What the Newspaper Industry's Death Teaches the Web About Surviving AI

In 1995, Craigslist went live. By 2000, it was operating in nine cities. By 2005, it had become the largest classified advertising site in the world. And between roughly 2000 and 2010, classified advertising — which had accounted for approximately 40% of newspaper revenue — essentially ceased to exist as a newspaper business line.

The newspaper industry’s reaction to Craigslist is one of the most instructive case studies in business history. The newspapers could see what was happening. They discussed it extensively in trade publications. Several major newspaper companies tried to build their own classified ad websites. None of them succeeded in any meaningful way. By 2008, when the financial crisis hit on top of the structural advertising collapse, hundreds of newspapers had closed or reduced to shells of themselves.

The Newspaper Association of America’s data tells the story in numbers. US newspaper advertising revenue was $49.4 billion in 2000. By 2012, it was $22.3 billion. By 2020, it was $8.8 billion. This is not a gradual decline. It’s a cliff. And the cliff was built over fifteen years of watching the disruption happen and responding primarily by asking the disruptors for money and occasionally suing them.

The question worth asking is not why the newspapers failed — that part is obvious. The question is why they failed to mount an effective response despite seeing the threat clearly for five years. The answer illuminates exactly what web publishers are getting wrong about AI right now.

The newspapers thought they were in the newspaper business. They weren’t.

Newspapers were bundles. The bundle contained: local news reporting, national and international wire service content, opinion columns, lifestyle features, sports scores, movie listings, classified advertising, display advertising from local businesses, and a physical delivery infrastructure that reached homes five or seven days per week. The bundle was sold to readers for a few dollars per week and to advertisers for rates based on the reader count. The bundle worked as long as all its components were only available together — when you wanted classified ads, you had to buy the newspaper that contained the classified ads.

The internet unbundled the components. Craig Newmark realized that what people actually wanted, specifically, was a way to sell their used furniture and find an apartment. They did not want the furniture listing to come packaged with a sports section and a local news report. Craigslist gave them the piece they wanted, free, without the bundle. Google gave them the news they wanted via search, without the bundle. The bundle fell apart because the components could now be obtained separately.

What newspapers were actually selling was: local information intermediation and trust, plus classified advertising infrastructure. Once someone else provided better classified advertising infrastructure (free, searchable, city-specific), and once someone else provided better information discovery (Google), the newspaper’s core value proposition collapsed. No amount of better newspaper-making could have fixed this. The problem was that the thing they were defending — the bundle — was not actually the valuable asset.

Web publishers are repeating this mistake with AI. They think they’re in the content business. They’re not.

The parallel is nearly exact. Web publishers produce articles, videos, and posts. They bundle these with advertising, distribute them through social media and search, and monetize based on pageview volume. The value of a pageview is based on the scarcity of the content — if people want this piece of information, they have to come to this website to get it. AI destroys that scarcity. If you can ask a language model “what are the best hiking trails near Denver” and get a synthesized answer that’s better than any individual hiking website’s article, you stop visiting hiking websites.

The panicked response from web publishers has been to demand payment from AI companies for training data, sue over copyright violations (with mixed legal results), and add paywalls. All of this is the newspaper industry building classified ad websites in 2002. It responds to the surface mechanism of disruption — “they’re taking our content” — without engaging with the deeper problem, which is: what are web publishers actually selling that AI cannot replicate?

The answer exists. It’s just not “content.”

What AI cannot replicate, at least not yet in any meaningful commercial sense, is trusted human curation with skin in the game.

The New York Times’s restaurant reviews carry weight because readers believe the critic has genuinely eaten at the restaurants, has a developed palate, has no financial relationship with the establishments being reviewed, and will lose professional reputation for false or slipshod assessments. These properties cannot be replicated by a language model trained on existing reviews. An AI can produce plausible restaurant criticism, but it hasn’t eaten the food, doesn’t have a reputation at stake, and can’t be sued for defamation or lose a job for producing consistently wrong assessments.

Consumer Reports still has value for the same reason. Wirecutter has value (or had value before its NYT acquisition complicated its independence) for the same reason. The Dispatch and The Atlantic retain subscribers not because they produce content — AI can produce content — but because readers trust specific writers with specific track records to tell them what to think about complicated topics where judgment matters more than information. That trust is built over time, is associated with specific people, and is not transferable to a system that produces anonymous-seeming output.

