Remember the days when searching online meant scrolling through a neat list of blue links and picking one? Nostalgic, isn't it? That model—type something, get a page of options, choose—was remarkably durable for thirty years. It wasn't elegant. It required effort. But it was at least legible: you could see what you were being shown, and you retained the final decision about which source to visit.
That model is collapsing. The rise of AI powered search—ChatGPT, Gemini, Claude, Perplexity—is changing how people discover information. Instead of browsing a feed curated to maximise your engagement, you ask a model a direct question, and it returns a synthesised answer. The form is different. But the underlying dynamics of power and mediation remain. And in some ways, they've intensified.
We replaced one gatekeeper with another. Platform algorithms kept you scrolling. AI models synthesise the answers before you can even ask the follow up question.
The Scale of the Shift
Google itself has rolled out an AI powered search mode that gives you conversational answers rather than a page of links. Early users have been typing in longer, more complex questions and getting synthesised answers—often without ever clicking through to a website. One SEO expert described it as "the biggest disruption in the history of search." I'd take that quote seriously. The digital ad world has a long history of understatement.
In the UK, search ads still make up nearly half of all digital ad spend—but experts predict that by 2028, about one-third of search advertising will happen on AI powered interfaces. For publishers and content creators who built their businesses around being discoverable in search results—from travel bloggers to national newspapers—this shift is an earthquake. When I talk to editorial teams at major publishers, the anxiety is palpable. Their traffic from search has been declining for two years and nobody has a confident answer for what replaces it.
From Feeds to Models: What Changes?
Platform-mediated discovery was built around engagement. The algorithm showed you what kept you clicking, even when that meant amplifying outrage, misinformation, or content specifically designed to provoke reaction. Model mediated discovery claims to be built around utility. The model answers your question rather than exploiting your attention.
This sounds like progress—and in some ways it is. When I need to understand a complex topic quickly, a well calibrated AI model is genuinely more useful than scrolling through ten search results and reading three of them incompletely. The friction reduction is real. The time saving is real.
But the shift from feeds to models doesn't eliminate the problem of mediated discovery. It restructures it. The question is no longer "what does the algorithm want you to click?" but "what does the model believe to be true?"—and those are two very different questions with two very different failure modes. Engagement optimised content is manipulative. Model curated content can be simply wrong—and confidently, fluently wrong in ways that are harder to detect than a misleading headline.
The Black Box Problem
When platforms mediate discovery, you can at least observe the feed. You know what you're being shown, even if you don't know exactly why. With models, the reasoning is hidden. You ask a question, the model consults its training data—a compressed, frozen snapshot of the internet from some point in the past—and returns an answer that feels authoritative, polished, and complete.
An AI answer doesn't feel like a curated search result. It feels like a conclusion. And conclusions are much harder to argue with than search results.
Models can be wrong. They hallucinate—a technical term for producing confident-sounding fabrications. They have embedded biases that reflect the data they were trained on. And unlike a feed that shows its sources, a model's answer often arrives without a paper trail. The user experience is excellent. The epistemics are alarming.
There's a real world case that illustrates the stakes. In 2023, a US lawyer submitted a legal brief citing six precedents his AI assistant had found. None of the cases existed. The AI had invented them—complete with plausible sounding names, dates, and holdings. The lawyer trusted the output because it arrived in the form of confident, structured information. That's the risk of model-mediated discovery applied to consequential decisions: the format communicates authority that the content doesn't always warrant.
What Happens to the Open Web?
The more important question is structural. Independent publishers are already feeling the impact. Vice Media, once valued at $5.7 billion, filed for bankruptcy in 2023 after steep declines in traffic and advertising revenue—attributed in part to changing discovery habits and over reliance on platforms. BuzzFeed shuttered its entire news division, citing the collapse of referral traffic from search and social platforms. Even legacy outlets like The Guardian have publicly challenged AI firms over content scraping, warning that these shifts threaten both editorial independence and financial sustainability.
When people get their information from models instead of visiting websites, the incentive to create content changes fundamentally. Publishers don't receive traffic. Creators don't get read. The value created by the open web gets extracted and repackaged by the companies that train models on it. It's a classic tragedy of the commons: individually rational behaviour at the model level (summarise everything) produces a collectively irrational outcome (destroy the ecosystem that generates the content worth summarising).
Model mediated discovery is a different bottleneck from the platform algorithm—but it concentrates power in even fewer hands. At least the platform ecosystem, for all its problems, was plural. There were many algorithms. There are very few frontier AI models. The narrowing of the discovery funnel is not incidental to this shift. It is its central logic. And as a society, we haven't yet reckoned seriously with what that concentration means.
