AI search has a clean story it tells about itself: you type a question, you get an answer, you’re done in seconds. No clicking through five tabs, no ad-stuffed listicles, no SEO slop. On purely functional terms, that story holds up. Perplexity, Google’s AI Overviews, and ChatGPT Search have all gotten noticeably better at synthesising accurate information quickly. The problem is what that efficiency is quietly doing to the infrastructure underneath it.
The web has always run on a bargain: publishers produce content, search engines send traffic, ad revenue covers the costs. It’s an imperfect system that spawned a decade of keyword-stuffed garbage, but it also funded legitimate journalism, documentation, independent research, and niche expertise that doesn’t exist anywhere else. AI search engines are drawing from that well without meaningfully replenishing it. When an AI Overview answers your question about medication interactions or tax rules, the physician’s article or the accountant’s explainer that trained or sourced the answer gets no visit, no engagement, no revenue signal.
Google has made noise about “AI Mode” driving incremental clicks to publishers, but the framing is telling - incremental, as in extra, as in distinct from the baseline traffic that used to just exist. The baseline is eroding.
The Citation Problem Is Real, But It’s the Wrong Argument
A lot of the pushback against AI search focuses on attribution - whether sources are cited properly, whether citations are accurate, whether the links actually go somewhere useful. These are legitimate complaints. Perplexity in particular has faced criticism for producing responses that lift text closely from original sources while burying the link behind a numbered footnote most users don’t click.
But getting the citation format right doesn’t fix the economics. A footnote nobody clicks is aesthetically fairer than no footnote, but it does the same financial work: none.
Publishers Are Running Out of Moves
The options available to publishers are narrowing fast. Blocking AI crawlers is technically possible via robots.txt, but enforcement is inconsistent and unverifiable - crawlers have been caught ignoring those directives. Paywalling content keeps it out of AI training pipelines to some extent, but also cuts organic discovery. Licensing deals exist, but only for publishers large enough to negotiate them.
Small outlets, independent writers, and domain-specific experts - the people producing the kind of specific, reliable information that AI search depends on - don’t have seats at that table.
What’s not clear yet is whether the internet can sustain the production of original knowledge once the traffic incentives for producing it are gone. AI search is optimised to extract value from a system it has no structural reason to maintain. That’s not a flaw in the product design. It’s just the logic of the model, running forward.