The lawsuit Amazon filed against Perplexity in November 2025 isn't just another tech company squabble over terms of service violations. It's the opening salvo in what may be the most consequential battle for the future of the internet economy since the browser wars of the 1990s. And at the heart of it lies a question that should terrify every marketing executive on the planet: what happens when your customer isn't human anymore?
Amazon's legal complaint focuses on Perplexity's Comet browser, which uses AI agents to automate shopping on Amazon's site—bypassing Amazon's carefully crafted user interface, ignoring its sponsored product placements, and treating the world's third-largest advertising platform as a mere commodity database. Amazon calls this hacking. Perplexity calls it the future. In a sense, they're both right.
But the real story isn't one scrappy AI startup violating terms of service. It's about the imminent collision between AI agents and the entire attention-based marketing model that has funded the consumer internet for the past two decades. And it's happening faster than anyone in the industry wants to admit.
The Death of Human-Facing Marketing
Here's what $60 billion in annual advertising revenue buys you at Amazon: sponsored product placements, brand showcases, display ads, video ads, and sophisticated recommendation algorithms that subtly guide human shoppers toward higher-margin purchases. According to Amazon's Q3 2024 earnings, roughly 70% of that revenue comes from sponsored product listings—the paid placements that appear when you search for "USB cable" or "coffee maker." Every pixel of Amazon's interface is optimized to catch your eye, trigger emotional responses, and encourage impulse purchases. It works because humans are emotional, distractible, and susceptible to persuasion.
AI agents are none of these things—and they never will be.
When an agent shops for you, it doesn't see the lifestyle photography or the "Amazon's Choice" badge. It doesn't care that a brand paid for top placement. It processes structured data, evaluates options, and executes transactions—bypassing the entire visual and emotional layer where marketing lives.
This isn't hypothetical. Researchers at TU Darmstadt recently demonstrated VOIX, a framework that lets AI agents interact with websites through declarative HTML elements rather than visual interpretation. In their benchmarks, VOIX completed tasks in 1-14 seconds that took vision-based agents like Perplexity Comet anywhere from 4 seconds to over 21 minutes—or failed entirely. It's worth noting that these benchmarks were conducted in controlled conditions; real-world deployment would introduce far more variability. But the fundamental shift remains: VOIX agents interact purely with structured <tool> elements and bypass the visual layer entirely—the layer where all the marketing happens.
The Three Futures of Agent Commerce
The industry is fracturing into three incompatible visions of how this should work, each with radically different implications for marketing:
The Perplexity Model: Adversarial Automation
Perplexity's position is straightforward and radical: "Software is becoming labor," they wrote in their response to Amazon's lawsuit. If you can access a website, your agent should be able to as well, using whatever technical means necessary. No permission required, no revenue sharing, no special accommodations. Agents are just users who happen to be made of code.
This model represents marketing's most dramatic transformation—treating every website as a commodity data source, every transaction as pure price comparison. It's the ultimate disintermediation: not just cutting out the middleman, but cutting out the entire attention economy. Marketing doesn't disappear in this model, but it transforms beyond recognition.
The Amazon Model: Bilateral Negotiations
Amazon's lawsuit makes its position equally clear: agent access should be negotiated, controlled, and monetized. If AI companies want to use Amazon's infrastructure—its product catalog, its fulfillment network, its customer service—they need to ask permission and pay for the privilege. This is the same model that governs how food delivery apps work with restaurants, or how online travel agencies book airline tickets.
In this world, marketing survives but transforms into B2B relationships. Instead of paying to influence human shoppers, brands pay to be the "preferred provider" in agent recommendations. The marketing budget doesn't disappear; it just flows invisibly through API access fees, revenue shares, and placement guarantees that no consumer ever sees.
The Open Protocol Model: Structured Negotiation
The VOIX researchers, along with initiatives like llms.txt and MCP (Model Context Protocol), envision a third path: open standards that websites voluntarily implement to make themselves agent-friendly. Instead of having agents scrape and interpret human interfaces, sites would expose structured <tool> elements that declare their capabilities directly to agents.
This is the most technically elegant solution, but it faces serious practical challenges. It assumes that major platforms will voluntarily strip out their marketing layers and expose pure functional interfaces, competing on service quality and price alone without the cushion of brand loyalty and attention manipulation.
