Designing for AI Agent Surfaces

My site currently serves about 13,100 requests a day to AI agents (a seven-day average) and about 325 daily human visits (a 28-day average), straight from my own edge logs – roughly a 40-to-1 ratio, most of it live retrieval: ChatGPT-User fetching a page right now because a person just asked their assistant something. That reader ignores every design decision I was trained to make. It sees no typography, no spacing scale, no motion. And yet it is not indifferent to design – it is ruthlessly sensitive to a different set of design decisions, ones most craft discussion never mentions. This post is the craft guide for the surface I actually serve most.

TL;DR

• On this site, the agent surface is the majority surface – 40 to 1. The volume is dominated by user-directed live retrieval: ChatGPT-User made ~6,700 requests yesterday against 4 from GPTBot, OpenAI’s training crawler. If your site serves developers or reference content, your logs likely lean the same way – but measure before you assume.
• Agents read different pages than humans click. My edge logs show comparison posts and niche technical pages drawing heavy agent traffic while sitting far below my human top 15 – the assistant reads them for the human who never arrives.
• Design for extraction, not immersion: a one-paragraph answer at the top, question-shaped headings, self-contained sections, tables for anything enumerable, and dates on every claim.
• The visual layer and the agent layer are not rivals. The same structure that machines extract cleanly is what humans skim: answer-first is good design for both readers.
• Verify what agents actually consume – serve clean markup, keep content out of client-side rendering, and watch your logs, because the agent audience is measurable in a way “brand impression” never was.

The Reader That Renders Nothing

Design craft assumes an eye. Typography, spacing, motion – the whole discipline steers attention through a visual field. The agent reader has no visual field. It gets your HTML as text, spends a strict token budget on it, extracts what answers the user’s question, and quotes or paraphrases you to someone who may never see your page.

That is not a degraded reading – it is a different medium, with its own craft variables:

• Position replaces visual hierarchy. In single-fetch pipelines, what appears early is what gets read before token budgets truncate; chunked retrieval can rescue a deep section that matches the query, but you do not choose which pipeline reads you, so a conclusion buried at 80% depth exists only sometimes.
• Structure replaces layout. Headings, lists, and tables are parsed as semantics, not styled as decoration. A of versions and dates is machine-legible in a way the same facts scattered through paragraphs are not.
• Self-containment replaces flow. Human prose builds across sections; extraction takes sections in isolation. A paragraph that begins “as mentioned above” exports as a dangling reference.
• Explicitness replaces context. “Last updated August 2026” travels with a quoted claim. Freshness implied by the surrounding page does not.

What the Agents Actually Read

The part of this I could not have guessed from theory: agents have their own taste in pages. My last seven days of edge logs, top pages by agent requests:

The reference guides lead – expected, and they top the human leaderboard too. The revealing rows are the comparisons: pages far down my human leaderboard that assistants pull constantly, because “which should I use” is what people ask assistants. The assistant reads the comparison for them and answers in chat. The human never visits; the content still did its work.

The composition matters too. Just under half of yesterday’s agent volume was ChatGPT-User alone – 6,718 requests, against 4 from GPTBot, OpenAI’s training crawler. Training crawls are not absent (Bytespider and ClaudeBot together ran about 28% of the day), but the single largest reader by far is not a scraper stocking a corpus. It is a person’s assistant, fetching a page because someone just asked.

The Craft Rules

What I now hold agent-facing pages to – which is to say, all of them:

1. Answer first, at full strength. The opening paragraph states the complete answer a reader came for – not a teaser, the answer. On this site it is a styled answer block; structurally it is the abstract that extraction quotes. If an assistant reads only your first 200 words, they should leave with your conclusion, not your throat-clearing.

2. Headings are questions or claims, not labels. “How the budget gets truncated” beats “Background.” Question-shaped headings match the query the assistant is answering, and each section under one should resolve it without leaning on its neighbors. (Reference pages get the strict version of this rule; essays – this one included – trade some of it for voice, a deliberate choice about audience rather than an oversight.)

3. Enumerable facts go in tables. Versions, dates, compatibility, prices, comparisons – anything with rows. A table survives extraction intact; the same facts as prose come back paraphrased, and paraphrase is where errors enter.

4. Date the perishable. Any claim that can rot carries its date inline – “as of August 2026” – because quoted text travels without your page’s context. The alternative is your two-year-old numbers presented by an assistant as current, under your name.

5. Keep the substance out of client-side rendering. Most retrieval agents read served HTML. Content that materializes via JavaScript is invisible to some of your largest readers – check what curl sees, because that is closer to the agent’s view than your browser is.

6. Let structure serve both readers. None of this fights visual design. Answer-first is how busy humans skim; question headings are scannable; tables beat prose for humans too. The disciplines converge, which is the reassuring part: the craft that serves extraction is mostly the craft that serves respect for the reader’s time.

Measuring a Readership You Cannot See

The strange gift of the agent audience is that it is more measurable than the human one, not less. Every retrieval is a logged request with a user agent. I track three things: total agent volume by family (who reads me), the agent-read leaderboard against the human one (what they read that humans do not), and the ratio’s trend. Those numbers changed my content strategy more than any pageview chart – they are why comparison pages and reference structure get investment here that “time on page” would never justify.

The uncomfortable corollary: none of the agent readership shows up in analytics built for humans. If your strategy runs on sessions and bounce rate, the 40:1 surface is invisible, and you are optimizing the minority channel with the majority’s budget.

FAQ

What is an AI agent surface?

The version of your site consumed by AI assistants and their retrieval agents – ChatGPT, Claude, Perplexity and their fetchers – which read your served HTML as structured text, extract answers under token budgets, and relay them to users who often never visit. On measured sites this surface can dwarf human traffic; mine runs roughly 40 requests to agents per human visit.

How do you design content for AI agents?

Answer-first openings that state the full conclusion, question-shaped headings with self-contained sections, tables for any enumerable facts, inline dates on perishable claims, and substance in served HTML rather than client-side rendering. The same structure serves human skimmers – the two audiences reward the same discipline.

Do AI agents read different pages than human visitors?

Yes, measurably. In my edge logs, reference guides and comparison pages draw thousands of agent requests weekly while some barely register human visits – assistants pull “X vs Y” content to answer decision questions in chat, so the page does its work without a visit. Comparing your agent-read leaderboard to your human leaderboard shows where the two audiences diverge.

Does designing for agents hurt the human experience?

No – the requirements converge. Answer-first structure, honest headings, tables, and dated claims are what respectful design for busy humans already looks like. The visual layer (typography, spacing, motion) remains fully human territory; the agent surface just makes the structural layer’s quality measurable.