The web publishers who are winning right now, in a nascent way, are the ones who have understood this. Substack’s success stories are not generic content producers; they’re specific writers with specific audiences who trust their specific judgment. The Pragmatic Engineer sells $250/year subscriptions to tens of thousands of engineers because Gergely Orosz has fifteen years of software engineering experience at Uber and Skype and a track record of being right about things that matter to engineers. You cannot synthesize that with training data.

Local news is a particularly sharp example of both the failure mode and the potential recovery. Local newspapers died largely because their advertising revenue collapsed — the classified ads went to Craigslist, the display ads went to Google. But local news was never primarily an advertising vehicle; it was an accountability mechanism. The Pulitzer Prize-winning investigative work done by local papers on police misconduct, corrupt city councils, and public health failures — none of that was replaceable by a classified ad website.

The local news organizations that have survived and occasionally thrived in the digital era have mostly done it by going back to the original value proposition: accountability journalism that no algorithm will produce because no algorithm has a legal reporter sitting in a county courthouse reading property records. The Texas Tribune, run as a nonprofit since 2009, covers Texas state government in the kind of depth that requires reporters who know the building, know the legislators, and will go read the actual bills. That’s not content. That’s institutional knowledge and accountability.

The web publishers who will survive AI are the ones who ask: what are we actually providing that isn’t “information that could be synthesized”? The ones who are trying to produce more content faster with AI assistance — to maintain pageview volume in an environment where pageviews are worth less — are fighting the last war. They’re building a better classified ad website in 2003.

The paywall question is a useful diagnostic here. Publishers who added paywalls and saw subscriptions grow — The Times, The Economist, The Athletic — are publishers whose readers were already trusting the brand, not the algorithm. The paywall forced a clarifying question: would users pay for this specifically? The ones who would were already experiencing the product as a trusted relationship rather than a commodity information source. The ones who wouldn’t were always just Google traffic arbitrage, and no paywall was going to convert them.

The uncomfortable version of this analysis: most web content is not worth preserving. Most of it is thin information packaging designed to rank in search results and accumulate advertising impressions. That content was already a kind of extractive activity — capturing value created by searchers’ attention without producing proportionate value in return. AI will destroy that business model, and honestly, the destruction is not obviously bad for the internet as a whole.

A large-scale SEO content operation in 2022 was producing articles at scale — often two hundred or three hundred per week — covering the full range of “what is the best [product category]” and “how to [common task]” queries. The economics required high volume because the revenue per pageview was measured in fractions of cents. When AI-generated answers started appearing at the top of search results pages in 2023, these operations saw traffic drop 40% or 50% within months. They responded by using AI to produce more content. They are now producing AI-generated content to compete with AI-generated content to capture traffic that doesn’t exist anymore. This is not a viable strategy.

What the internet will lose, if publishers don’t adapt, is the accountability and curation function — the specific human judgments that carry credibility because they’re attached to identifiable people with identifiable track records who will suffer consequences for being consistently wrong. That loss would be genuinely harmful. The loss of low-quality SEO content would not.

The newspapers that survived the internet transition — The New York Times, The Washington Post (for a while, at least), The Guardian, The Financial Times — did it by accepting that they were not the bundle and were not the advertising vehicle. They were the trusted brand with named reporters who had reputations to maintain. The ones that tried to compete on volume lost. The ones that competed on trust sometimes won. The Times’s digital subscription business crossed 10 million subscribers in 2022. That’s not a pageview business. That’s a trust business.

The web’s version of the Times model is visible and growing. Heather Cox Richardson’s “Letters from an American” Substack had over 2 million subscribers at its peak. Not because she was the fastest producer of US political history content — she wasn’t. Because her readers trusted her specific reading of events, her specific voice, her specific interpretive framework built over decades of academic work. That’s the asset. AI cannot replicate it because the asset is not the writing — it’s the person behind the writing and the track record that person has built.

The lesson is available. It’s thirty years old. The only question is whether web publishers will look at the newspaper industry’s failure and see themselves, or spend the next decade building classified ad websites.