Search is turning into a conversation, and publishers' traffic answers are changing

Search did not break. The people asking questions now expect the machine to answer them, not to hand them ten blue links. That shift shows up in product data, in academic work and in the traffic figures publishers have been staring at all year.

According to Search Engine Journal, a court filing dated 28 September quotes an OpenAI engineer saying users "won't click" links. That is one line in a legal document, and it is a claim by one company's employee, not an industry average. It also landed in a week when the supporting evidence got louder.

What the numbers say

Google reported in May 2026 that AI Mode had passed one billion monthly users, according to the search vendor Manticore, which cited Google's own blog in a post published on 20 August. Manticore's point is narrower than it sounds: full-text search still works, it just does not get a shopper to an answer. Keyword search, vector search and a language model each solve a different part of a query, the company argues, and combining them beats picking one.

For publishers, the more uncomfortable number sits on the other side of the click. Search Engine Land reported on 24 September that Google search referrals to publishers had fallen about 40% year on year, one of several datapoints in a decline that had already been running for months. The Next Web reported on 28 September that Cloudflare's Matthew Prince argued bots do not click on ads, which frames the same problem from the infrastructure side. Neither outlet's figure is a census. Treat both as directional.

The pattern is not confined to news. Alex Blumenfeld, Jonathon Hazell, Chen Lian and Andreas Schaab published NBER Working Paper 35793 in September. From financial markets they estimate that AI had raised the market's expected present value of software engineering productivity by the equivalent of a permanent 32.6% increase between November 2022 and December 2025, with a GDP effect of 3.6% in the baseline. By mid-2026, they write, that effect had more than doubled. The mechanism they describe is not search, but it is the same underlying move: tools that used to return material now return conclusions.

MongoDB's interim CEO used the company's 29 September platform announcement to make the operational version of the argument. In a blog post coinciding with MongoDB 9.0, Atlas Infinite and the public preview of Atlas Agent Engine, he wrote that agents need access to a business's live state, because yesterday's inventory or an old account balance can turn a sound inference into a bad action. The post does not mention publishers. It does explain why the retrieval layer is being rebuilt around freshness and action rather than indexing.

The research layer is not standing still either

Two NBER papers published in September point in opposite directions for anyone hoping the answer is simply "trust the model." Matthew Schwartz, Isaiah Andrews and Jesse M. Shapiro, in Working Paper 35782, ran an open-source LLM workflow across 4,452 published replication packages from five economics journals. It flagged discrepancies in 3,460 articles or their appendices. It also cut computation time by more than a factor of ten in 496 articles and produced an extension in 923. The authors disclose that Schwartz worked as a contractor for Anthropic during the project.

The counterweight is methodological. Doug Turnbull's write-up of an LLM-as-judge experiment on the WANDS furniture dataset found that letting the model answer "neither" pushed precision from 75.08% to 85.38%, while recall collapsed from 100% to 17.10% on 1,000 pairs. Swapping the two products and re-asking improved consistency. By current standards the experiment is small and old. It is still the cleanest public demonstration that a judge which always answers is not the same as a judge that answers well.

Elsewhere the week's releases look like plumbing. Loro's Zixuan Chen argued on 21 September that CRDTs are not enough on their own, because a conversation index can converge while the message body has not, which leaves a user staring at a notification with nothing behind it. OpenJev, a GitHub project published on 29 September, reproduces the interface pattern of TypeSafe's Jev with open-weight models and reports 290 ms for one question and 328 ms for 27 in its own harness. Oxilite, also on 29 September, ports an Oxigraph-compatible SPARQL engine onto SQLite so it can run on Cloudflare D1.

None of that resolves the publishers' problem. It does show where the engineering effort is going: into systems that return a decision, a probability or a compiled answer, rather than a page of results for a human to sift. If a court filing in September is right that users will not click, the question for anyone whose business model was built on the click is what replaces it.