Google Pays About 100 Publishers for AI Content, and Most Get Very Little

On 30 September, The Decoder reported that Google is paying about 100 digital publishers for content used in AI Overviews, AI Mode and the Gemini chatbot. Some of those publishers do not know how their payments are calculated. The pilot programme launched less than a year ago, according to The Information, and publishers can track usage and earnings inside Google Search Console.

The spread is enormous.

For several small and midsize blogs and websites, the payments amount to less than 0.1 percent of ad revenue, The Information reported. One publisher received $50,000 to $60,000 over a few months, while another earns more than $1 million a year. Small sites received less than $1,000 over several months. Content on niche topics with strong followings, such as anime and gaming, appears to earn more, according to the report.

Publishers cannot see the formula

Payments are supposed to depend on how much each source contributes to an AI answer. But some participants told The Information they do not know how Google calculates them, and that payments can change from month to month without explanation. That opacity matters, because the amounts are already the subject of an antitrust file. In July 2025 independent publishers filed a complaint with the European Commission over AI Overviews. In September 2025 Rolling Stone parent Penske Media sued Google over lost traffic and ad revenue. The European Commission opened an antitrust investigation in December 2025 into whether Google imposes unfair terms by using publishers' content for AI features without adequate payment or a genuine way to opt out.

Some larger publishers are refusing to join the programme in an attempt to push Google to pay more, according to The Information. Their traffic is already falling, and multiple studies show AI Overviews sharply reduce visits to the open web. The refusal is a bet that collective pressure works better than an individual deal. So far there is little evidence it has moved Google's rates.

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 Google uses their work in an AI answer. The Decoder noted that tracking each source's exact contribution to an answer would be difficult, if not impossible, and that managing those payments would add overhead on top of the fees themselves.

Traffic is the bigger number

Payment totals are one side of the ledger. The other is the traffic that no longer arrives. Recent headlines tracked by this desk point in one direction: an antitrust filing says Google cannibalises publisher traffic, small publishers are hit hardest by search declines, and one September analysis put the year-on-year decline for publishers at 40 percent. Those figures sit outside the primary documents in this dossier and should be treated as context, not as confirmed measurements.

What the dossier does confirm is the mechanism. AI Overviews answer questions on the results page, so the click that used to go to a publisher does not happen. Google then pays a subset of publishers for the content that fed the answer, at rates it sets, under terms it writes, with an opt-out it controls. The Decoder described the arrangement as a divide-and-conquer approach that creates a prisoner's dilemma: a few benefit, most get little or nothing, and publishers that walk away put little pressure on Google, because other sources fill the gap.

Individual deals also keep publishers negotiating separately rather than collectively.

The structural argument is that if Google licensed everything at a fair rate, its margins would take a hit. So it licenses selectively instead, choosing which content is worth what. That is not a new strategy. The Decoder noted it has served the company for decades as it drew more content from the web into search, and that AI Overviews push the trend much further, much faster.

The money is moving to the tooling layer

While publishers argue over fractions of a percent, capital is flowing into the infrastructure that builds and runs AI systems. On Wednesday, TechCrunch reported that 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. The round was co-led by Antonio Gracias of Valar Equity Partners and Gavin Baker of Atreides Management, with Sequoia Capital participating alongside former Sequoia partner Roelof Botha, who invested personally and joined the board. Flow names Anduril, Rivian, Joby Aviation, General Motors PPU, RV Tech and Stoke Space as customers.

The same day, The Decoder reported that OpenAI and Synopsys signed a multi-year strategic partnership to build a specialised chip design model called GPT-Synopsys. Synopsys makes electronic design automation tools, the software engineers use to design chips. OpenAI is licensing those tools so the model can reason about chip design and verification and operate the tools directly, with engineers delegating objectives and reviewing output. The model will run on OpenAI's infrastructure. Both companies said customer data will not be used for training and will be stored encrypted. Early tests with semiconductor customers are already underway, and the two will market the product together and share revenue. Synopsys CEO Sassine Ghazi said AI could significantly speed up the design process; OpenAI co-founder Greg Brockman framed the partnership as a path to better chips and better AI.

The pattern repeats lower down the stack. Quail, an inference engine built by Modal with Carnegie Mellon University's Full Stack Data Lab, reported hitting over a billion tokens processed per minute per H100 GPU on one multi-join AI-SQL query, more than 10x faster than a vLLM baseline on the same hardware, and under 6 cents per billion tokens on Modal. On a new benchmark for AI-SQL queries it ran 1.84x faster than vLLM, geometrically averaged over tasks. Separately, an open source project called Magnitude, posted to GitHub on 30 September, claims up to 2x faster inference than llama.cpp by compiling and tuning kernels on the user's own hardware, with 92 percent faster decode on Metal and 19 percent on CUDA.

None of this pays a publisher's bills. It does explain where the value is being captured.

Security researchers keep finding the same opening

There is one more thread worth pulling. On 30 September, researchers at VUSEC published Branch Target Reuse, a new Spectre-v2 attack against just-in-time compilers. The insight is that modern CPUs restore architectural code coherence after self-modification but do not necessarily invalidate stale indirect branch prediction entries. In JIT engines those stale targets can outlive the original code and be reused when the code cache is repopulated, producing a speculative execute-after-free primitive. 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 walking the kernel task list to extract a root password hash. They also bypassed the bpf_jit_harden constant-blinding option.

That is a different problem from Google's publisher programme, but it lands in the same week and on the same infrastructure: the browsers, runtimes and kernels that AI features increasingly depend on. The dossier also carries a report from machine.news, dated 30 September, about a developer claiming to be an Apple engineer who built a public GitHub project that manipulates internal activations in Alibaba's open-weight Qwen3 models to induce pain-like states, and who deleted a promotional post after criticism. The write-up says the work is based on a paper titled The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It, whose authors do not claim the models consciously feel pain.

For publishers, the practical question remains narrower. About 100 are inside Google's programme. The rest are watching search traffic fall, waiting for a court or a regulator to decide what their content is worth.