AI Search Is a Selection Layer

Losing in AI Search Isn't Dropping a Rank — It's Vanishing From the Answer

The definition the manual opens with governs everything after it: AI search is not a new traffic channel but a selection layer inserted between the user and the web. A traditional engine exposes candidates and hands the choosing to the user. An AI answer engine chooses first, then shows a synthesized answer built from a small set of retrieved sources.

The difference shows up in how failure feels.

The user never learns that the omitted sources existed — the choice set was compressed before comparison began. You didn't slip to page two; you were never there.

There's a second premise-breaker: the web page is no longer the unit of competition. AI answer engines retrieve passages, resolve entities, compare source confidence and synthesize. The model scores units smaller than the page and larger than the page at the same time. One good paragraph isn't enough — on a page the crawler can't reach, that paragraph doesn't exist either.

The Four-Layer Model — GEO Is a Question of Order, Not Tactics

The manual organizes GEO into four layers plus a multiplier. The ordering is the thesis.

The dependency runs one way. If layer 1 fails, nothing above it matters. If layer 2 fails, the engine arrives but doesn't know where to look. If layer 3 fails, it reads the page but misidentifies it. If layer 4 fails, it understands the page but finds nothing worth quoting.

Most sites we audit have a layer 1 or layer 3 problem. Most content investments go to layer 4. That's why most sites are not cited by AI engines — they're investing in the wrong layer.

— The GEO Readiness Manual, ch. 1

Judged by that diagnosis, "we're not cited by AI, so let's publish more articles" is the least efficient move available.

Design on the Assumption That AI Crawlers Don't Execute JavaScript

The manual calls Access the "zero layer": not the most sophisticated, but where the most damaging mistakes happen. Of the three free chapters, this is the one that lands hardest on Japanese sites.

Premise: Googlebot Is Effectively the Only Renderer You Can Count On

After surveying engine behavior as of July 2026, the manual lands on a practical conclusion: assume AI crawlers do not render JavaScript.

• —Googlebot: renders, though rendering can be delayed
• —Applebot: Apple documents that it may render — don't make it your only path
• —OAI-SearchBot: evidence is inconsistent; do not assume rendering
• —PerplexityBot / Claude bots: treat as basic HTML retrieval

In other words: if the body text isn't in the initial HTML the server returns, that text doesn't exist. The user sees the post-JavaScript screen; the crawler sees the first HTML response. Ranked by AI crawlability: SSG ≒ SSR ≒ ISR ＞＞ CSR (not crawlable). A CSR-only SPA is a blank page to AI search (see our measurements of how SPAs appear to crawlers).

403 Is the Silent Killer

The causes the manual lists:

• —Cloudflare bot-fighting modes (challenging or blocking non-browser user-agents)
• —WAF rules that block unknown user-agents
• —Bot detection that requires completing a JavaScript challenge
• —Rate limiting that returns 403 instead of 429
• —IP or geo-based blocking

Don't guess — measure. Change the user-agent and read the status code.

curl -s -o /dev/null -w "%{http_code}\n" -A "OAI-SearchBot"    https://example.com/
curl -s -o /dev/null -w "%{http_code}\n" -A "PerplexityBot"    https://example.com/
curl -s -o /dev/null -w "%{http_code}\n" -A "Claude-SearchBot" https://example.com/
curl -s -o /dev/null -w "%{http_code}\n" -A "Googlebot"        https://example.com/
curl -s -o /dev/null -w "%{http_code}\n" -A "Applebot"         https://example.com/

If anything other than 200 comes back, that engine cannot see you — regardless of what your robots.txt says.

Consent Banners and Paywalls

A consent banner should be a visual overlay, never a content gate. Serving a reduced page to crawlers that lack a consent cookie is, from the model's side, serving an empty page. The body must exist in the DOM even while the banner shows. The same applies to full-screen interstitials: in DOM order, main content first and interstitials last — bringing them to the front is CSS's job.

The Passages That Get Cited, and the Ones That Don't

The Quotability chapter is written as mechanics, not style advice. In the manual's framing, the model doesn't rewrite passages extensively — it prefers passages it can use with light editing. Format therefore decides selection.

The manual's own example makes it concrete. "The first issue is robots.txt. The second is JavaScript rendering. The third is schema." can't be extracted — each item needs its neighbors. But "robots.txt blocking GPTBot or PerplexityBot — often inherited from outdated anti-scraping rules; the site is invisible by default" stands alone, so it can be lifted.

Hedging and Attribution Are Not the Same Thing

The most valuable distinction in the chapter is this one.

Both look like ways of avoiding a flat assertion, yet their extractability is opposite. When you're uncertain, the manual says, write the second, not the first. That converges — from the mechanical side — with everything said about E-E-A-T.

The heading guidance is equally concrete: use question format, stay descriptive rather than clever, include the section's key term, never skip levels, and keep one H1 that matches the headline in your structured data. Clever headings engage humans and confuse semantic retrieval.

Applied to Japanese Sites, Three Things Come First

The manual assumes English-speaking sites. Reordered for what we actually see on Japanese sites — small and mid-size corporate sites, and sites built by web production shops — the priorities come out like this.

1. 1

   Fetch status codes as each AI crawler

   Highest return per minute spent, and it belongs ahead of reviewing robots.txt. Japanese sites frequently stack an imported theme, a security plugin and host-level bot protection, which produces the "allowed on paper, blocked in practice" state. Five minutes.

2. 2

   Confirm the body text is in the initial HTML on key pages

   Pipe curl into grep for your h1 or lead sentence. With headless CMSes and page builders now common, a page that looks perfect while its raw HTML is empty is not unusual.

3. 3

   Rewrite the first sentence of existing articles before publishing new ones

   Better return than a new post. Most Japanese articles open with "In recent years, as AI technology has advanced…". That paragraph is crawled, indexed and embedded — and never retrieved, because it sits far from every useful query. Delete it and open with the answer.

Conversely, there's no reason to start with llms.txt. The manual states plainly that the evidence for llms.txt is weaker than the evidence for access, schema and content structure, and ranks it fourth in its own layer. Create it because it takes thirty minutes — just don't reorder the queue around it. That matches our own position.

Don't Turn the Manual's Numbers Into Your Own Thresholds

The manual is unusually honest about evidence. Chapter 1 declares five tiers — official documentation, private benchmark observations, public reports, manual prompt testing, and inference from RAG mechanics — and commits to writing facts as facts, observations as observations, and black-box behavior as probability. It states that GeoReady measures readiness rather than guaranteeing placement, and says outright that a vendor promising guaranteed citations is selling certainty the market does not provide.

With that said, a note worth making: some specific numbers in the manual arrive without a source.

• "Target TTFB under 200ms" — Google's general guidance is under 800ms
• "Passages of 40–120 words", "chunks of 200–500 tokens" — plausible as tendencies, but unsourced
• "headline under 110 characters" — this traces to the old AMP requirement; current Google documentation states no limit

They're useful as heuristics and unusable as pass/fail criteria. We align our own audit thresholds with Google's documentation because optimizing a third-party score doesn't move the search reality. The value of this manual is not in its numbers — it's in the ordering of the layers and the explanation of the mechanics.