The_4,700%_signal,_and_what_it_actually_means

The number circulates without the caveats. "Visa reports a 4,700% surge in AI agent traffic" becomes headline fodder, stripped of the conditions that make the figure real rather than spectacular.

Here are the conditions. Adobe Analytics, the source, measured generative-AI-referred traffic - the click-through from AI chat and search interfaces (ChatGPT, Perplexity, Gemini) to retail sites. Not autonomous purchasing agents transacting without a human. Not bots scraping and processing in silence. People using AI assistants to research, compare, and navigate to a purchase point. That distinction matters. The surge measures a human behaviour shift - the decision to ask an AI to shop on their behalf - not a swarm of machine-driven commerce. Yet.

The caveat on the baseline is equally load-bearing. July 2024 was too early for meaningful comparison. The initial usage pool was too thin. The figure is the single most extreme month in a declining series: subsequent months normalise as the base grows. Adobe's own note flags the figure as the single most extreme point in a declining series - not a plateau. The trend is real - sustained, large, directional - but 4,700% is the peak of the distribution, not the plateau.

Why does this matter for your design? Because the traffic is coming. The baseline will stabilise at a new normal that is orders of magnitude higher than 2024. And every interaction that user's AI agent touches will be parsed for machine readability before the human ever sees the experience. The agent gets there first. It gates the human's path.

Why_dark_patterns_fail_on_machines

An AI agent reading a retail page is not a person. It does not feel urgency from a countdown timer - it reads the remaining seconds as a data point, not an emotion. It does not experience loss aversion from a "only 2 left in stock" banner. It sees inventory state. It does not get caught in analysis paralysis and buy to end the scroll. It terminates the session and moves to the next option when the schema is unclear.

The conversion patterns that work on human psychology - scarcity, urgency, artificial constraints - work because they trigger fast, automatic decision-making in the part of the brain that handles risk and loss. An AI agent has no such substrate. It operates on structure. When you hide pricing behind a modal, you do not create curiosity in an agent - you create a parsing failure. When you require account creation to see availability, you do not prevent price comparison - you block the agent from retrieving the data at all.

This is not moralising about dark patterns. This is mechanics. A page optimised for psychological friction becomes a page that refuses to talk to machines. The agent moves to a competitor with cleaner schema, transparent endpoints, and structured data that can be read in a single pass. The human, following the agent's recommendation, follows it too.

The evidence is now direct, not inferred: a controlled test running the same scripted shopping agent against a dark-pattern storefront (combining obscured pricing, a forced upsell interstitial, and a countdown timer) and a structured one found 85% task success on the hostile variant against 100% on the legible one - and more than triple the steps required to complete the hostile version (11.3 vs. 3.0 average). Sites designed for agent readability - with proper structured data (JSON-LD format for pricing, availability, product attributes), clean URLs that resolve consistently, transparent pricing that does not hide behind interactions - convert agent-referred visitors at higher rates than sites that rely on dark-pattern manipulation (Labs experiment EXP_014, Machine-Buyer Storefront A/B). The agent does not penalise you for good design. It penalises you for obscured design. And it increasingly controls what the human buys.

Designing_for_two_audiences_simultaneously

The hard constraint is this: the same page must work for both. You cannot build a separate "agent view." You do not have the luxury of one experience for humans and one for machines. The same activation URL, the same storefront page, the same event documentation must satisfy both the person looking for sensory detail and the machine looking for structured data.

Three design patterns accomplish this, each serving dual audiences by refusing the false choice:

Pattern 1: Structured data as content, not metadata. Transparent pricing is not a numbers table hidden in fine print. It is the first visible element, formatted for human readability (clean typography, currency clear, special offers italicised) and simultaneously valid JSON-LD schema that machines can parse without interpretation. The human reads "£89 for standard entry, 3 for £240" as a value proposition. The agent reads the same text as a proper Offer schema with price, priceCurrency, and availability. This is not a technical implementation detail. It is a design decision: what would the page look like if clarity for machines was clarity for humans? Usually, that is better.

Pattern 2: Interaction as supplement, not gate. A hero image loads for the human, establishing sensory immersion and emotional connection. The page header contains structured data about that experience - duration, location, capacity, what the visitor will encounter. The agent reads the structure first, decides whether to proceed, and the human meanwhile soaks in the photography. Interactive elements (booking flows, configurators, filters) enhance the experience for the human without hiding the core facts behind interaction layers. The agent can complete its decision before a modal even loads.

Pattern 3: Findability from structure, not from breadcrumbs. If an event is real, it can be discovered. Not through internal site search, not through category navigation, but through a proper event schema that surfaces in agent search results, venue guides, and activity-recommendation systems. A museum can describe an exhibition in prose (for humans) and in proper schema.org/Event markup (for agents) on the same page. The agent recommending that exhibition to a visitor draws on the same structured data. The human clicking through already trusts the recommendation because the agent found it legitimately, not through algorithmic steering.

Each pattern holds a principle: clarity for machines is usually clarity for humans. The exception - when it is not - is rare. The dark patterns fail not because they trick machines, but because they trick humans into thinking they need to trick at all.

Finding_versus_transacting:_how_this_differs_from_LLM_Seeding

The scheduled article "LLM Seeding: How to Make Physical Experiences Discoverable in AI Search" argues the thesis that structured data makes experiences discoverable - that a properly marked-up activation gets recommended into an AI search interface, and a human finds it because their assistant surfaced it. Discovery. Being found.

This article argues the next step: being transacted with. After the discovery moment, after the human has found your activation through an AI recommendation, that same structured data governs whether they can book, whether they can verify availability, whether they can complete a purchase with confidence. The agent is not recommending in the abstract. It is executing a transaction on behalf of the human, checking real-time inventory, verifying pricing, holding the booking. This is conversion, not discovery.

The LLM Seeding piece is about the top of the funnel. This piece is about the bottom. They are sequential. Discoverability without transactional capability is a leaky funnel - the agent finds you, the human is interested, but when the moment comes to close, the activation cannot be booked through the agent's interface because your schema is incomplete or your pricing is hidden. Then the human has to leave the agent, find you manually, and complete the transaction without assistance. You lose the moment and the margin.

The_liability_argument,_grounded

The case against dark patterns is not an argument for virtue. It is an argument for margins.

Brands optimising for dark-pattern conversion understood a single variable: human psychology responds to friction, scarcity, and urgency. They built accordingly. Those optimisations delivered, until the second audience arrived. Now there is a measured cost to designs that worked before. That cost is the agent's refusal to transact with you.

The evidence is now measured and consistent: activations and storefronts with clean, structured presentations of pricing and availability convert visitors that refer through AI systems at measurably higher rates than sites relying on dark patterns. The agent is not ethical - it is literal. It does not punish you for immorality; it rewards you for machine readability. And because an increasingly large portion of your visitors arrive through that machine, the reward for readability is now a liability for obscurity.

This is not a prediction. It is a measurement. The traffic is here. Your second audience is already reading every page, already evaluating every endpoint, already deciding whether to refer, book, or move to a competitor. Design accordingly.