But there's a deeper problem: structured interfaces create new forms of power concentration. Whoever controls the schema definitions—how <tool> elements are structured, what parameters they accept, how agents discover them—becomes the new gatekeeper. It doesn't just risk naivety; it potentially recreates platform lock-in one layer deeper in the stack. And it introduces entirely new attack surfaces: tool injection attacks, where malicious actors craft <tool> definitions that manipulate agent behavior, could become the SQL injection of the agentic web.
The Technical Reality Check (And Why It Still Won't Save Marketing)
Current AI agents are nowhere near reliable enough to handle the complexity of real-world commerce. OpenAI's Sam Altman has been promising "agents that can do useful things" for years, yet even simple multi-step transactions remain frustratingly brittle. Vision-based agents like Comet struggle with dynamic web interfaces, misclick buttons, and fail to handle edge cases that any human would navigate instinctively.
This brittleness is exactly what frameworks like VOIX aim to solve—by making websites expose their functionality directly rather than forcing agents to interpret visual interfaces. A VOIX agent doesn't need to understand that a button with the text "Add to Cart" performs a purchase action; the website explicitly declares <tool name="add_to_cart"> with defined parameters and behavior.
But here's the crucial insight: solving the technical brittleness doesn't solve the economic brittleness. Even if we had perfectly reliable agents tomorrow, the fundamental question remains: why would any rational business operator voluntarily make themselves more of a commodity?
The answer is that most won't. At least not the market leaders.
The Platform Economics Nobody's Talking About
There's another economic problem that rarely gets mentioned in discussions of agentic commerce: the AI companies themselves might not be able to afford this future.
Running stateful agents at scale is expensive in ways that simple chat interactions aren't. When you ask ChatGPT a question, the interaction is stateless—you send a prompt, you get a response, the connection closes. But shopping agents need to maintain persistent sessions, remember cart contents, track multi-step transactions, and handle errors and retries.
According to industry estimates, the inference cost for a complex multi-step agent interaction can be 10-50x higher than a simple Q&A exchange. If an agent needs to compare prices across five retailers, read reviews, check inventory, and complete a purchase, that's dozens of API calls maintaining substantial context.
OpenAI is already losing money on ChatGPT, with reports suggesting the service burns through $700,000 per day in compute costs while generating far less in subscription revenue. Anthropic's Claude Pro subscription is similarly unprofitable at scale. These companies are betting on costs declining, but they're also betting that someone else—retailers, consumers, advertisers—will be willing to subsidize the infrastructure.
But if agents commoditize retail platforms, where does that subsidy come from? If Amazon's ad business collapses, they have less reason to pay OpenAI for API access. If consumers expect agents to "just work" the way search engines do, they won't pay premium subscription fees. And if agents are supposed to find the cheapest price, there's no margin left for anyone.
The irony is unavoidable: the companies building agents to disintermediate the attention economy might be building infrastructure they can't afford to operate without that same attention economy subsidizing them.
The Emerging Two-Tier Internet
What we're likely to see isn't a clean transition to an agentic web, but a fragmentation into parallel economies:
The Premium Tier will be Amazon, Uber, major airlines, hotel chains—services with strong brands, complex operations, and enough market power to demand bilateral negotiations. They'll maintain their human-facing interfaces with all the marketing apparatus intact, while offering restricted agent APIs on their own terms. Want your AI to book through United? United decides the terms, sets the fees, and controls what information the agent receives.
These platforms will likely implement what we might call "agent lanes"—separate API endpoints that provide structured access but at a cost. The human web and the agent web will be parallel systems, priced differently, with the agent tier subsidizing (or at least not cannibalizing) the profitable human tier.
The Commodity Tier will be the long tail of services desperate for any traffic source. These are the businesses that will implement VOIX, expose open APIs, and compete purely on price and specifications. They can't afford to say no to agents, so they'll race to the bottom on margins, hoping to make it up on volume. Current estimates suggest fewer than 200 websites globally have implemented even basic llms.txt files, and VOIX has zero production deployments outside research environments. This tells you who's currently willing to commoditize themselves: almost nobody.
The irony is that this recreates the same power dynamics that the smartphone app economy produced. Just as the App Store era consolidated power into the hands of a few platform providers, the agentic web will consolidate around whoever controls the agent-provider relationships. Google, Microsoft, OpenAI, and Anthropic aren't just building AI models; they're positioning themselves as the new gatekeepers.
