The Open Markets Institute, a Washington-based group that campaigns against corporate monopolies, today published a 71-page report arguing that the companies which already run the internet's main gateways are positioned to take control of AI agents - the software assistants that choose products, complete purchases, book travel and write messages on a person's behalf. The report, "Same Gatekeepers, New Gateway: AI Agents, Monopoly Power, and Democratic Risk," names Google, Amazon, Microsoft and Meta as the firms best placed to extend their control of cloud computing, foundation models, operating systems, browsers, search and social platforms into the agent market, and sets out a policy programme meant to stop that from happening.

The report is an advocacy document. Its findings and recommendations are the positions of Open Markets, a group whose stated mission, according to its announcement, is to address threats to democracy, individual liberty and economic opportunity "posed by corporate monopolies and concentrated economic power." It was written by Max von Thun, Director of Open Markets Institute Europe, Dr. Courtney C. Radsch, Director of the Center for Media and Digital Governance at Open Markets, George Colville and Dora Wierzbicka, with Barry Lynn as project director and editor. The document is dated November 2026 and carries 186 footnotes. None of the companies it criticises is quoted responding to it.

In Short

The Open Markets Institute, an anti-monopoly advocacy group, published a report today arguing that Google, Amazon, Microsoft and Meta could use their control of search, browsers, phones, cloud computing and messaging apps to dominate AI agents, the assistants that shop, book and write on a user's behalf. It matters to merchants, publishers and advertisers because an agent that picks a single product, rather than showing a list, decides who gets the sale, and the group argues that companies building these agents have strong reasons to favour their own services and paying partners. The report asks regulators in the US, Europe and elsewhere for loyalty duties toward users, guaranteed access for rival agents, easy switching and, in some cases, forced separation of cloud businesses from AI; these are proposals from one advocacy group rather than rules, and the report presents no new data of its own.

What the report argues

The central claim is simple. Search engines and social feeds still leave the final choice to the person using them; an agent increasingly makes that choice. "It does not present options so much as it selects them," the report states. "It does not surface information so much as it filters, synthesizes and interprets it." For Open Markets, whoever owns the agent therefore owns something closer to the decision than any previous intermediary did.

The announcement condenses the argument into three forces that could concentrate the agent market quickly. The first is control over distribution: dominant firms can use operating systems, browsers, search engines and platforms to put their own agents in front of users and disadvantage independent rivals. The second is data-driven lock-in: as agents accumulate personal information, preferences and transaction histories, switching providers could become harder, rewarding early leaders. The third is control over access and standards: firms that run the infrastructure, protocols and services agents depend on could decide which competitors, merchants and publishers take part, and on what terms.

The body of the report adds network effects, advertising and a web of investments as reinforcing factors, then devotes roughly 17 pages to harms ranging from surveillance pricing to algorithmic collusion. Its definition of an agent is technical: a large language model plus an "agent harness," the software that carries out the tool calls the model requests. It adopts programmer Simon Willison's line - "An LLM agent runs tools in a loop to achieve a goal" - and concentrates on general-purpose assistants such as ChatGPT, Claude, Gemini, Muse and Copilot.

"Big Tech already controls much of the infrastructure that powers the internet. Now those same companies want to control the AI agents that will increasingly decide what people buy, read, and do," von Thun said in the announcement. "If policymakers fail to act, we risk transforming today's digital gatekeepers with something even more dangerous: corporations that make choices for us. Europe has a central role to play in ensuring that this doesn't happen." Radsch added: "As agentic systems grow more sophisticated and concentrated there is a serious risk that predictive personalization will become anticipatory persuasion."

According to the report, Meta's introduction of its Muse agent "marks the first true mass market product." PPC Land reported that Meta launched Muse in the US on September 8, 2026, with WhatsApp as one of its conversational surfaces, running on a dedicated virtual machine with its own browser and checking out through Link by Stripe.

