Google Pays About 100 Publishers for AI Answers, and Some Get Almost Nothing

The pilot launched less than a year ago. Publishers can see in Google Search Console how often their content is pulled into an AI answer, and how much they earned for it. What they cannot see is the formula.

Some participants told The Information they do not know how Google calculates the payments. The amounts can shift from month to month without explanation. The spread is wide. One publisher received $50,000 to $60,000 over a few months. Another earns more than $1 million a year. Small sites received less than $1,000 over several months, according to The Decoder's write-up of the report. Content on niche topics with strong followings, anime and gaming among them, appears to earn more. That is the deal on offer nine months into a programme that Google has never had to run at all. It lands at the end of a September in which the company's relationship with the open web was already being renegotiated in courtrooms and regulators' offices on three continents.

What publishers are weighing

Some larger publishers are refusing to join the programme to push Google to pay more, The Information reported. Their traffic is already falling, and multiple studies show AI Overviews sharply reduce visits to the open web. That is the trade being made explicit: accept a rate you did not negotiate, or hold out while your referral traffic erodes anyway.

The structural problem is that individual deals keep publishers negotiating separately rather than collectively. A few benefit, most get little, and publishers that walk away put limited pressure on Google because other sources fill the gap. The Decoder's Matthias Bastian described the arrangement as a divide-and-conquer approach that creates a prisoner's dilemma for publishers. It is hard to argue with the label when the payment model lets Google decide what content is worth.

Google's answer to the criticism has been a combination of selective licensing deals and an opt-out feature. Tempo.co English reported on 29 September that publishers are pushing for controlled access rather than an all-or-nothing switch. PPC Land's 29 September piece on Google search chief Nick Fox sits in the same conversation. The Information's original story ran on 29 September, with syndicated versions appearing at Investing.com and Investing.com Canada on 30 September and a summary at inkorr.com on 29 September.

Regulators are not waiting for the pilot to mature. 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, according to The Decoder. The Commission 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 ruling that could reset the price

A German court has ruled that AI Overviews are Google's own content, not summaries of existing material, The Decoder reported. 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. Each source's exact contribution to an answer is hard to track. Managing the payments would add overhead, and the fees themselves could cut into Google's margins. The court's framing turns an opaque revenue-share pilot into a potential liability.

None of this is happening in a stable traffic market. The search referral declines that make the payments matter have been documented for months. Reports in March and June tracked small publishers losing as much as 60 percent of search traffic. A February antitrust filing accused Google of cannibalising publisher traffic, and by 24 September one tracker put the decline for publishers at 40 percent year on year. Microsoft's decision in September to add an "ask a follow-up" feature to Copilot results, reported on 7 September, points the same direction for Bing referrals.

Meanwhile the rest of the industry is reorganising around AI-assisted work in ways that assume fewer clicks. Arrowfly, the parent company of The Robot Report, launched an AI for Engineers platform on 30 September. It cited a network of more than 20 engineering brands reaching 2.1 million verified industry readers and 29 million website visitors, with a conference set for 20 to 22 September 2027 in Henderson, Nevada. It is a media business built for a subscription-and-events future, not a pageview one.

The technical side of the same shift keeps accelerating. OpenAI and Synopsys signed a multi-year partnership on 30 September to build GPT-Synopsys, an AI model for chip design that runs on OpenAI infrastructure with customer data excluded from training, per The Decoder. Valor, Atreides and Sequoia backed hardware-design startup Flow Engineering at a $750 million valuation on Wednesday, TechCrunch reported. On the same day, the VUSEC research group published Branch Target Reuse, a Spectre-v2 attack against JIT engines in Firefox's SpiderMonkey, Oracle GraalVM and the Linux kernel that leaks 8 bytes per second. Facebook released OpenZL v0.3.0, which the project says decompresses 144 percent faster than Zstandard at equivalent settings.

Smaller, more practical items filled out the week. Modal published Quail, an inference engine that it says hits more than a billion tokens per minute per H100 GPU on one multi-join AI-SQL query, under 6 cents per billion tokens on its own platform. Elastic's search labs tested TypeSafe AI's Jev model as an ecommerce reranker on 250 Amazon Shopping Queries, lifting nDCG@10 from 0.9351 to 0.9565 at $0.042 per million input tokens. Magnitude, a YC S25 company, shipped an open source inference engine for agents that it claims runs open models up to 2x faster than llama.cpp, with 92 percent faster decode on Metal.

And the labour questions underneath all of it are being argued in public. A software engineer writing at martinelli.ch on 30 September described AI coding agents collapsing specialised roles back into one all-rounder, with the specification becoming the most important artefact in a project. Pooyam, an engineer working on LLM training and inference performance at the GPU and TPU level, wrote the same day that implementation cost is heading toward zero and that the remaining bottleneck is deciding what to build. Gerben Rijpkema, an AI engineer with a control engineering background, argued that the noisy, messy data is exactly where LLMs earn their place, which is why narrow use-case-specific architectures succeed where general tools fail.

For publishers, the arithmetic is less philosophical. A payment that rounds to less than 0.1 percent of ad revenue is not a licensing deal in any meaningful sense. It is a rounding error attached to a product that is actively reducing the traffic those ads once ran against. Until the German ruling is tested elsewhere, or Brussels forces a formula into the open, the terms stay where they have been since the pilot began: set by Google, published nowhere, and renegotiated one publisher at a time.