In 2004, Larry Page told Google's prospective shareholders that a well-functioning society needed abundant, free and unbiased access to high-quality information. For twenty years, Google more or less delivered on that promise by acting as a gateway: you asked a question, it handed you a list of doors, and you chose which one to walk through.
That model is dead. Nobody at Google will say so in those words, but the operational data says it for them.
Google's public position, restated as recently as last August by search VP Liz Reid, is that the company sends out a "relatively stable" number of clicks and that reports of declining referral traffic are inaccurate. That is the theatre.
The pulldown is Cloudflare's. In its July 2026 report marking a year since it started restricting AI crawlers by default, Cloudflare found that agent traffic has now crossed 50% of all internet traffic, a first in the medium's history. More tellingly: of every sixty minutes a person now spends looking for information online, only fifteen are spent on the open web. The rest happens inside a chat window, one prompt in, one synthesised answer out, no doors, no walking through them.
The categories hit hardest, mostly news and media at first, have since spread to retail, software, IT and finance, with some losing up to 40% of human traffic in under a year. Wikipedia is down 8% on human visitors over the same period, even as bot scraping for AI training climbs. The Wikimedia Foundation has started charging AI firms for training access and pushing its own app and social channels, because it can no longer assume Google will keep sending readers its way.
None of this makes Google uniquely villainous. It makes Google uniquely positioned: Cloudflare's own figures put Google at roughly 88% of referral traffic, and unlike most other AI firms, Google runs a single mixed-use crawler that blends search indexing and AI training into one undifferentiated stream. Publishers cannot see which purpose a given crawl serves, and cannot grant access to one without granting it to the other. That is not an oversight. It is a structural advantage, and it is the reason the UK's Competition and Markets Authority had to legislate transparency into existence rather than wait for Google to volunteer it.
For anyone tempted to read this as a knockout blow: the CMA's own notes to editors are explicit that SMS designation "does not imply that it has acted anti-competitively." Google has nine months to comply, with six-month reporting cycles, but the core commercial incentive remains untouched: keep the user inside the walled garden. The requirement addresses the symptom Google was caught concealing. It does not address the disease.
Start with what synthesis actually does, because the word is doing a lot of quiet work in Google's favour. A traditional search result is a list; you see ten different sources disagreeing with each other, or corroborating each other, and the disagreement itself is information. A synthesised answer collapses that list into one voice. It picks a position, smooths over the sources that contradicted it, and hands you a paragraph with no seams showing. You lose the disagreement, which means you lose the signal that would have told you the question was contested in the first place.
That collapsing is not a side effect; it is the product. Ten minutes comparing sources becomes ten seconds reading a conclusion. Those ten seconds feel efficient precisely because the friction that drives healthy scepticism, weighing one outlet against another, catching a lone unverified claim, has been erased for you, invisibly, by a system with no obligation to show its reasoning and a direct interest in keeping you satisfied with the answer it already gave you. Confidence, in a synthesised answer, is a design choice, not a measure of how settled the underlying question actually was.
Layer the sourcing opacity on top of that and the problem compounds. An AI-mediated answer is not a neutral summary. It is a single point of curation, produced by an organisation with a direct commercial interest in you not leaving its property, running on a crawler that will not tell you whether the content beneath the answer was indexed for discovery or harvested for training. You are trusting one intermediary's judgement about what counted as a reliable source, with no visibility into the selection, no view of what got left out, and, until nine months from now at the earliest, without even a reliable link back to check its working.
That is a governance problem before it is a technology problem. Verification requires being able to see the sources, see where they disagreed, and choose for yourself which one to trust. An answer you cannot decompose back into its parts is not information; it is a conclusion wearing information's clothes.
This is not an anti-AI argument. I use AI daily, for drafting, for research, for exactly the kind of synthesis Google is selling. The difference is that when I use it, I know which model produced the answer, I can see what it retrieved and when, and nothing about the arrangement depends on an intermediary's undisclosed incentive to keep me from clicking away.
Self-hosted metasearch does the same job Google used to do honestly: it queries multiple engines, returns links, and lets you decide where to go. It does not synthesise an answer on your behalf, which is precisely the point. You get the doors, not someone else's account of what is supposedly behind them. Lantern, the SearXNG instance behind this publication's own research, exists for exactly this reason: not because self-hosting is fashionable, but because an auditable search layer is a precondition for auditable conclusions.
If your publication, your business or your own thinking depends on being able to say where a fact came from, the tool you use to find that fact needs to answer to you, not to a shareholder letter written twenty years ago and quietly abandoned since.
Run your own search engine. Or at least know exactly whose you are trusting, and why.
Cross-reference: We Built a Search Engine This Afternoon
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy, with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