Why model quality matters less than reach

Open Markets argues that competition has moved away from the models themselves, citing diminishing returns from scaling and open-weight models, many from China, that it says are commoditising the broader model market. "As models become increasingly interchangeable for many use cases," the report says, the advantages of leading model developers are being eroded, and differentiation shifts toward "effective application, distribution, and integration."

That shift favours companies that already sit in front of billions of users through search, browsers, operating systems, productivity software, email, messaging, marketplaces and social media. The report cites examples it treats as early evidence:

  • Meta integrating Muse into WhatsApp and Facebook Messenger, and revising WhatsApp's terms to stop rival AI providers distributing their assistants there. According to the report, the AI button and search features in WhatsApp still route exclusively to Meta's own assistant.
  • Meta giving Muse proprietary integrations with Instagram and Facebook Marketplace that no rival has.
  • Google turning search into a Gemini product, starting with AI Overviews and expanding into AI Mode, and planning, according to the report, to eventually replace the traditional search interface with a mix of chatbot and agentic features.
  • Google's "Personal Intelligence," which lets Gemini connect to Gmail, Google Photos, Search and YouTube history, and its Gemini Spark assistant, which promises to manage a user's "digital life" from email to shopping to invoicing.
  • Microsoft's "Agent Mode," which builds Copilot into Word, Excel, Teams and PowerPoint.
  • xAI embedding Grok as a dedicated tab in X, and Meta experimenting with something similar in Threads.

The report calls such conduct "anti-competitive bundling, tying, and self-preferencing practices" that are "severely distorting nascent competition in the agentic market." That is the group's characterisation; none of the examples it lists has been found unlawful in the agent context.

PPC Land's own reporting fills in some of those details. Gemini Spark launched at Google I/O on May 19, 2026 as a background agent built on Gemini 3.5 Flash, connected to Workspace and to third-party services; it remains unavailable in the European Economic Area, Switzerland and the UK. A Google privacy policy version dated October 16, 2026 added an exception to a July 2021 promise, allowing Search to use data from connected apps such as Gmail and Google Photos to personalise content, ads included, for users who link those services.

The French competition authority reached a related, more cautious conclusion in July. Its advisory opinion of July 17, 2026, found that OpenAI, Google and Anthropic held over 84% of the global AI agent sector in May 2026, citing Sensor Tower data, and flagged default integration, data advantages and switching difficulty as lock-in risks.

Data, network effects and switching costs

The second leg of the argument is a feedback loop. Interaction data is of limited use when a foundation model is first trained, the report says, but "extremely valuable" for fine-tuning, and most leading assistants' terms of use allow input data to be used for training. More users produce more data, which improves the service, which attracts more users - a wheel that firms with a captive audience on day one are best placed to spin.

Network effects compound the effect. The more businesses integrate with a given assistant, the more useful it becomes to users; the more users it has, the more businesses want to integrate. The report warns that if agents become two-sided platforms, "These network effects risk creating or entrenching gatekeepers with market power likely exceeding even the control wielded by Apple over its App Store or Amazon over its Marketplace," in part because an agent can intermediate across many markets at once and shapes discovery in ways even less visible than a search ranking.

The report is candid that the danger is prospective. "At the time of this report's publication, lock-in to specific agentic assistants remains low, at least on the consumer side," it says, noting that Anthropic and OpenAI let users export their data and make it easy to import context from a rival assistant. Open Markets wants those options written into law. The lock-in it anticipates would come from conversation histories, inferred preferences, exclusive integrations and custom workflows that would be costly to rebuild elsewhere.

Protocols as the new chokepoint

The section most relevant to merchants concerns standards. Because scraping and browser automation struggle with the unstructured web, the report says, the industry is building protocols that let agents query inventory, manage carts and pay without visiting pages designed for humans. "Whoever writes these rules will shape the open web, the agentic web, and online markets for years to come," the authors write.

The report's precedents are robots.txt, which shaped access to the web for decades on adoption alone and was only formally adopted by the Internet Engineering Task Force in 2022, and Google's AMP, which the report says tied preferential search placement to the use of Google's own format.

