The AI search traffic apocalypse is a story most working journalists, marketers, and comms pros know by heart, and I’ve written about it plenty. The premise is simple: If your business runs on funneling readers from search engines to a website, AI summaries are cutting that funnel off at the top, sending the audience instead to a paragraph the bot wrote about your work rather than the work itself.
The compensation question hangs over all of it. Publishers may eventually get paid for the material that AI systems are scraping and citing. They may not. But the more pressing shift is happening in plain sight. The battle for attention is moving toward new victory conditions, namely whose information gets cited most prominently in an AI summary. AI presence doesn’t replace traffic, but it does function as the new proxy for relevance and authority.
In my first Fast Company column, I argued the incentive system this creates is healthier than the one we’ve lived under for over a decade. Search and social trained media to chase engagement, which produced the fire hose of listicles, outrage bait, and formulaic explainers like “What time is the Super Bowl?” If AI systems are the arena, and if they really do reward well-sourced, domain-specific content more than social heat, that could lead to a resurgence of good journalism, at least directionally.
Enough time has passed to start checking the theory against actual data. AI search engines have been operational for more than a year, usage is rising fast, and researchers are starting to publish on what these systems actually cite. The early signal is encouraging, but the caveats are real.
AI’s reading list
A pair of recent studies looked at millions of AI citations, the term for a source that an AI summary both names and links. The data found that AI systems treat LinkedIn as one of the most authoritative sources on the internet. Research from Meltwater, a communications intelligence company, showed LinkedIn as the second most-cited source overall in AI summaries (after YouTube), and a separate study from Semrush, a search-data analytics company, concurs, also putting it at No. 2, closely behind Reddit.
The Meltwater data also point to why LinkedIn is a decent indicator of substance. Individual members (rather than brand or company accounts) drove most of the citations, structured content like newsletters and posts performed best, and more than half of the citations went to members with fewer than 10,000 followers. Likewise, Semrush found that the most-cited LinkedIn posts had only modest engagement on the network itself. That’s strong evidence against a simple popularity model.
The harder finding is what AI systems do when structure is missing. When you drill down into academic papers that zero in on exactly how large language models prioritize information, like this one from the Canadian AI company Cohere, they show that LLMs will miss key facts when an article lacks clear titles and headings. A separate paper from Stanford University goes further, showing AI search systems strongly favor the beginning and the end of documents over the middle. If the meat of your reporting sits in paragraph seven, the bot may never reach it.
Put those findings together and AI systems look as gameable as search and social were, just along different axes. An article optimized for machines, with declarative ledes, clean Q&A, and consistent titling throughout—but otherwise empty of substance—could theoretically outscore a piece with original reporting that lives in the middle. AI systems reward retrievable substance, not necessarily the most insightful or information-dense content.
Visibility to AI engines isn’t enough on its own. You have to lead the bots to the good information instead of hoping they’ll find it. This is the whole idea behind GEO, or generative engine optimization, and it can feel at odds with what makes good writing good. Clever titles, narrative hooks, the slow backing-into of a topic through a scene: humans love that work; machines mostly don’t.
The human edge in machine search
Look back at which sources are ranking highest: LinkedIn, YouTube, Reddit. That mix suggests the best content is a blend of machine-friendly formatting and the human element. AI doesn’t always cite the most engaged posts, but the Semrush data also shows that frequent posting and an established following still help. LinkedIn’s own internal guidance points in the same direction. So engagement still matters, just less directly than it did in the previous era.
Demoting raw engagement is progress. The structural bias is something working journalists and content creators can put to use. It signals that content based on original reporting or insights needs to do several things: explain concepts clearly and quickly, include machine-friendly structures like subheadings, and connect the dots with other sources the AI is reading by referencing them by name.
The play, then, is to make good work easier for machines to read without sanding off what made it valuable to humans in the first place. The next incentive system will have its own bad habits, and there’s no doubt many people will try to exploit them. But if AI search continues to weight original facts, named sources, clear context, and demonstrated expertise over outrage and raw engagement, that’s an opening worth taking seriously. The winners in an AI-mediated future shouldn’t simply be the loudest accounts or the best-formatted posts. They should be the people who know something real and can demonstrate its worth to both audiences they’re now writing for, human and machine.
A version of this column appears in Fast Company.
