Search traffic losses hit publishers as AI answers reshape the results page

Search engines are not broken. They are being asked to do a different job. That is the argument in a Manticore Search post published on 29 September. The post describes a shopper typing a full sentence into a search field and expecting a recommendation, not a page of blue links. According to the company, Google reported in May 2026 that AI Mode had surpassed one billion monthly users. People are asking longer, more complex questions that did not fit conventional search before.

The shift matters commercially because the click is the unit publishers sell. If the engine resolves the query itself, the page that used to receive the visit does not get one.

Manticore's own write-up is careful here. It does not claim that full-text search has decayed. It says keyword search still performs well on exact names, SKUs, product codes, brands and keywords. What changed, in its account, is user expectation. The same query can contain several tasks at once: extract constraints such as colour and waterproofing, recognise that the shopper means running rather than walking, weigh price, size and availability, then find products and explain the differences. Manticore's answer is hybrid retrieval: full-text plus vector search plus a language model that writes the final answer using retrieved products as context, with the products still visible so a reader can check the source of a claim. The company built a demo on ConvApparel, a conversation dataset about apparel that after clean-up contained 82,524 products.

That framing is a useful counterweight to the gloomier reading of the same trend.

For publishers, the gloomier reading has numbers behind it. Google search traffic declines have been reported at roughly 40% year on year for publishers as of late September 2026, and smaller sites have been described as seeing drops of up to 60%. An antitrust filing has argued that Google cannibalises publisher traffic. None of this is settled or universal. The figures come from different methodologies and different sample sets, so they should not be averaged into a single industry number.

The pattern is not confined to one company. Microsoft has been adding an ask a follow-up feature to Copilot results, which could further reduce referral traffic. Cloudflare's Matthew Prince has argued publicly about bots and paying for the web, a debate that runs alongside the question of human referrals. Publishers themselves have been modelling a future with significantly less Google search traffic. Digiday's September 2026 Publishing Summit recap describes an industry rebuilding for a post-search era.

What the numbers actually say

UK audience data offers a partial counterpoint. Press Gazette reported on 28 September that the Oxford Mail led audience growth among the 50 biggest UK news websites and tripled time spent in August. That is one title, in one market, in one month, and it does not disprove a broad referral decline. But it does show that the aggregate hides a wide spread.

The mechanism behind the spread is visible in the search stack itself. Manticore describes three distinct jobs that used to be collapsed into one ranked list: word search for exact matching, semantic search for meaning, and conversation for explanation. A system that combines them can answer a product question without a click. A system that only matches words still sends the user somewhere.

It is far more effective to combine them than to choose between them.

That is Manticore's conclusion about search architecture, not a prediction about publishing economics, and the two should not be confused.

There is a second, quieter pressure on the referral economy: the quality of the ranking itself. A software engineer's post republished on 28 September describes a pairwise evaluation method in which a language model judges which of two products is more relevant to a query, then judges again with the two products swapped. Forcing a decision on the WANDS furniture dataset produced precision of 75.08% at 100% recall across 1,000 pairs. Allowing the model to answer neither when evidence was insufficient raised precision to 85.38% but cut recall to 17.10%. The author's point is that a single pass is not enough; position bias has to be checked by reversing the order.

That is an evaluation detail, but it points at something structural. Search systems are being tuned to decide relevance with less human labelling, and the same models are being used to answer the query outright.

The infrastructure underneath

The rest of the stack is moving at the same time. MongoDB used its 29 September announcement to launch MongoDB 9.0, Atlas Infinite and Atlas Agent Engine. An interim CEO note argues that agents now take actions, not just answer questions, and that stale data turns a sound inference into a bad action. The pitch is that agents need live operational state rather than periodic copies.

Whether that changes publisher referrals is indirect, but the direction is consistent: software is being rebuilt around systems that act on retrieved context rather than hand a user a list of destinations.

On the security side, the VUSEC group at Vrije Universiteit Amsterdam published research on 29 September describing Branch Target Reuse, a Spectre-v2 style attack against just-in-time compilers. The team analysed Linux cBPF, Oracle GraalVM and SpiderMonkey, the JIT engine in Firefox, and built two end-to-end exploits against the Linux kernel, leaking 8 bytes per second and, in their demo, recovering a root password hash. The attack relies on stale branch prediction entries surviving code reallocation. The Linux kernel can harden cBPF at runtime with the bpf_jit_harden option, which applies constant blinding; the researchers note the option is off by default.

None of this addresses the publisher question directly. It does show how much of the plumbing beneath the browser is being reworked at once.

Two research papers published on 29 September sit further out but frame the same transition. An NBER working paper by Matthew Schwartz, Isaiah Andrews and Jesse M. Shapiro describes an LLM workflow that reproduces, improves and extends published economics research using replication packages. Across 4,452 replication packages from five economics journals, the workflow flagged discrepancies in 3,460 articles or their appendices; in 496 articles it cut computation time by more than a factor of 10 at similar or greater accuracy; in 923 it produced an extension not present in the original. A separate NBER paper by Alex Blumenfeld, Jonathon Hazell, Chen Lian and Andreas Schaab estimates that from November 2022 to December 2025, AI raised the market's expected present value of software engineering productivity by the equivalent of a permanent 32.6% increase, with a GDP effect of 3.6% in the baseline and 6.5% when R&D productivity is included. By mid-2026, the authors write, that effect had more than doubled relative to the end of 2025.

Those are macro estimates built from financial market data, not measurements of publisher revenue, and they carry the usual caveats attached to working papers. They are also a reminder that the productivity story and the traffic story are two sides of the same deployment.

What to watch

The most concrete signal in the dossier is not a forecast. It is the one-billion-user figure for AI Mode, reported by Google in May 2026 and repeated by Manticore on 29 September, sitting next to reported year-on-year publisher search declines of around 40%. If both hold, the referral economy is being repriced rather than destroyed, and the titles that survive will be the ones that stop treating search as a distribution channel and start treating it as a competitor.

On the same day, VDE published figures on international electrical engineering students in Germany. That is a different subject, but it uses the same logic: measure what a system produces, not what it promises. According to VDE calculations, every euro invested in an international electrical engineering student generates roughly 16 euros in added value over the long term, yet more than half of international master's graduates, 55.5%, leave Germany after finishing their studies.

Publishers have the harder measurement problem. They can count clicks. They cannot easily count the readers who got their answer somewhere else.