Several agent protocols already exist. Anthropic developed the Model Context Protocol (MCP) in 2024; it is now maintained by the Agentic AI Foundation under the Linux Foundation and lets developers connect AI tools to external data sources and software. Google developed A2A for agent-to-agent communication. Open Markets is broadly positive about MCP, saying such protocols lower integration and switching costs, though a footnote adds that the Linux Foundation "has long been criticized for favoring the interests and priorities of its highest-paying members," which it identifies as large technology companies.

Its harder criticism falls on commerce protocols. The Agentic Commerce Protocol, it says, is "open, but developed and controlled by OpenAI and Stripe." Google's Universal Commerce Protocol (UCP) draws the sharpest language in the report. While UCP is technically open source, Open Markets argues, a merchant that wants products to appear in Gemini and AI Mode must have an active Google Merchant Center account and use Google Pay, after which Google invites it to use Google's tools to "optimize [their] presence in the conversational commerce era." The report's verdict: "What is presented on paper as an open-source solution is in fact an attempt to lock in merchants - and their advertising - as part of Google's latest effort to entrench its advertising monopoly."

PPC Land's coverage supports part of that description and leaves the rest as the group's interpretation. Google launched UCP on January 11, 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by more than 20 companies including Adyen, Mastercard, Stripe and Visa. When Google's UCP checkout help page appeared on March 2, 2026, it described Merchant Center as the central hub for product data and said shoppers pay with Google Pay using payment methods stored in Google Wallet, with the merchant's payment provider required to accept a Google Pay token. The feature was US-only and limited to selected merchants. Whether that amounts to locking merchants into advertising is a judgement: Merchant Center is also the feed console behind Google's Shopping ads, but the help page as reported by PPC Land did not make ad spending a condition of UCP checkout.

UCP has drawn pricing criticism before. Two days after its launch, Groundwork's Lindsay Owens described it as a surveillance pricing risk and Google called the pricing claims inaccurate, adding that merchants are barred from showing higher prices on Google than on their own sites.

Steering at the protocol layer is the report's broader concern. Because Google builds both a leading agent and a leading commerce protocol, it says, Google "can dictate which products are surfaced to consumers using its agentic assistant, favoring merchants that adopt its standards and excluding those that do not." The report offers no measurement showing that this happens in practice.

Advertising inside the agent

Open Markets treats advertising as the business model most likely to pull agents away from their users. "For AI assistants not sold as enterprise solutions or subscriptions, advertising is an inevitable monetization strategy," it says, because agents are interactive, deeply personalised and directly drive decisions.

The report traces the 2026 debate in some detail. OpenAI introduced advertising to ChatGPT in February 2026, according to the report, despite chief executive Sam Altman having previously called ads in AI services "uniquely disturbing." The decision prompted OpenAI researcher Zoë Hitzig to leave and write a New York Times op-ed comparing OpenAI to Facebook. PPC Land reported that OpenAI announced its ad tests for the US free and Go tiers on January 16, 2026, and that Reuters reported on April 9, 2026 that OpenAI told investors to expect $2.5 billion of ad revenue in 2026 and $100 billion by 2030, with Barclays separately projecting $102 billion.

Anthropic used the moment to present Claude as "acting unambiguously in our users' interests," the report notes, including with Super Bowl commercials mocking AI advertising. PPC Land reported that Anthropic ran four spots under the "A Time and a Place" campaign, two of them scheduled for Super Bowl 60 on February 8, 2026, and pledged to keep Claude ad-free on all subscription tiers. Its IPO filing of October 26, 2026 described no advertising business. Perplexity also dropped advertising in early 2026, the report says, citing the Financial Times.

The authors take no side's word for it. "It is important not to take companies' statements or actions at face value, particularly when it comes to whether they are for or against advertising," they write, pointing out that Anthropic and Perplexity rely largely on subscription revenue and that the backlash against OpenAI gave both a way to stand out. Their proposed answer is binding rules that apply "whether they adopt advertising or not."

