A law firm partner receives a call. The caller says: “My AI assistant shortlisted three firms. You were on the list. I’ve already read about your background, your specialism, and the kinds of cases you typically handle. I just need to confirm you’re taking new clients before I book.”
That call happened because the firm was on an AI’s consideration set. The partner didn’t win it with a clever headline, a Google ranking, or a well-designed contact form. They won it because an AI system, evaluating options on a user’s behalf, had enough independently verifiable information about the firm to include it in a recommendation.
The two firms that didn’t get called weren’t invisible on Google. They were ineligible for the shortlist. Different problem. Different solution.
This is the shift that matters. Not SEO versus AI. Not Google versus ChatGPT. Ranked versus recommended. And the architecture that determines which side of that line you’re on is being built right now.
What four separate research pieces are telling you
Four significant pieces of analysis have landed in recent months. Each one valuable in isolation. The real insight is the pattern they form together.
Duane Forrester at SEJ — the person who built Bing Webmaster Tools and launched Schema.org at Microsoft — describes a four-layer machine-readable brand architecture. JSON-LD as a fact layer. Entity relationship mapping that expresses how things connect. Content API endpoints for programmatic access. Provenance metadata as the tiebreaker when retrieval systems evaluate conflicting claims. Pages with valid structured data are 2.3× more likely to appear in Google AI Overviews (Forrester, SEJ, 2026).
Chris Green’s Web Almanac analysis gives us the most honest sentence in any of this research: “The web is really messy. Really messy.” LLMs.txt adoption is at 2% of the HTTP Archive dataset, largely tool-driven. SE Ranking’s analysis of 300,000 domains found no statistical correlation between llms.txt presence and AI citation frequency — removing it from the predictive model actually improved accuracy. This doesn’t mean llms.txt is worthless. It means it is what it always was: a clean, low-noise path for AI crawlers. A sitemap for the machine-readable era. Not a citation strategy.
Slobodan Manic places the December 9th 2025 Linux Foundation announcement correctly: the Agentic AI Foundation, with AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI building shared infrastructure rather than competing standards. MCP from Anthropic now has 97 million monthly SDK downloads and over 10,000 published servers. This is the W3C moment for agentic AI — the protocols being established now will define how AI interacts with the web for the next decade.
Purna Virji at SEJ argues that trust is the new ranking factor. She’s right about the phenomenon. What she’s describing as new has actually been encoded in Google’s quality guidelines since 2014 as Experience, Expertise, Authoritativeness, and Trustworthiness. What is genuinely new is the consequence structure. Google evaluated trust at crawl time. The effect was graduated: rank 4th, still get clicks. AI agents evaluate trust at query time. The effect is binary: on the consideration set or not, and that determination happens before any human visits your site.
Rand Fishkin’s SparkToro research captures this precisely: ask AI systems for brand recommendations repeatedly and you get significant variance — different brands, different orders, different list lengths. But inside that noise, a stable core consideration set. The same handful of businesses appearing consistently. Those are the businesses the AI evaluates as safe to surface.
The building most businesses are standing in
The most useful way to understand where your business stands is to think about a building with four floors. Each floor depends on the one below it. The sequence is not optional.
The ground floor was built in 2010. Most businesses have lived there comfortably. Technical hygiene, basic structured data, crawlability. Familiar. Solid. But not sufficient.
The first floor has been under construction since 2022. Content extractability. Whether your content can be pulled from its context and still make sense. Whether your statistics have their sources in the same sentence. Whether your claims are attributed to a named person. Many businesses have started this work. Most have left it half-finished. Princeton GEO research found content with clear structural signals saw up to 40% higher visibility in AI-generated responses.
The second floor is where commercial decisions are being made. Trust. Selection. The editorial corroboration that makes a business safe for an AI to recommend. This floor is built not by what you say about yourself, but by what independent sources say about you. Seer Interactive found AI-cited traffic converts at 14.2% versus 2.8% for standard organic. The reason is simple: when an AI recommends you, the recommendation has already been made before the visitor arrives.
Most businesses are trying to reach the second floor via a temporary ladder. A few press releases. A Wikipedia page with thin content. Some Google reviews. Self-published thought leadership the AI correctly identifies as first-person claims. The ladder holds sometimes. But it is not a floor.
The third floor is being framed above you as you read this. MCP, WebMCP, agentic commerce protocols. For most businesses, this is a 2026-27 planning horizon. But the structural decisions you make on Floors 1 through 3 determine whether your business is accessible when agents start operating on the third floor.