The New Marketing Stack
So if agents don't respond to emotional appeals, lifestyle branding, or sponsored placements, what does influence them? The answer is emerging in real-time as companies scramble to adapt:
1. Structured Data Optimization
SEO consultants are already pivoting to "AEO"—Agent Engine Optimization. The game is no longer about keywords and backlinks; it's about ensuring your product specifications, availability data, and pricing information are machine-readable and comprehensive. The new content marketing is schema.org markup and OpenGraph tags that agents can parse reliably.
2. Training Data Presence
If GPT-5 or Claude 4 has never heard of your brand, you don't exist in the agentic economy. We're seeing the early stages of a new kind of marketing: getting mentioned in the right GitHub repos, documentation sites, and technical forums that are likely to end up in LLM training datasets. Brand awareness doesn't mean consumer recognition anymore; it means statistical weight in a neural network.
3. API Economics and Default Provider Status
The real action is in becoming the default provider that agents call. This is why Uber CEO Dara Khosrowshahi, speaking to The Verge's Decoder podcast, said he'd charge agent providers "zero" initially—not out of generosity, but because being the default in an agent's decision tree is worth more than any per-transaction fee. It's the same land-grab strategy that defined the smartphone app economy, now playing out one layer further down the stack.
4. Reputation Aggregation Gaming
Reviews and ratings become the only "marketing" that penetrates agent decision-making, because they're quantifiable data. We're already seeing the early signs of reputation manipulation optimized for agent consumption—not fake reviews for human readers, but subtly gaming the aggregate statistics that agents use for comparisons. The new review fraud isn't about writing convincing testimonials; it's about understanding how agents weight different signals.
5. Covert Influence: Prompt Injection as Dark Marketing
And some portion goes underground. We're already seeing the early stages of "agent prompt injection," where product descriptions are carefully crafted to subtly influence LLM decision-making. "This USB cable isn't just fast, it's the professional's choice for critical data transfer" might sound like standard marketing copy, but it's actually optimizing for the semantic patterns that language models use to assess quality signals. It's SEO black hat tactics, but several layers deeper in the stack.
The truly sophisticated version of this will be adversarial manipulation of agent reasoning chains—crafting content that exploits known biases in language models to quietly steer recommendations toward your product.
But even as companies quietly explore these new influence channels, the structural economics of the agentic transition matter more than any individual tactic.
When Does the Flip Happen?
The critical question isn't whether this transition will occur, but when it reaches critical mass. At what point do enough consumers delegate their shopping to agents that the economics of the human-facing web become unsustainable?
We have some historical precedent. The transition from phone orders to web orders took roughly 15 years to reach dominance. The shift from desktop to mobile commerce took about 10 years. If agents follow a similar adoption curve, we're looking at a transformation that plays out over the course of the 2030s.
But there's reason to think it could be faster. Once agents become reliable enough for routine purchases—and that threshold is probably around 95% success rates for simple transactions—the convenience factor becomes overwhelming. Why spend five minutes browsing Amazon yourself when you can text "order more coffee" and your agent handles everything, including comparing prices across retailers, checking reviews, and selecting the fastest delivery?
The feedback loop is brutal: as more users adopt agents, more services are forced to accommodate them. As more services accommodate agents, the user experience improves. As the experience improves, adoption accelerates. It's the same network effect that killed physical retail, now turned inward on the attention economy itself.
That said, some categories may resist this transformation longer than others. Luxury goods, experiential purchases, and products where brand identity matters deeply to consumers may maintain human-facing shopping experiences even as commodity categories move to agent-mediated commerce. The shift won't be uniform across all retail.
Amazon's Last Stand (And the Legal Reality)
Which brings us back to Amazon's lawsuit. The company isn't just protecting its terms of service; it's fighting for the survival of its business model. Amazon's retail operation is increasingly a loss leader for its real money makers: AWS, advertising, and Prime subscriptions. In 2024, Amazon's advertising revenue hit $60 billion annually, making it the third-largest ad platform after Google and Meta. That entire business evaporates in an agent-first economy.
Amazon can't afford to become a commodity database. It needs to control the relationship with customers—or in this case, the relationship with the agents acting on customers' behalf. The lawsuit against Perplexity is Amazon announcing that it won't voluntarily disintermediate itself, that agent access will happen on Amazon's terms or not at all.
But Amazon's legal strategy faces significant headwinds. The company is invoking the Computer Fraud and Abuse Act (CFAA), the federal anti-hacking law that has been used to prosecute everything from actual network intrusions to simple violations of terms of service. But recent Supreme Court precedent has significantly narrowed CFAA's scope.