Google, the report says, "has shown far fewer reservations," with ads already integrated into AI Overviews and AI Mode while the Gemini app remains ad-free "(for now)." Google announced shopping ads for AI Mode on February 11, 2026, and said on May 19 that AI Mode had passed 1 billion monthly active users. Muse does not carry ads, the report notes, but advertising is the basis of Meta's business. PPC Land reported in October that Profound extracted Muse's instructions, which require well-known merchants to rank among the top five picks, and that Meta extended Muse to small businesses with connectors to their Meta ad accounts on September 29.

The report sketches how paid influence could work. A luggage company that pays for placement would see its products surfaced first when a user asks for a recommendation or delegates the purchase entirely. A more elaborate version would be a deal under which an agent routes every travel-related task through one booking platform. In a competitive market, the report concedes, exclusive arrangements are "a familiar if contentious commercial practice"; in a concentrated one, it argues, they operate "through invisible preferences built into the system they rely on."

Stakes, partnerships and acquihires

Open Markets argues that the apparent variety of AI companies masks commercial dependence. It says Microsoft holds a 27% stake in OpenAI and, under a revenue-sharing clause, receives 20% of OpenAI's revenue; the news report it cites for the revenue share is headlined as describing a cap on those payments. Anthropic, according to page 28 of the report, received $13 billion from Amazon and $10 billion from Google, giving each an approximate 15% stake, and committed to spend $100 billion on Amazon Web Services over the coming decade.

The report is inconsistent on those numbers. Page 54 says Amazon "has invested a total of $8 billion in Anthropic" and that Google holds a 14% stake. PPC Land's reading of Amazon's filings reconciles the Amazon figures: Amazon invested $8 billion in Anthropic convertible notes between the third quarter of 2023 and the fourth quarter of 2025, and put another $5 billion into Anthropic preferred stock after March 31, 2026. The same PPC Land article found no ownership percentage for Amazon or Google in published accounts of Anthropic's S-1, citing analyst estimates of 10% to 15% for Alphabet.

Nvidia, according to the report, holds stakes in OpenAI, xAI, Reflection AI, Mistral and several "neocloud" operators, and plans to acquire Hugging Face; the report cites a Financial Times article of September 3, 2026 headlined as a $13 billion deal. Talent deals form the third strand: Google hiring Windsurf's leadership, Meta bringing in Scale AI's team, Microsoft absorbing Inflection AI and Nvidia taking much of Groq's talent and assets. These "acquihires," the report says, capture talent "without triggering antitrust scrutiny."

Meta's purchase of Manus, a Chinese agent start-up, is the report's example of regulators becoming more alert. Meta closed the roughly $2 billion acquisition on December 29, 2025, and China's National Development and Reform Commission ordered it unwound on April 27, 2026. The report notes that Beijing acted on national security rather than competition grounds.

The harms Open Markets expects

Most of the report's length goes to consequences. The authors divide them into societal and economic harms and acknowledge that some are hypothetical, arguing that "both the commercial incentives and the technological capabilities to exploit them already exist."

Steering and self-dealing

The report imagines Gemini Spark directing users to Google Flights and Google Shopping or favouring Pixel phones, Copilot steering toward LinkedIn, GitHub and Azure, and Muse favouring Facebook Marketplace. Rivals, it says, "are likely to find themselves at a systemic yet imperceptible disadvantage" because of the architecture of the agent rather than the quality of their products. These are scenarios, not documented incidents. The report's documented precedents come from older markets: the European Commission's EUR 2.42 billion fine over Google's comparison shopping service and the FTC's ongoing lawsuit against Amazon, which covers preferential placement.

Surveillance pricing

Agents could make individually tailored prices far easier, the report argues, because a single assistant collapses data that today has to be stitched together from cookies and devices. It cites Amazon's six-day DVD pricing trial in 2000, a 2023 study cited by the FTC that found grocery stores personalising prices at the zip code level, and a Dutch Authority for Consumers and Markets finding that the platform Wish personalised prices without disclosure. Its hypothetical examples include a higher price when an agent detects urgent need, such as medication or last-minute travel, and data sold to an airline that lets it charge more to someone booking for a funeral. "It is currently impossible to audit agentic AI recommendations," the report states.

