{"slug": "the-answers-are-free-because-somebody-already-paid-for-them", "title": "The Answers Are Free Because Somebody Already Paid for Them", "summary": "A randomized browser-extension trial found that Google AI Overviews cut outbound clicks to websites by 39.8 percent on queries where they appear, with zero-click searches rising from 54 to 73 percent, while the share of tracked US keywords showing an Overview grew from 5.93 percent in September 2024 to 61.43 percent in August 2026. The analysis also documents collateral damage to independent publishers, citing a Berlin living guide that lost roughly 75 percent of its traffic and now faces closure, leaving stale pages that AI systems may still cite. The piece argues generative engine optimization offers little recourse, noting llms.txt showed no measurable benefit and AI assistants sent sites about 0.3 percent of traffic.", "body_md": "Every model you use was trained on a corpus somebody wrote for free, and this week three signals showed that the people doing that writing are being paid less, publishing less, and in one case erased on a ten-day notice.\n\n## 1. A Randomized Trial Finally Priced the Click Economy. It Came Back at Minus Forty Percent.\n\nThe strongest evidence yet is an actual experiment, not another correlation study. Researchers shipped a browser extension that randomly assigned US desktop users into two groups — one saw Google normally, the other had AI Overviews hidden — then [watched two weeks of their real searches](https://revscope.co/blog/search-seo-geo-2026/) in early 2026. On queries that would show an Overview, the group that saw one averaged 0.37 clicks out to websites; the other 0.62. A 39.8 percent reduction. Searches ending with no outbound click went from 54 percent to 73 percent. Ad clicks didn't move, which tells you which part of the page Google was careful not to disturb.\n\nScale matters. The share of tracked US keywords showing an Overview went from 5.93 percent in September 2024 to 61.43 percent in August 2026, rising across every intent class — informational to 70 percent, even transactional to 29. Over the same window the featured snippet fell from 8.96 percent of result pages to 0.44, and the forums block from 21 percent to 7. There is no safe keyword tier to retreat into: low-volume, low-difficulty terms show Overviews slightly *more* often than competitive ones.\n\nThe industry's answer is GEO, generative engine optimization, and the honest reading is that it is SEO with two footnotes. About 38 percent of Overview citations come from the top ten organic results anyway, a study of nearly 300,000 domains found no measurable benefit from publishing an llms.txt file, and AI assistants sent websites roughly 0.3 percent of their traffic. You cannot optimize your way back to a click the interface no longer needs to give you.\n\n**Why it matters:**\n\n- **For ICs:** If your product's growth model assumes organic search, that assumption carries a measured 18.5 percent haircut on total Google clicks, and it is getting worse. Check your referrer logs before someone asks in planning.\n- **For leaders:** Fund first-party distribution — docs, email, in-product surfaces — as infrastructure, not marketing. The channel you rented is being repossessed.\n- **For founders:** Content-led acquisition is a declining asset class. Price it that way instead of assuming last year's funnel.\n- A randomized trial beats every vendor dashboard. Insist on one before accepting any claim about what AI did to your traffic.\n\n## 2. The People Who Wrote the Answers Are Quietly Closing Up Shop\n\nThe macro number has a human shape. One well-regarded independent guide to living in Berlin — running since 2017, its author's full-time living since 2020, cited by federal and municipal offices, the basis for two government portals — [has lost about 75 percent of its traffic](https://nicolasbouliane.com/blog/death-by-ai) to AI Overviews. Its author now spends his days preparing an exit instead of writing guides. The immigration office uses his data about the immigration office, and he can no longer afford to maintain it.\n\nThe failure mode matters more than the sob story, because sites like this rarely vanish. They get paywalled, or stay online and stop being updated — producing a corpus that looks exactly the same size as before and quietly stops being true, which is the worst possible outcome for anything trained on it. Your assistant will cite a 2025 procedure with total confidence because the page is still there, still indexed, and nobody is paid to correct it. The supply of accurate, specific, maintained public knowledge turned down this year, and models see that on a training lag.