5層GEOファネルは、AIの可視性を下から順に診断します—技術的アクセス、引用、Share of Model、リファラル、収益—これにより、実際に問題のある層を修正し、上位層に過剰投資することを防ぎます。
2026/07/05
8 分で読む
更新日 2026/07/13
The 5-layer GEO funnel is a maturity model for AI visibility that reads from the bottom up: technical foundation, content and citation, AI visibility, traffic and conversion, and business ROI. Each layer answers one blunt question about whether generative engines can reach your content, trust it, surface it, refer people from it, and finally turn it into revenue. Unlike a marketing funnel built to capture clicks, this one is built for a world where ChatGPT, Perplexity, and Gemini answer users directly and the click often never happens — so it measures depth of AI understanding rather than sessions alone. Its real value is diagnostic: it shows you which layer is actually broken before you spend on the one above it.
Key takeaways
The funnel is dependency-ordered from the bottom: infrastructure gates citation, citation gates visibility, visibility gates referral, and referral gates revenue — so audit bottom-up, not top-down.
Layer 1 is pure machine access: an llms.txt file, JSON-LD schema, deliberate crawler rules for GPTBot, ClaudeBot, and Google-Extended, plus IndexNow for instant re-indexing.
Layer 3 is where brands actually compete. Share of Model, AIGVR, sentiment, and generative position decide whether an engine recommends you or a rival — it is the AI-era equivalent of share of voice.
Because most AI answers are zero-click, Layers 4 and 5 track assisted conversions, product-card activation, and model-attributed revenue rather than raw referral traffic.
A single weak layer caps everything above it: perfect content earns no citations if crawlers are blocked, and high visibility earns no revenue if the product card never activates.
Read the funnel bottom to top
Most teams inherit the funnel from advertising, where you pour awareness in the top and squeeze conversions out the bottom. The GEO funnel inverts the intuition. You build from the foundation up, and you diagnose from the foundation up, because a weak lower layer silently caps every layer above it. The order:
Technical foundation (infrastructure) — can AI agents reach and parse your content?
Content and citation (trust) — is your content worth quoting as a source?
AI visibility (strategy) — how much of the AI conversation do you own?
Traffic and conversion (connection) — does that visibility move real people to act?
Business ROI (commercial value) — what is it worth in revenue?
If you're new to the discipline, start with what GEO is and before you wire up the layers below.
If a crawler can't reach your pages, or can't read their structure, nothing above this layer exists. This is the plumbing, and in 2026 it has four moving parts.
llms.txt: a plain-text file at your root that points models to your highest-value content in a clean, token-efficient form, the way robots.txt guides traditional crawlers. See the llms.txt proposal.
Structured data: JSON-LD schema for Product, FAQ, HowTo, and Organization gives models the context behind your words, not just the words. It is the difference between "a page mentioning a price" and "a product priced at $49."
Crawler access: robots.txt rules that deliberately allow or block agents such as GPTBot, ClaudeBot, PerplexityBot, and Google-Extended. Blocking them keeps you out of both training and retrieval; allowing them is a prerequisite for everything else. OpenAI documents its crawler user agents openly.
IndexNow: instead of waiting days for a recrawl, IndexNow pings participating engines the moment content changes, which matters when your pricing or availability shifts.
Get this wrong and your brand is invisible to the retrieval systems that assemble AI answers, no matter how good your content is.
Layer 2 — Content and citation
Does AI trust my content enough to cite it?
Indexing gets you into the library. Citation gets you quoted. This is the layer most brands skip, and it's where authority is won. Track four things:
Total citations: how often your pages are named as a source in AI answers.
Content extraction rate: the share of your content models actually lift into their responses.
Grounding queries: the hidden searches an engine runs to verify a claim. Align your pages with those questions and you become the source it lands on.
Citation drift: how often you get swapped for a competitor in the same prompt over time — a stability signal, not a vanity count.