How Forrester’s architecture and the Four Floors relate
Forrester’s four layers are horizontal and architectural: what technical infrastructure to build. The Four Floors are vertical and sequential: what order to build in, what the commercial consequence of each floor is, and why the sequence is non-negotiable.
| Forrester’s layer | Four Floors equivalent | What that means in practice |
|---|---|---|
| Layer 1: JSON-LD structured data | Floor 1 + Floor 2 schema | Schema is both entity foundation and content provenance. The @id graph pattern connects them. |
| Layer 2: Entity relationship mapping | Floor 1 (databases) + Floor 3 (corroboration) | Your entity graph is built partly from your own data, partly from independent sources that verify it. |
| Layer 3: Content API endpoints | Floor 4 (pre-MCP entry point) | Structured, programmatic access is the architecture behind MCP integration. |
| Layer 4: Provenance metadata | CITATE C3 + C4 at content level | Timestamps in JSON-LD are infrastructure provenance. Named statistics with inline sources are content provenance. Both are required. |
Where the Four Floors goes beyond Forrester: Floor 3 has no equivalent in his model. He treats provenance as a technical metadata problem — timestamps, version numbers. The Four Floors recognises that trust is also an earned off-page problem that infrastructure alone cannot solve. You can have perfect structured data and fail Floor 3 because no independent source corroborates what you claim about yourself.
Where CITATE specifically fills Forrester’s gap: his Layer 4 provenance is described at the infrastructure level. CITATE operationalises the same standard at the content level. A statistic without its source named in the same sentence fails Forrester’s Layer 4 regardless of what the JSON-LD says.
The three questions AI asks about your business
Underneath all the frameworks and acronyms, there are three questions an AI system answers when evaluating whether to recommend a business:
Can I find you? Floor 1 — entity foundation, Bing indexation, consistent structured identity across all surfaces AI consults.
Can I extract and verify what you say? Floor 2 — content extractability, CITATE, named attribution, provenance.
Do independent sources agree? Floor 3 — trust, selection, editorial corroboration.
Most content optimisation focuses on question 2. Most technical SEO focuses on question 1. Almost no one is systematically working on question 3 — because it’s hard, it takes time, and it doesn’t appear in any dashboard. But question 3 is where the shortlist is built.
Priority by business type
| Business type | This week | This quarter | 3-6 months |
|---|---|---|---|
| Professional services (law, accountancy, consulting) | CITATE audit on 3 key service pages. Bing indexation check. Name the partner/director as a named entity on every page. | Wikidata entity. Chambers / Legal 500 / sector directory listing. LegalService JSON-LD schema. | 15 editorial targets in your sector. This is the floor where your shortlist is built. |
| B2B SaaS / Tech | Bing indexation — Copilot and ChatGPT Search depend on it. Competitor displacement pages, CITATE-structured. JavaScript rendering audit. | G2, Capterra, Clutch reviews and listings. SoftwareApplication JSON-LD. Wikidata + Crunchbase entity completion. | Developer and editorial publications in your category. MCP: high relevance, plan integration architecture. |
| Enterprise | robots.txt audit across all subdomains. JSON-LD @id graph pattern — treat schema as machine-facing fact layer, not rich snippets. | One structured content endpoint for most-compared information. Provenance metadata on every public-facing product claim. | Analyst notes, industry publications, trade press at scale. MCP and WebMCP: architecture planning now. |
| Local SME | Google Business Profile — complete, accurate, active. Apple Business Connect. NAP consistency across every directory. | Bing Places for Business. Basic Organization + LocalBusiness JSON-LD. Reviews — volume, recency, response rate. | Local press and community editorial mentions. MCP: Floor 4 work, Floors 1-3 first. |
The sequence is non-negotiable
Floor 1 first. If Bing cannot find your key pages, everything else is irrelevant. If your entity is absent from Wikidata, AI systems cannot corroborate who you are. If your robots.txt is blocking the wrong crawlers, you’re invisible to systems that matter.
Floor 2 second. CITATE your three most important commercial pages this quarter. Not everything at once — just the pages where being cited matters most. Add named attribution. Write statistics with their sources in the same sentence. Rewrite your opening paragraphs so they stand alone.
Floor 3 ongoing. Identify fifteen editorial publications in your sector that carry genuine authority. Build a presence you didn’t write. Ensure your review platform presence is current, cross-platform, and actively managed. Get your entity into the databases that matter for your industry.
Floor 4 when ready. Monitor the MCP standard. Understand the agentic commerce protocols. Know which of your business processes are genuinely agent-automatable. But don’t implement before Floors 1-3 are solid — the lift doesn’t stop at unprepared buildings.
One more thing worth saying plainly
This industry has a habit of making this feel more complicated than it is. New acronyms. Competing frameworks. Cryptic signals from people who benefit from the confusion. It serves them. It doesn’t serve you.
What’s happening is straightforward. AI systems are making recommendations before humans ask for them. The businesses that get recommended consistently are the ones an AI can find, extract from, and verify through independent sources. The infrastructure being built — MCP, agentic protocols, structured content APIs — is real and significant, but it is the fourth floor of a building most businesses haven’t finished the second floor of yet.
Start with what you can audit, fix, and measure this week. The building is being constructed around you regardless. The question is which floor you’re on when the agents start using the lift.
For your current four-floor position, the AI Visibility Audit maps where you stand across ChatGPT, Perplexity, Copilot, and Google AI Overviews. For the content standard that makes Floor 2 work, see the CITATE framework. For the full architecture model, see the AI Discovery Stack.