\n\nGoogle has signaled the end state: AI answers replacing results entirely, with ads inside them. In that world there is no referral economy left to split, and whatever the next model generation knows about configuring a router or upgrading a framework comes from a snapshot of the last era in which somebody could make rent writing it down.\n\n**Why it matters:**\n\n- **For ICs:** Treat confident answers about procedures, pricing and APIs as stale by default. Date-check the primary source; the model has no idea when its page stopped being maintained.\n- **For leaders:** Your internal corpus is now a differentiated asset. Runbooks, incident writeups and architecture decisions are the part of your context no vendor's crawl will ever replace.\n- **For founders:** There is a real business in maintained, verifiable, machine-readable reference data. Freshness is the moat, because freshness is precisely what just got defunded.\n\n## 3. Apple Is Deleting Two Decades of Its Own Support Archive Tomorrow\n\nOn November 10 Apple announced that its Support Community — where users have troubleshot Apple products for roughly twenty years — would go read-only immediately and close permanently on November 20. A volunteer archiving crew is [running a distributed grab against the clock](https://wiki.archiveteam.org/index.php/Apple_Discussions) to save it, so the record of what actually went wrong with real hardware survives somewhere. Apple's suggested replacement is its own support site: first-party, curated, and conspicuously free of the thread where someone explains the fix that shipped documentation says is unnecessary.\n\nIt is the same squeeze hitting independent publishers. As the numbers go up in AI, [they go down for independent educators, publishers and community conferences](https://css-tricks.com/when-everything-has-changed/) — the institutions that produced the tutorials every working engineer learned from. Add vendor-controlled docs replacing third-party Q&A and you get a knowledge base with no failure history in it, which is the half that makes debugging possible.\n\nSo mirror what you depend on while mirroring is still an option, and pay the independents your team reads — a course budget is cheap next to rebuilding that knowledge in-house.\n\n**Why it matters:**\n\n- **For ICs:** Keep local copies of the docs for your critical dependencies. A link is not an archive, and read-only is one quarter away from gone.\n- **For leaders:** Audit which external knowledge sources your team relies on and which have a business model. The ones that don't are a dependency with no SLA.\n- **For founders:** If your product's answers come from a public corpus, you have supply-chain risk you haven't modeled. Own or license the data you cannot do without.\n\n## The Verdict: Real or Hype?\n\n**AI answers collapsing referral traffic → Real.** A randomized trial with a 40 percent effect and a 54-to-73 percent jump in zero-click searches is not a vibe. **GEO as the replacement for SEO → Hype.** Citations track organic rank, llms.txt does nothing, and the whole channel is a third of a percent of traffic. **The shrinking public corpus → Real but early.** Answer quality hasn't visibly degraded yet, but supply turned down this year and training lags reality by a year or more.", "url": "https://wpnews.pro/news/the-answers-are-free-because-somebody-already-paid-for-them", "canonical_source": "https://fromtheterminal.substack.com/p/the-answers-are-free-because-somebody", "published_at": "2026-11-20 17:27:00+00:00", "updated_at": "2026-11-20 18:15:04.243802+00:00", "lang": "en", "topics": ["ai-search", "generative-engine-optimization", "ai-crawlers", "artificial-intelligence"], "entities": ["Google", "AI Overviews", "Nicolas Bouliane"], "also_reported_by": [], "takeaway": "Hiding AI Overviews in a randomized two-week trial raised outbound clicks from 0.37 to 0.62 per affected query, a 39.8 percent drop attributable to the feature, with zero-click searches rising from 54 to 73 percent.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/the-answers-are-free-because-somebody-already-paid-for-them", "markdown": "https://wpnews.pro/news/the-answers-are-free-because-somebody-already-paid-for-them.md", "text": "https://wpnews.pro/news/the-answers-are-free-because-somebody-already-paid-for-them.txt", "jsonld": "https://wpnews.pro/news/the-answers-are-free-because-somebody-already-paid-for-them.jsonld"}}