Tolls on sellers

Amazon's effective commission rate for third-party sellers reached an estimated 45% in 2023, up from 19% in 2014, as its e-commerce market share grew from roughly 22% to roughly 40%, according to estimates the report cites from the Institute for Local Self-Reliance and others. Citing McKinsey, it says $3 trillion to $5 trillion of global retail revenue could flow through agent-mediated commerce by 2030. And citing The Information, it says OpenAI's Instant Checkout charged retailers a 4% sales commission before the initiative "floundered."

On timing, the report says OpenAI launched Instant Checkout in February 2026. PPC Land reported OpenAI announcing Instant Checkout and the Agentic Commerce Protocol on September 29, 2025, with merchants paying a fee on completed purchases. By March 2026, Walmart said in-chat conversion through ChatGPT checkout was three times lower than click-outs to its own site.

The report also accuses leading AI firms of the copying that Amazon has been accused of in its marketplace. According to the report, which cites The Information, Anthropic partnered with Canva and Figma to integrate their services into Claude and later built its own competing service. "It is naïve not to expect that successful apps or integrations built on AI gatekeepers' infrastructure will be copied and unfairly disadvantaged in this way," it says. It adds that OpenAI cut off Cursor after Cursor's acquisition was announced in August 2026, and that Anthropic ended Windsurf's direct access to Claude models with under five days' notice amid reports that OpenAI would buy it.

Correlated behaviour, outages and collusion

Three further risks stem from many users relying on the same model. The first is correlated behaviour: the report points to the 2010 Flash Crash, a 36-minute, $1 trillion stock market collapse driven by similar trading algorithms, and notes that retail brokers such as Robinhood now let users set up agent-driven trading. The second is a single point of failure, illustrated by the 2021 AWS outage, which the report says cost billions of dollars in lost commerce, and by prompt injection attacks that could hijack many accounts at once. Among its sources is the Dutch data protection authority's warning about OpenClaw agents of February 12, 2026.

The third is algorithmic collusion. The report recalls the 2011 pricing-bot war that listed the biology book "The Making of a Fly" at over $23 million, and the 2025 settlement in which 26 property management firms using RealPage software paid $142 million over alleged rent coordination. It argues that a market served by a few similar agents would give sellers' pricing algorithms a stable signal to align on, and could also push buyer agents toward the same tactics, a reverse "hub and spoke" effect that would work against sellers.

Publishers and information

The publisher section repeats figures PPC Land has covered. Citing Chartbeat data reported by Axios on March 17, 2026, the report says search referrals to small publishers fell 60% between 2023 and 2025, against 47% for mid-sized and 22% for large publishers. PPC Land reported the same Chartbeat declines, based on weekly tracking from January 2024 to February 2026. The report adds that AI applications account for less than 1% of publisher traffic, that Google still delivers roughly 85% of external referral traffic, and that conservative estimates put industry losses from zero-click AI search at $2 billion a year.

On the regulatory side, the report points to the European Commission's antitrust investigation, opened on December 9, 2025, into whether Google used publisher content for AI Overviews and AI Mode without fair compensation, and to the UK requirement that Google offer publishers an AI opt-out without retaliation. PPC Land reported that the CMA imposed that Publisher Conduct Requirement on June 3, 2026, with its main obligations applying from December 3, 2026.

The wider democratic argument rests on examples such as xAI's Grok briefly censoring unflattering mentions of Donald Trump and Elon Musk in early 2025, UK AI Security Institute research linking persuasiveness to lower accuracy, and a study of 21 large language models showing less variation in preferences than humans. Two claims in its section on state capture carry no footnote: that the US government temporarily limited the release of Anthropic's Fable 5 model to US citizens, and that OpenAI has proposed the government receive a 5% ownership stake in leading AI companies. The executive summary also refers, without a citation, to unauthorised hacks by AI agents in 2026 affecting the Australian healthcare system, the US Securities and Exchange Commission and Hugging Face.

