Build for Agents, Not Just Humans — The Operator Playbook Google Didn’t Publish

The Quiet Announcement Most Operators Missed

On May 1, Google’s web.dev published a guide titled “Build agent-friendly websites.” No keynote. No I/O moment. A single blog post on a developer documentation site.

The message inside is the most consequential thing Google has said about the open web in 2026: websites must now be designed for two audiences — humans and AI agents. Not as a thought experiment. As practical engineering guidance, on Google’s own developer site, telling builders to stop assuming a human visitor.

I’ve been writing about this shift since 2023, when I published an 88-slide deck titled AI’s Imminent Impact on SEO — before GEO and AEO existed as categories. The agentic shift is no longer a thesis. It’s now Google’s developer documentation.

But the guide stops at the floor. The interesting work is everything on top of it.

Google’s web.dev guide is HTML hygiene for agents. It is necessary. It is nowhere near sufficient.

The Gap — What Google Didn’t Say

Google’s guide answers one question: how do agents see my site?

It does not answer the questions that determine whether you actually get cited:

• How do agents decide which site to cite when six options are equally readable?
• How do they extract a chunk worth quoting?
• How do they distinguish my brand from a competitor’s noise?
• How do they verify my claims when sources disagree?
• Why do they cite some sites 53% of the time and identical-looking competitors 0%?

Those questions are answered in content architecture, schema integrity, brand citability, and citation graph density — none of which are in Google’s guide.

This is where the operator playbook starts. Google handed builders HTML hygiene. The work that actually changes citation rate is layered on top of that hygiene, not contained within it.

In the 14 months I’ve spent rebuilding indexableai.com from the ground up as an agent-native site, and in the audits we’ve run on $1B+ brand portfolios, the pattern is consistent: brands that nail Google’s web.dev floor still lose AI search if they skip the seven layers above it.

Layer-by-Layer Walkthrough

Layer 1 — JSON-LD Schema Is Your Agent Contract

Google’s web.dev guide skips schema entirely. That is the most important omission. JSON-LD is the machine-readable contract between your brand and a language model. It tells an agent who you are, what you sell, who works there, what claims you’ve made, where you operate.

When an agent must choose between citing a page with rich Organization + Product + Article + FAQPage schema versus a page with no schema at all, the structured page wins by default. We inject seven JSON-LD types site-wide on indexableai.com: Organization, WebSite, Service, Article, BreadcrumbList, Person, and FAQPage. It is not a nice-to-have. It is the meta-text agents read first.

Layer 2 — Stable URLs Are Stable Identity

Agents do not follow redirect chains the way humans do. A URL that 301s through three hops is a URL the agent treats as unstable — and unstable URLs get downranked in citation. Canonicalize aggressively. Pick the URL you want cited in AI search and make every alias collapse into it. Treat URLs as identity, not as routing.

Layer 3 — Zero JavaScript on Critical Paths

This is the hardest layer for most teams to swallow. Agents can execute JavaScript — but most don’t reliably, and the ones that do penalize JS-rendered content for latency and inconsistency. The open agent-readiness scanners on the market today (Cloudflare’s public scanner is the cleanest baseline) score JS-heavy pages in the 20s out of 100. A pure HTML page scores in the 90s.

We removed all JavaScript from /pricing, /how-it-works, /about, and /why. The score went from a starting baseline to 100/100. Vanilla HTML is an unfair advantage in 2026.

Layer 4 — Codename-Led Content. Facts, Not Adjectives.

Agents extract claims, not vibes. Every product page, agent page, and pricing page on indexableai.com leads with a codename and a measurable claim. Our SEO Manager agent, our Content Engineer, our Technical SEO — each has a codename, a concrete responsibility, and a numeric output. Agents quote facts. They don’t quote “industry-leading” or “next-generation.” Strip the adjectives. Lead with the codename and the numeric claim.

Layer 5 — Citation Density and Chunk Shape

A 42,971-citation analysis (Shashko, 2025) found that the median AI citation is 10 words long, and that structured content (lists, tables, definitions) is cited at a 91.3% advantage over flowing prose. Top 35% of citations originate from content positioned in the first third of the page.

The implication is direct: agents extract chunks, not articles. Write each page as a sequence of self-contained, copy-pasteable chunks. The first sentence of every section should be a complete claim, readable out of context. Lead with the chunk, then explain.

Layer 6 — Citation Graph as Authority Moat

This is the hidden layer. Google’s web.dev guide treats every site as if it stands alone. It does not. Agents weight citations from other authoritative sites pointing back at you — precisely like Google’s PageRank, but applied to the citation graph instead of the link graph.

A new domain can earn agent trust by being cited by 27 internal pages on its own site, plus a smaller set of trusted external sources. Three days ago, we shipped 27 internal inbound links to our flagship article on indexableai.com. That is not vanity SEO. It is citation graph engineering.

Layer 7 — WebMCP Isn’t the Destination. It’s Table Stakes.

Google’s preview of WebMCP is real and important, but it solves agent action — letting agents call your tools as functions. It does not solve agent citation — getting your brand mentioned in the first place.

WebMCP is the steering wheel. Layers 1 through 6 are the engine. Brands that ship WebMCP without the engine are giving agents a steering wheel attached to nothing. Sign up for the preview — but do not let it become a distraction from the foundational work.

The 14-Day Agent-Readiness Sprint

You don’t need to overhaul your stack. You need a 14-day sprint with seven owners and one objective: score 90+ on the open agent-readiness scanner before Google I/O on May 19-20.

Days 1-2 — Run the scanner and get your baseline

Most enterprise sites we audit start in the 20s out of 100. The number is shocking the first time. Document it. The baseline becomes the proof of progress.

Days 3-5 — Strip JavaScript from your top 5 highest-traffic landing pages

The hardest step. The biggest score lift. Replace JS-rendered content with static HTML. Defer JS to non-critical interactions only. Most sites move from 25 to 60 on this step alone.

Days 6-8 — Inject Organization + Service + Article JSON-LD site-wide

One schema file, included in the site-wide <head>. Validate every type with Google’s Rich Results Test before deploy. Score climbs from 60 to 80.

Days 9-10 — Audit URL structure

Collapse all redirect chains. Pick a canonical for every duplicate. Redirect everything to the chosen canonical with a single 301 hop. No chains.

Days 11-12 — Rewrite top 5 page H1s and first paragraphs

Codename-led. Fact-led. Chunk-shaped. Every section opens with a self-contained claim. Adjectives go in the marketing department. Codenames and metrics live on the page.

Days 13-14 — Inject internal inbound links from existing high-traffic pages to flagship pages

Citation graph in motion. Identify your top 20 highest-traffic existing pages, and add one contextual link from each into your most important new flagship article. Score lands at 90+.

By end of Day 14, you should be scoring 90+ on the open agent-readiness scanner. By Day 21, you should see citation rate climb in your AI visibility tracker for category prompts.