- The Daily Mail's desktop click-through fell from 25.23% to 2.79% the moment Google put an AI Overview above its link — an 89% collapse. The content still gets read. The publisher just stops getting paid for it. The referral economy that funded journalism is over. 🧵
- It's not one outlet. Google now shows an AI Overview on 48% of searches; when it does, the link gets clicked in just ~8% of visits (Pew). Publisher search traffic fell ~33% globally in the year to Nov 2025. 69% of searches now end in zero clicks. A structural cliff, not a slump.
- The obvious answer — "build our own AI" — has scar tissue. The Washington Post shipped Ask The Post AI, then an AI podcast that invented quotes and ran its own commentary AS the paper's position. The standards editor's leaked Slack (Semafor, Dec 2025) is the risk in one message.
- The technical failure was one missing step: citation verification. Every answer must ground each claim in a real, linked source article — or refuse to answer. Ask FT does this: it runs on Claude, cites FT journalism only, mandatory. No citation, no answer. That's the bar.
- The other hard part nobody sells: entity resolution. In a 30-year archive, "Mr. Musk", "Elon Musk" and "the Tesla CEO" are one person but three nodes. GraphRAG has to collapse them, or multi-hop questions return garbage. This is the unsexy 60% of the build.
- News questions are also temporal. "Who chaired the council when the stadium deal passed?" needs time-stamped edges, not vector similarity. Temporal RAG with valid_start/valid_end metadata is the gap between a real archive engine and a chat widget over embeddings.
- A SaaS vendor drops a widget on your site for $60K–$120K: no entity resolution, no temporal reasoning, your archive in their cloud. A big SI builds it properly for $1.5M–$5M and a discovery phase longer than your runway. Mid-tier publishers fit neither option.
- Defense alone loses. Run a parallel licensing play: Cloudflare Pay Per Crawl (Jan 2026) now default-blocks AI crawlers across ~20% of web traffic — the first infrastructure-level bot tax. Plus the News/Media Alliance + ProRata pool: 2,200 publishers, 50/50 on attributed answers.
- The honest take: run BOTH — your own cited chatbot for retention AND the licensing pools for leakage. Ask FT proved the chatbot lifts retention, it doesn't cannibalize subscriptions. Most consultancies pick one. Both are correct. Stop letting Google rent your archive for free.
- If you run a publisher AI engine: are you enforcing citations at generation time, or bolting fact-checks on after the model speaks? After WaPo, that gap is a board-level liability — not an engineering preference. Which side is your stack on? #AISearch #RAG
- We wrote up the full publisher AI playbook — citation enforcement, GraphRAG entity resolution, temporal reasoning, and the dual licensing strategy — here: https://veriprajna.com/solutions/conversational-ai-for-publishers