Architecture · Agent-facing surface

The channel that converts best cannot read your site.

Adobe measures AI-referred shoppers converting better than every other channel, and up to 46% of retailer sites that machines cannot read. Cloudflare measures more than half of all web requests coming from automated agents. The public surface became an interface. Nobody on the org chart owns it.

Field briefing · HB-2026-0016 The Hikari Blue team · Operator notes from Austin 5 min read

More than half of the requests hitting the web are now machines, not people. Most enterprise pages are still written for the other half.

The prevailing read treats AI assistants as a new marketing channel. Get cited, win the referral, count the traffic. That read stops one layer too early. An assistant does not browse a page. It fetches it, parses it, and restates it. What it cannot parse, it omits or fills in from somewhere else.

What the infrastructure just measured

Cloudflare put numbers on the shift on July 1. Automated agents and bots now drive more than half of all web requests. Search engines reach roughly twice as much information as the leading AI companies do (Cloudflare, July 2026). That gap is not appetite. It is access and readability.

The same announcement moves the default. From September 15, new domains on Cloudflare allow search crawlers but block training and agent crawlers on pages that carry ads. Crawlers that cannot separate those uses are blocked outright on those pages. Cloudflare is also turning Pay Per Crawl into Pay Per Use, paying publishers when their content appears in an AI answer. The immediate scope is narrow. The direction is not. Agent access to your content is becoming a priced, revocable setting held at the edge.

The channel that converts is the one you cannot serve

Adobe Analytics measures more than a trillion visits to US retail sites. Around Prime Day, AI-referred traffic grew 98.3 percent year over year. It converted 50.7 percent better than non-AI channels, with a 33 percent higher add-to-cart rate (Adobe Analytics, June 2026). The agent arrives having already done the qualification work. Then it meets the page. Adobe puts the share of retailer sites that machines cannot read at up to 46 percent in some cases.

That is not a visibility problem. It is a loss at the bottom of the funnel, and the causes are mechanical. Product data rendered by JavaScript the crawler does not execute. Price and availability behind a call it cannot make. Terms and conditions inside a PDF.

The public surface of the enterprise stopped being a marketing asset. It became an interface, and it now serves machines first. No one on the org chart owns it. Hikari Blue · operator note

In a regulated firm, the exposure is different

What the machine restates about your product is what the customer acts on. The CFPB set the principle out in 2023: when a firm's chatbot gives a customer inaccurate information, the standard the firm is held to does not move (CFPB, June 2023). That finding covered a firm's own chatbot. The exposure now runs through assistants the firm does not control, reading a surface the firm does publish.

Read the mechanics for a bank, an insurer, a pharmaceutical company. If the rate table, the exclusion, or the contraindication is unreadable, the assistant answers from an older or third-party copy. The customer hears it under your name. Whether that becomes your legal liability is unsettled, and we flag it as our reading rather than settled law. Whether it is already your operating problem is not in question.

What the leaders publish on purpose

Autodesk shipped a read-only Product Help MCP server in April. It exposes help content across more than 110 products to agents running outside Autodesk software. The stated reason is plain: an agent paired with real domain knowledge does less guessing (Autodesk, April 2026). Adobe shipped the same pattern for commerce, giving agents live access to catalog, price, and inventory.

The split between leaders and laggards is clean. Laggards tune pages for humans and hope the crawler copes. Leaders publish a machine-facing surface with an owner, a version, and a record of what was served. One is a marketing tactic. The other is architecture.

What stops it

Ownership. The agent-facing surface is cut three ways today. Marketing owns the content. Infrastructure owns the crawler policy at the CDN. Legal owns what may be said. None of the three owns the outcome, and none is measured on it.

Two numbers make it measurable. The share of revenue-bearing pages an agent can fully parse. The share of AI-referred sessions that end in a completed action. Both are cheap to instrument and neither sits in a marketing dashboard today.

The question to bring to the next operating review

Do not ask how the brand ranks in AI answers.

Ask who owns what the machine reads, and what it tells your customer when the page does not answer.

A setting nobody owns gets decided by whoever ships the default. On September 15, someone else ships it.

The Hikari Blue team · Austin, July 2026

More from the Newsroom

See all articles in the Newsroom →

Map your exposure before the board does.

Thirty minutes with an operator. No slides.

Direct call with one of the partners. We listen, we structure, and we tell you what your agent-facing surface would look like (machine-readable product and policy data, crawler posture, ownership, record of what was served) for your specific stack and regulatory perimeter.

Book a 30-minute working session