Citation source analysis: source type distribution and the domains AI engines cite most — Source: GEOly AI (app.geoly.ai)
The work here is unglamorous: original data, expert quotes, dense and clearly structured pages that are easy to extract. Generic blog filler doesn't get cited. A structured citation analysis tells you which domains an engine already trusts in your category so you can earn placement on them.
Layer 3 — AI visibility
How much of the AI conversation do I own?
This is the strategic core, the equivalent of share of voice in traditional media. Four metrics matter:
AIGVR: a 0–100 visibility score across thousands of category-relevant prompts — the closest thing GEO has to a ranking.
Share of Model: your slice of brand mentions versus competitors inside a specific engine (owning 40% of "best CRM" answers in ChatGPT, say).
Sentiment: not just whether you're mentioned, but whether the engine recommends you, stays neutral, or steers users away.
Generative position: when an AI lists "top five" options, being first instead of fifth changes everything downstream.
AI search visibility dashboard tracking mention rate, AIGVR and Share of Model across ChatGPT, Gemini and other AI engines — Source: GEOly AI (app.geoly.ai)
Platforms like GEOly AI measure these across seven engines — ChatGPT, Gemini, Perplexity, Copilot, Grok, Google AI Mode, and Google AI Overview — so you can see the same prompt resolve differently in each. Compare tools first in this roundup of AI SEO platforms.
Layer 4 — Traffic and conversion
Is AI visibility driving real traffic and action?
Most AI interactions are zero-click: the user gets an answer and never visits your site. That doesn't mean visibility is worthless — it means you measure it differently. Watch AI referral sessions from engines that do send traffic, assisted conversions where an AI answer touched the journey earlier, and branded-search lift after a wave of AI recommendations. For commerce brands, the signal moves inside the engine: product-card activation and agentic checkout, where the model surfaces your product and the transaction begins without a homepage visit at all. Share of card is the emerging metric for that shelf.
Layer 5 — Business ROI
What is this worth to the business?
The top of the funnel is money. Model-attributed revenue ties AI-sourced sessions and assisted conversions back to a number your CFO recognizes. Layer it with acquisition cost for AI-sourced customers and the lifetime value of those customers, and GEO stops being a brand-awareness line item and becomes a channel with a return you can defend in a budget review.
Putting the funnel to work
Run it as a diagnostic, bottom-up. Start with a GEO audit — GEOly's is a 29-point check — to find your lowest broken layer. Fix that first: there's no return on Share of Model work while GPTBot is blocked at Layer 1. Then climb. For teams that want this automated, GEOly exposes the same data through an MCP server, CLI, and Skills, so an agent can pull your AIGVR, citations, and Share of Model on a schedule. You can try it on a free three-day trial at app.geoly.ai or compare plans on pricing.
How is the GEO funnel different from a marketing funnel?
A marketing funnel tracks a human moving from awareness to purchase across your owned channels. The GEO funnel tracks whether a machine can reach, trust, and surface you inside someone else's interface, then whether that produces action and revenue. The biggest shift is that the middle of the journey — discovery and recommendation — now happens inside the model, often with no click, so you measure Share of Model and citations instead of impressions and CTR.
Which layer should I start with?
The lowest one that's broken. Crawler access and llms.txt at Layer 1 are cheap and gate everything above them, so confirm those first. Once agents can reach you, most of the compounding work sits at Layers 2 and 3 — earning citations and growing Share of Model. Chasing ROI attribution at Layer 5 before those are healthy just measures a problem you haven't fixed.
Do I still need llms.txt if I already have a sitemap and schema?
They solve different problems. A sitemap lists URLs and schema describes a page's meaning, but llms.txt curates a short, token-efficient path to your best content specifically for language models. It's not universally honored yet in 2026, but it's low-cost to publish and increasingly read, so it belongs in Layer 1 alongside the others rather than replacing them.
How do I measure Layer 3 when AI answers are personalized?
You sample. Instead of trusting a single prompt, tools run each category prompt many times across engines and accounts to produce a distribution, then report AIGVR and Share of Model as rates rather than one-off screenshots. That sampling is what turns a personalized, non-deterministic answer into a metric you can trend week over week. See how AI visibility metrics are defined.
