Google pays roughly 100 publishers for AI answers, and most get almost nothing

That is the awkward core of a pilot programme that launched less than a year ago and still has no published rate card. The Decoder summarised the reporting on 30 September. Publishers can see in Google Search Console how often their content is pulled into an AI answer and how much they have earned for it. What they cannot see is the formula behind the number.

The spread is enormous. One publisher received $50,000 to $60,000 over a few months. Another earns more than $1 million a year. Several small and midsize blogs and websites have collected less than $1,000 across several months, and for them the payments work out at less than 0.1 percent of ad revenue.

Some participants told The Information they do not know how Google calculates the amounts, and that the payments can move from month to month without explanation. The reporting also suggests that niche topics with strong followings, anime and gaming among them, tend to earn more. Google's answer to the whole arrangement so far is a set of selective licensing deals, an opt-out feature and a payment model in which the company decides what content is worth.

Traffic is falling while the deals stay individual

The money question matters because the underlying traffic trend is bad and getting worse for the open web. Multiple studies cited in the coverage show AI Overviews sharply reduce visits to source sites, and some larger publishers are refusing to join the pilot in an attempt to force Google to pay more.

That refusal is a bet, and the structure of the programme makes it a hard one. Deals are negotiated publisher by publisher, so nobody is bargaining collectively. A site that walks away puts limited pressure on Google because other sources fill the gap in an AI answer. A site that stays accepts the rate it is offered. The Decoder described this as a prisoner's dilemma for publishers, and noted that the divide-and-conquer approach has served Google well for decades as it drew more web content into search to monetise.

There is regulatory pressure in the background. Independent publishers filed a complaint with the European Commission over AI Overviews in July 2025. Rolling Stone parent Penske Media sued Google that September over lost traffic and ad revenue. The Commission then opened an antitrust investigation in December 2025, examining whether Google imposes unfair terms by using publishers' content for AI features without adequate payment or a genuine way to opt out.

A German court has also ruled that AI Overviews are Google's own content rather than summaries of existing material. If that interpretation gains wider acceptance, publishers would have a legal basis to demand licensing fees whenever their work appears in an AI answer. The practical difficulty is obvious: tracking each source's exact contribution to an answer is hard, possibly impossible, and administering per-source payments would add overhead on top of fees that could cut into Google's margins.

The rest of the automation stack is moving faster than the money

While the licensing fight grinds through regulators and courts, the tooling around AI-assisted work is being rebuilt at a much faster clip, and much of it lands directly on engineers.

On 30 September, OpenAI and Synopsys said they had signed a multi-year partnership to build a specialised chip design model called GPT-Synopsys, according to The Decoder. The model combines OpenAI's AI technology with Synopsys' electronic design automation tools, which OpenAI is licensing. The stated goal is a system that can reason about chip design and verification and operate Synopsys' tools directly, with engineers delegating objectives and reviewing output. It runs on OpenAI's infrastructure; both companies say customer data will not be used for training and will be stored encrypted, and that early tests with semiconductor customers are already underway. Synopsys CEO Sassine Ghazi says AI could significantly speed up the design process.

The investment side is moving too. Flow Engineering, a three-year-old San Francisco startup selling AI tools for hardware design, raised a $50 million Series B at a $750 million valuation, TechCrunch reported on 30 September. The round was co-led by Antonio Gracias of Valor Equity Partners and Gavin Baker of Atreides Management, with Sequoia Capital participating alongside former Sequoia partner Roelof Botha as an individual investor. Botha has joined the board. Flow names Anduril, Rivian, Joby Aviation and others as customers.

Infrastructure claims are getting louder as well. Modal published benchmarks on 30 September for Quail, a query-aware inference layer it built for AI-SQL workloads, claiming more than a billion tokens processed per minute per H100 GPU on one multi-join query, which it says is over 10x faster than its vLLM baseline on the same hardware and works out at under 6 cents per billion tokens on Modal. On Modal's new AI-SQL benchmark, Quail runs 1.84x faster than vLLM geometrically averaged over tasks.

Smaller pieces of the same shift are shipping quietly. Facebook's OpenZL project released v0.3.0 on 29 September with a new LZ engine and a Compression Transformer, a neural selector that builds numeric compression graphs on the fly. The release notes claim decompression 144 percent faster than Zstandard at equivalent settings, and the Transformer compresses on average 35 percent better than zstd -19 across 868 families of numeric streams. Magnitude, a YC S25 startup, launched an open source inference engine on GitHub that compiles and tunes kernels on the user's own device; its README claims up to 2x faster than llama.cpp, with 92 percent faster decode on Metal and 19 percent on CUDA, and 27 percent less memory per agent.

Security research is also piling up. On 30 September, VUSEC published Branch Target Reuse, a Spectre-v2 attack class targeting just-in-time compilers in browsers, language runtimes and the Linux kernel. The team built two end-to-end exploits against the Linux kernel and demonstrated leaking the root password hash from a running su process at 8 bytes per second, with a full walk of the kernel task list and page tables.

What the people doing the work are saying

Opinion is splitting on what all this automation does to the job. Writing on martinelli.ch on 30 September, one consultant argued that AI coding agents push software engineers back toward the 1990s model where one person owns requirements, architecture, code and production, and that the specification becomes the most important artefact in a project. A separate post on pooyam.dev the same day is blunter: the author, who works on LLM training and inference performance on GPUs and TPUs, says the cost of implementation is heading toward zero and describes an agent that retrained part of a model to improve speculative decoding acceptance length, something he had not thought to forbid.

A third account, on rtfm.co.ua, describes the same drift from the other side: the agent becomes Mycroft to your Holmes, and the human ends up as what the author calls meat middleware between the agent and the deploy button. Gerben Rijpkema, an AI engineer, argues on Substack that the successful projects are the ones with use-case-specific architectures, borrowed from control engineering's assumption that you cannot fully predict the system and have to design around the noise.

None of that resolves the publisher question. Google is still setting the terms, 100 publishers are still inside the pilot, and the European Commission is still investigating.