Google pays around 100 publishers for AI content, and the rates vary wildly

The pilot programme is less than a year old. Publishers can see how often Google uses their content and what they earn through Google Search Console, The Decoder reported on 30 September, citing The Information. What they cannot see, several participants told The Information, is how the number is calculated.

The range of outcomes is wide. For several small and midsize blogs and websites, payments came in below 0.1% of ad revenue. Some small sites received less than $1,000 over several months. At the other end, one publisher received $50,000 to $60,000 over a few months, while another earns more than $1 million a year. Niche topics with strong followings, such as anime and gaming, appear to earn more, according to the same report. Payments depend on how much each source contributes to an AI answer. Participants also told The Information that the amounts can shift from month to month without explanation.

Not everyone is signing.

The legal pressure around the programme

Some larger publishers are refusing to join in order to push Google to pay more, The Decoder reported, at a time when their search traffic is already falling and multiple studies show AI Overviews sharply reduce visits to the open web. The regulatory track is already open. Independent publishers filed a complaint with the European Commission over AI Overviews in July 2025, and Rolling Stone parent Penske Media sued Google that September over lost traffic and ad revenue. In December 2025 the Commission opened an antitrust investigation into 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 ruled that AI Overviews are Google's own content, not summaries of existing material, according to The Decoder. If that interpretation spreads, publishers would have a legal basis to demand licensing fees whenever their work appears in an AI answer. Tracking each source's exact contribution would be hard, administering the payments would add overhead, and the fees could cut into Google's margins.

For now Google sets the terms through selective licensing deals, an opt-out feature and a payment model that lets it decide what content is worth. The Decoder's Matthias Bastian describes this as a divide-and-conquer approach that creates a prisoner's dilemma: a few publishers benefit, most get little or nothing, and those who walk away put little pressure on Google because other sources fill the gap. The context is not new. Search referrals have been moving against publishers for years: reporting aggregated in the dossier points to a 40% year-on-year decline in Google Search traffic for publishers as of 24 September 2026, and earlier coverage of small sites seeing drops of around 60%. None of that is part of the September pilot, but it is the backdrop against which publishers are being asked to accept the new rates.

What the wider engineering conversation says

The AI-versus-publisher story is not only about money. On 30 September, Martinelli published a piece arguing that AI coding agents are pushing software engineers back toward the generalist role of the 1990s, where one person handled requirements, architecture, code and production. The argument matters for media because the same shift is happening inside newsrooms and platform teams: fewer handovers, more automation, and a specification that becomes the most important artefact in the project. A second post that day, by Gerben Rijpkema, makes a related point from control engineering: successful AI systems are built around a specific problem, while generic out-of-the-box tools tend to fail. "In the projects that were a success, I ended up designing use-case-specific architectures, whereas the failed ones were built on top of out-of-the-box AI tools that promised to solve all my AI engineering challenges with a one-size-fits-all solution," he wrote on 30 September. That is a fair description of the licensing problem too. Google's programme is a single pipeline serving many kinds of publishers, from anime blogs to large newsrooms, and the payouts reflect a calculation nobody outside the company can inspect.

Other 30 September releases show how fast the surrounding tooling is moving.

Modal published benchmarks for Quail, a query-aware inference layer that it says reaches over a billion tokens processed per minute per H100 GPU on one multi-join query, more than 10x its vLLM baseline on the same hardware and under 6 cents per billion tokens on Modal. On its AI-SQL benchmark, Modal reports Quail at 1.84x faster than vLLM geometrically averaged. Elastic, meanwhile, tested TypeSafe AI's Jev model as a search reranker on 250 Amazon Shopping Queries and 4,754 judged query/product pairs, lifting nDCG@10 from 0.9351 to 0.9565 and Exact MRR from 0.9201 to 0.9616. Jev is priced at $0.042 per million input tokens, with output tokens free. Cheaper inference on the platform side does not answer the publisher question. If anything it makes large-scale content reuse more affordable for the companies doing it.

Smaller signals, same week

On 30 September, the VUSEC research group at VU Amsterdam published Branch Target Reuse, a Spectre-v2 attack against just-in-time compilers. The team built two end-to-end exploits against the Linux kernel and leaked 8 bytes per second, enough with pointer chasing to extract the root password hash from a running "su" process. The write-up notes that the Linux kernel's bpf_jit_harden option, which blinds immediate values, is off by default.

Facebook's OpenZL project shipped v0.3.0 on 29 September with a new LZ engine and a neural Compression Transformer. On 868 families of numeric streams, 34,737 files, 17.9 GB, the Transformer compresses on average 35% better than zstd -19 and lands within 1.1% of trained OpenZL graphs, per the release notes. Flow Engineering, a three-year-old San Francisco startup selling AI tools for hardware design, announced a $50 million Series B at a $750 million valuation on Wednesday, 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 and former Sequoia partner Roelof Botha participating. Anduril, Rivian, Joby Aviation, General Motors PPU, RV Tech and Stoke Space are named as customers.

OpenAI and Synopsys, meanwhile, signed a multi-year partnership to build GPT-Synopsys, a model for chip design that combines OpenAI's technology with Synopsys' electronic design automation tools, The Decoder reported on 30 September. Customer data will not be used for training and will be stored encrypted, both companies said, and early tests with semiconductor customers are already underway.

For publishers, none of this changes the core asymmetry. Google decides the price, the terms and the opt-out, and the market for the content it uses is not a market in any normal sense. The Information's reporting, relayed by The Decoder, is the clearest public picture so far of what that arrangement pays out, and the picture is not flattering.