The buyer's first stop is AI, not your site

For a growing subset of buyers, the first stop is AI, not your site. They open ChatGPT or Claude, ask the question, get a shortlist of three tools — and your website never enters the journey at all. The Agent Layer is the work of being present at the moment AI does the shortlisting. It is also, frankly, the most uneven stage in the methodology — high-leverage for some clients, structural future-proofing for others, irrelevant for a few.

For SaaS and AI tools, this work moves citation rates now. For local services, almost none of it applies. The work here is segmenting first, building second.

The honesty cuts both ways. For SaaS products, AI tools, and B2B brands whose buyers research via AI agents, this work moves citation rates now. For local services, relationship-driven sales cycles, and consumer brands evaluated through human discovery, almost none of it applies.

Why this stage exists at all

The buyer journey is moving upstream of the URL. Increasingly, when a buyer needs a tool, they don't open Google — they open Claude or ChatGPT, ask the question conversationally, get a shortlist, and pick from it. The URL is incidental.

That shift has produced a new surface AI consumes: AI tool directories, Custom GPTs, MCP servers, and the machine-readable documentation that lets AI agents reason about whether a product fits. The Distribution pillar calls this the agent-readable layer. Once you give it a name and a build sequence, it becomes The Agent Layer — the set of surfaces where the brand sits inside the buyer's AI workflow rather than alongside it. A few things distinguish this from every other stage in the methodology:

Impact is segment-specific. Foundation, Authority, and most of Distribution compound for almost every enterprise. The Agent Layer compounds dramatically for a subset and barely registers for the rest.

Selling it as universal is overselling. Skipping it for the wrong clients is just good consulting.

The literature is genuinely thin. Custom GPTs, MCP servers, and AI-native discovery surfaces are 2024–2026 inventions, and no mature playbook exists yet. Anyone selling a “complete Agent Layer methodology” with case studies and citation-impact data is, honestly, selling something that doesn't exist with that confidence.

The early-mover advantage is real and time-bound. The brands building this infrastructure now will have a structural advantage by 2027–2028 — by which point the surfaces will be more developed, more competitive, and harder to enter cheaply. The window for landing positions in Theresanaiforthat or earning early credibility in Anthropic's MCP registry is open now and won't be open forever.

The playbook

Five moves, in order. Total time varies sharply by which moves apply — anywhere from 8–15 hours for a non-SaaS client running the lighter pieces, to 80+ hours for a SaaS client shipping an MCP server. The segmentation in Move 1 determines the rest.

Every stage in the methodology starts with diagnostic. The Agent Layer takes this further than any other stage, because the cost of building the wrong surface for the wrong client isn't just wasted budget — a Custom GPT nobody uses is a credibility cost, an MCP server with no genuine use case is a press release the engineers will resent. Six client buckets, mapped to which surfaces apply:

The output of Move 1 is a one-page segmentation note: which surfaces apply, which the team is explicitly skipping, and why. The “why we're not doing this” decisions are as important as the “why we are” decisions — they're what protects the engagement from creeping into work that doesn't earn its place.

This is the most universally applicable component for any tech-adjacent client. AI tool directories function similarly to traditional business directories but are increasingly cited by AI systems when buyers ask “what's the best tool for X.”

The directories that matter, in rough order of citation impact:

Worth flagging honestly: there are dozens of AI tool directories. Most are noise. Submitting to all of them inflates a count but doesn't move citations. The six above plus any category-specific aggregators the citation audit surfaces are what's worth the time.

For each directory: consistent canonical data drawn from The Verification Stack (same name, description, category), the full description length available, the primary category that matches buyer-intent prompts from The Brief, a quality logo and screenshots, and a link to the canonical landing page rather than the homepage.

The honest assessment of Custom GPTs: most don't move citation rates, and most aren't worth building. The ones that do work — branded expert tools, not promotional bots — only work when the brand has genuine methodology to package as a tool. A useful filter, three questions:

If all three answers are yes, build it. Patterns that work: a pricing software company building a “Pricing Strategy GPT” trained on their methodology; a marketing agency building a “Campaign Brief Builder” that walks users through their proprietary framework; a B2B tool building a “Setup Helper” for configuration. If any of the three is no, skip it. Realistic expectations to set up front: most GPTs get 100–1,000 conversations in the first month if actively promoted, 10–100 if not. The top GPTs in popular categories get six figures of weekly conversations, but they earn that through sustained promotion and genuine utility — not through a launch tweet. The build itself is straightforward; the work that matters is in the design: a clear narrow purpose, a knowledge base that gives the GPT something specific to do, and conversation starters that signal what kind of question to ask.

MCP (Model Context Protocol) is Anthropic's open protocol for connecting external tools and data to AI assistants. Launched in late 2024, it has matured through 2025–2026 and is now genuinely emerging as a discovery surface — particularly for SaaS products with a programmatic interface and a technical user base.

When MCP fits: the client is a SaaS product with an API, the buyer would benefit from accessing the product from inside Claude, Cursor, or another MCP-supporting AI client, and the product is something users would want to query from inside an AI conversation — a CRM, an analytics tool, a project management app, a knowledge base. When it doesn't fit: local services, pure content brands, products with no programmatic interface, or SaaS where the buyer is non-technical. The work is a real engineering project, not a marketing sprint — typically 30–80+ hours including testing, documentation, and submission to relevant registries. Distribution surfaces:

The honest framing: MCP is real, adoption is growing, and early adopters will benefit. But shipping an MCP server because the founder wants the announcement rather than because users want the integration produces a maintenance burden with no payoff. Build it when there's a genuine use case. Don't build it as a press release.

The least glamorous and possibly the most durable move in the stage. AI agents — both consumer-facing assistants like ChatGPT and Claude, and the wave of vertical AI agents emerging through 2026 — need to consume documentation to reason about whether a product fits a buyer's need. Most enterprise documentation is built for humans clicking through a hierarchy. AI agents read differently. The work, in rough order of impact:

The buyer never sees it. The AI does, and that's the entire point.

What to measure

Five Agent Layer signals. Some are still emerging and tooling is sparse — but they're concrete enough to track meaningfully. Report quarterly; the Agent Layer compounds on a longer cycle than the rest of the methodology, and monthly reporting on a quarterly-moving stage trains stakeholders to read normal variance as failure.

Common failure modes

Five patterns. Each one is the result of selling this stage as something it isn't, or skipping work that isn't glamorous.