Regulators are already acting, the report says

Open Markets says policymakers still have time to act and that some enforcers have started to. Its main European example is WhatsApp. According to the report, Meta changed WhatsApp's Business terms in October 2025 to ban third-party AI assistants, Italy's competition authority moved for interim measures a month later, and in early 2026 the Commission's own plan for interim measures led Meta "to reverse the changes and allow rival chatbots access in Europe."

PPC Land's coverage shows a longer sequence. The Commission sent Meta a Statement of Objections and notice of intended interim measures on February 9, 2026 in case AT.41034. Meta then introduced a pricing framework in March, which the Commission said in a supplementary Statement of Objections on April 15 was equivalent in effect to the original ban. On June 9, the Commission imposed interim measures ordering Meta to restore free access for third-party general-purpose AI assistants on the terms in place before October 15, 2025, within five working days.

The report says that in July 2026 the Commission issued measures under the Digital Markets Act requiring Google to give rival AI services fair access to the Android features Gemini already uses; the footnote for that statement cites the Commission's announcement opening specification proceedings on Google's interoperability and search data obligations.

It also cites Amazon's lawsuit against Perplexity, filed on November 4, 2025, over the Comet browser agent, and Amazon's decision to block Meta's Muse from placing orders on Amazon.com, reported on September 20, 2026. Open Markets does not take a side on whether Amazon's security concerns are justified. Its point is structural: Amazon's terms reserve "sole discretion" to refuse access to agents, so "even if an agent is perfectly safe and secure, Amazon can still refuse access." That conflict of interest is why it wants independent certification.

In the US, the report cites Senator Mark Warner's discussion draft of an AI AGENTS Act, which it says would introduce duties of loyalty, interoperability requirements and privacy protections for the agent market, and the proposed Preventing Algorithmic Collusion Act, which would bar pricing algorithms that rely on non-public competitor data. In the UK, the Digital Regulation Cooperation Forum published a foresight paper on agentic AI on March 31, 2026, concluding that agents do not fall outside existing UK regimes.

The recommendations

Open Markets groups its remedies under three headings - making agents serve their users, creating an open market for agents, and using competition policy - and frames them as a roadmap for US, European and other policymakers. None is law.

Duties to users. Providers would owe a binding duty of care modelled on the fiduciary duties of financial advisers and lawyers, disclose material conflicts of interest, and face a rebuttable "presumption of disloyalty" that puts the burden on them to show regulators they are not favouring their own commercial interests. Exploiting sensitive information such as finances, sexual orientation or health for commercial gain would be banned. Agents that file taxes, recommend investments or give health advice would meet the standards of licensed professionals doing the same work.

Limits on surveillance. Collection would be restricted to what is necessary, sensitive biometric, cognitive and neural data would be off limits or tightly conditioned, and users could view and delete the "memory" an agent holds. Providers would offer low-data versions, and enterprise-style no-training protections would extend to individuals. The report says the GDPR already embodies many of these principles but has been "significantly underenforced."

Interoperability. Vertical interoperability would give agents fair access to email, calendars, messaging apps, operating systems and productivity tools. Horizontal interoperability would let users move conversation and transaction history, inferred preferences and documents between providers in a machine-readable format.

Equal access. Dominant platforms would owe all agents fair, reasonable and non-discriminatory access, so that a company could not offer an API to rivals while giving its own agent faster responses or richer data. The same obligation would apply to dominant agents themselves, which would have to give merchants, publishers and other businesses equal access and could not arbitrarily restrict their visibility.

Independent certification. Rather than letting platforms decide which agents are safe, a neutral body would set technical and security standards, following the model of the UK's Open Banking framework, under which the CMA set up an independent body to govern third-party access to banks.

