser or reader could then use those signals, combined with locally held preferences, to decide how prominently to surface, summarize, or emphasize different sections. Some browsers might use LLMs to interpret semantic signals and generate summaries — Chrome seems likely to take this path, given their AI investments — while privacy-focused browsers like Firefox might use simpler, transparent rules. If a browser doesn’t support personalization or a reader switches it off, all the content would be displayed. Those differences would become part of how readers choose between browsers, and the standard itself would be a progressive enhancement rather than a replacement for how we structure articles today.

There’s a long history of developers trying to build semantic signals into HTML. To be clear, I’m not arguing for what we used to call the semantic web. Instead, just as RSS succeeded as a standard because you could get something up and running in an afternoon, this new standard needs to be incredibly lightweight and easy to build into both a CMS and a browser.

One embodiment might be to use existing schema.org vocabulary to describe content semantically — what places, topics, and government actions are mentioned — without making assumptions about reader demographics. In this version, a section about Oakland’s city council vote on ICE collaboration would be tagged with areaServed="Oakland, CA" and about="immigration enforcement policy", not with assumptions about which type of person cares about it. The browser, informed by the reader’s own expressed interests and location, decides what's relevant.

You might disagree that schema.org is the right markup, and that’s fine. Those are implementation details. I think the most important facets of personalization markup are:

• It’s very simple (a developer can implement it in an afternoon)
• It’s progressive enhancement (a browser can render an article even if it doesn’t support personalization)
• It happens in-browser, on-site (the publisher retains control over the overall publishing context)

Of course, AI agents will continue to summarize and rewrite journalism regardless of the existence of any standards. But for readers who want unmediated access to journalism while still getting personalization, and for publishers who want to maintain direct relationships with those readers, this approach offers a path forward that neither requires publishers to build complex systems nor surrenders control to platforms. It also makes blocking AI agents a clearer choice, because the same personalization can take place in less extractive, publisher-friendly environments.

Browsers can also use this as a way to compete with vendors like OpenAI. The web is already compelling, and this proposal brings more of the personalization that AI enables into that more familiar environment. It’s a way for Mozilla and even Google to use their pre-existing web-first stance as an advantage. And because it’s all web-based, it’s easy to create a shim using a browser extension even before browsers themselves choose to adopt it.

So, no. The article isn’t dead. But how it lives on is in question.

Personalization is coming whether newsrooms build it or not. What's in question is where it lives, and who controls it. Built around surveillance and intermediaries, it erodes trust and weakens journalism’s public role. Built into reader-controlled tools on the open web, it can do the opposite: make journalism more relevant without making it more extractive.

At its core, the question isn’t really about AI, browsers, or markup. It’s about agency. Who decides what’s relevant? Who holds the data about what readers care about? And who controls the relationship between journalists and the people they serve?

Those choices will shape not just the future of the article, but the future of journalism itself — and with it, how all of us learn about the world.