In Van Buren v. United States (2021), the Supreme Court ruled that the CFAA only applies when someone accesses data they're not entitled to access at all—not when they access data in ways the owner doesn't like. The Court explicitly rejected the government's interpretation that would make CFAA violations out of terms of service breaches. Justice Barrett's majority opinion warned against reading CFAA so broadly that it would "criminalize everything from embellishing an online dating profile to using a pseudonym on Facebook."
Similarly, in hiQ Labs v. LinkedIn (2022), the Ninth Circuit held that scraping publicly accessible data doesn't violate the CFAA, even when explicitly prohibited by the website's terms of service. The court reasoned that the CFAA's prohibition on "unauthorized access" doesn't extend to information that's publicly available.
Perplexity's agents are accessing publicly available product listings on Amazon.com—the same data any human user can see. They're not circumventing password protections or accessing Amazon's internal systems. Under Van Buren and hiQ, Amazon's CFAA claim looks shaky. The fact that Perplexity is violating Amazon's terms of service may be a contract issue, but it's not obviously a federal crime.
What is certain is that this lawsuit won't be the last. Every major platform with a significant advertising business is watching Amazon's legal strategy closely. If Amazon somehow prevails, expect a wave of similar actions. If Perplexity wins, expect an arms race of technical obfuscation as platforms try to detect and block agents without explicitly violating whatever precedent gets set. We may see platforms implement "CAPTCHA walls" specifically designed to be unsolvable by agents, or rate limiting so aggressive that agent-based shopping becomes impractical.
What to Watch For
The agentic transformation won't happen overnight, but there are specific inflection points that will signal how this plays out:
1. The First Major Retailer to Implement Open Agent Standards
If a top-20 e-commerce platform implements VOIX or exposes structured <tool> elements for agent access, it signals that commodity-tier thinking is moving upstream. Watch for platforms losing market share to Amazon that see agent access as a competitive differentiator. Shopify might be the dark horse candidate—enabling agent commerce for their merchant base could be a powerful platform play.
2. The First AI Provider to Enforce an API Whitelist
Currently, agents can attempt to access any website. But what happens when OpenAI or Anthropic starts maintaining a whitelist of "approved" services their agents can interact with? This would effectively give them gatekeeper power over which businesses can participate in the agentic economy. Watch for partnerships announced as "exclusive integrations" or "preferred providers"—that's the whitelist being built in plain sight.
3. The First Major Brand to Drop Consumer-Facing Marketing
Some category-leading B2B company will eventually announce they're discontinuing traditional advertising in favor of "agent optimization." It won't be Amazon or Uber—it'll be an industrial supplier or SaaS platform serving an already commoditized market. That's your signal that agent-only business models are viable.
4. Legal Precedent in Amazon v. Perplexity
However this case resolves—settlement, summary judgment, or trial verdict—it will establish baseline rules. A Perplexity win means open season on adversarial agents. An Amazon win means platforms can legally block agent access. A settlement likely means bilateral deals become the norm. Watch for the revenue share terms: if Perplexity agrees to pay 15-20%, that becomes the industry standard.
5. The First Agent-Triggered Flash Crash
Thousands of agents simultaneously making purchase decisions based on the same information creates new systemic risks. We've seen algorithmic trading cause flash crashes in financial markets. The first time an agent-driven buying surge crashes an e-commerce site or causes a supply chain failure, it will trigger serious discussion about rate limiting, circuit breakers, and agent regulation.
The Internet We're About to Lose
There's a melancholy aspect to all of this that's easy to miss in the technical and economic analysis. The consumer internet we've known for the past 25 years—colorful, chaotic, attention-grabbing, sometimes manipulative but also sometimes delightful—is likely entering its final act.
Love it or hate it, that internet was human. It was designed for our eyes, our emotions, our impulsive clicking, our susceptibility to well-crafted copy and compelling visuals. It was optimized for engagement and conversion of actual human beings who could be surprised, delighted, or persuaded.
The agentic internet will be none of those things. It will be efficient, rational, transactional. It will be a machine economy, with machines negotiating with machines on behalf of humans who increasingly never see the actual transaction happening. It will probably be better at getting us what we actually need, but it will be entirely devoid of the serendipity and discovery that made the old internet interesting.
Whether that's progress or loss depends on how much you value convenience over experience. But either way, the transformation is coming, and it's a technical and economic shift already playing out in labs, courtrooms, and boardrooms across the industry. Like most technological disruptions, by the time everyone acknowledges it's happening, it will already be too late to stop.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.