Competition enforcement. Authorities would use interim measures early, block mergers touching agents and adjacent markets such as chips, cloud and devices, and treat minority stakes, partnerships and acquihires as de facto mergers where warranted. Structural remedies would include divestments and line-of-business restrictions: Open Markets repeats its earlier call for Amazon, Google and Microsoft to divest their cloud businesses and says regulators could even consider barring large technology firms from owning AI agents at all.

Algorithmic collusion. Algorithms that systematically distort competition would face a rebuttable presumption of collusion, with audit and logging duties. Agents could not use non-public data from business customers to price, rank or promote those customers' competitors, and regulators would get audit access to run tests such as "calibrated-regret testing."

What the report does not establish

The report contains no original measurement. Its evidence comes from news coverage, company announcements, academic papers and third-party estimates such as Chartbeat, the Institute for Local Self-Reliance and McKinsey. Several of its most vivid harms are explicitly hypothetical. It concedes that consumer lock-in is low today and that open protocols such as MCP are an "important source of optimism."

There are also errors and gaps. The Amazon and Google figures for Anthropic differ between pages 28 and 54. The February 2026 date the report gives for Instant Checkout's launch does not match OpenAI's September 2025 announcement. Its account of the WhatsApp case compresses five months of objections, fee changes and a formal order into a single reversal. Some claims about government action and agent hacks carry no source. In the Cursor episode, the report's text names xAI as the buyer while the OpenAI statement it cites refers to SpaceX, and its publisher chart labels the Chartbeat period 2023 to 2025, while the underlying data ran from January 2024 to February 2026. The authors credit Figure 9, the publisher referral chart, as "Created with the assistance of Claude AI (Anthropic)," a product of one of the companies the report examines.

The central policy ideas are also not new to Open Markets: the cloud divestment proposal draws on its earlier reports, and the announcement says the work builds on Radsch's "The Battle for Cognitive Liberty in the Age of Corporate AI" and Senior Fellow Sally Hubbard's "Who Do AI Agents Work For?"

Why this matters for marketers

For advertisers, retailers and publishers, the report is a map of where the next intermediary fees and ranking rules could come from. Agents are already shopping. Meta's Muse buys through a browser and Link by Stripe, Google's UCP routes checkout through Merchant Center data and Google Pay, and ChatGPT has carried ads since early 2026. Each of these products sets terms for who gets surfaced and on what conditions, and Open Markets argues those terms will drift toward the platform's own interests unless law intervenes.

The questions are concrete. Will agents disclose when a recommendation is paid? Who decides which agents may shop on a retailer's site, as Amazon has done with Perplexity and Muse? Will commerce protocols stay neutral when their owners also sell the ads? PPC Land has documented each of these disputes over the past year; today's report argues they are symptoms of one structural problem. Whether regulators agree is another matter, and the US measures it cites are still proposals.

Timeline

Summary

Who: The Open Markets Institute, a Washington-based anti-monopoly advocacy group. The report was written by Max von Thun, Dr. Courtney C. Radsch, George Colville and Dora Wierzbicka and edited by Barry Lynn. It focuses on Google, Amazon, Microsoft and Meta, with further discussion of OpenAI, Anthropic, Nvidia, xAI and Apple.

What: A 71-page report, "Same Gatekeepers, New Gateway: AI Agents, Monopoly Power, and Democratic Risk," arguing that control of distribution, data and standards could let incumbent technology companies dominate AI agents. It recommends fiduciary-style loyalty duties, professional standards for agents, surveillance limits, vertical and horizontal interoperability, fair and non-discriminatory access, independent certification, tougher merger control, structural separation including cloud divestment, and new rules on algorithmic collusion. It contains no original data and has internal inconsistencies on Anthropic investment figures.

When: Published today, November 9, 2026. The document is dated November 2026.

Where: Issued from Washington, D.C., with recommendations aimed at US, European and other regulators. The enforcement examples it cites come mainly from the EU, Italy, the UK and the US.

Why: Open Markets argues that agents differ from search and social media because they make choices rather than present them, so whoever controls them could steer purchases, prices and information at a scale beyond today's platforms. It contends that policymakers still have time to act while the agent market, its protocols and its business models remain unsettled.