MCP server development that lets AI agents find, read and try your product
An agent-ready brand is a product that an AI assistant or coding agent can find, read, try and explain when a developer asks which tool to use. MCP server development is one part of it: a server that lets Claude Code, Cursor and ChatGPT query your docs directly. We test what agents say about you, build the layer they read, and measure your pick rate monthly.
Alpha: we take a few teams at a time. Three taps and your site, then we reply within two business days.
How an AI answer picks who to name
Which your category tool should I use for the job?
An example AI answer
The sources it read
Clients
Their stories ran in
Outlet coverage from our client reporting. Logos belong to their owners; no endorsement implied.
In short
An agent-ready brand is one that AI assistants and coding agents can find, read, try and explain when a developer asks which tool to use. Green Flag Digital tests what agents say about your product today, builds the layer they read (markdown pages, llms.txt, schema and a docs MCP server) and re-runs the test every month.
Why developers now ask an agent which tool to use
G2 surveyed 1,076 B2B software buyers in March 2026: 51 percent start research with an AI chatbot more often than Google, and 69 percent said a chatbot led them to a different vendor than they expected (reference 1). On December 17, 2025, OpenAI opened an app directory inside ChatGPT, and its Apps SDK is built on the Model Context Protocol (references 2 and 3). Webflow's developer docs already tell agents to add .md to any page URL for clean markdown, and point Claude Code and Cursor to a docs MCP server (reference 5).
Joe wrote in a July 2026 essay: The biggest, loudest, and most controversial systems of today point to the reality of tomorrow.
What an AI agent sees when a developer asks about your category
In a December 2025 walkthrough, Joe tested a GFD roundup of cybersecurity marketing agencies: I've done searches like this a lot, and this format is a really popular search format. We wrote a pretty thorough post. I like the way we structured it, in the sense that we went pretty deep and broke it out into a bunch of different categories based on different best-for types. If we go to an incognito window and search for best cybersecurity marketing agencies, Green Flag Digital actually doesn't show up until page three. If I go and search Bing for that same thing, Green Flag Digital ranks number one, which I think is very interesting. The point is there that ChatGPT did pick us up, and gave us citations multiple times, even though we're not on page one in Google.
His read: ChatGPT is separate from Google's ranking, but they are very tightly aligned with Bing. So it does still seem that ChatGPT is using Bing for its index and ranking. That's why ranking on Bing, or at least having them in mind and using their tools, is key. Then the catch: If we go into Google Analytics, we see zero from ChatGPT. Either nobody's clicking, or it's all being grouped as direct, or it can't be tracked. Both are worth keeping in mind.
So we ask the agents directly. One thing I've been testing is going directly into ChatGPT and just seeing what comes up. A more advanced iteration of this would be running a ChatGPT search for all of them and giving them their exact position. That idea became the agent pick test.
Agents guess when you give them little to read. Joe saw it while testing a custom GPT: When we give it just the title and stuff, it's not super accurate. I think it makes sense to give it a little bit more information rather than nothing.
What we do
What does MCP server development for marketing include?
Map: the agent pick test
We write fixed prompts in your developers' words, like "which [tool] should I use for [job]" and "[your product] vs [competitor]". We run them in ChatGPT, Claude, Gemini, Claude Code and Cursor, and record what each agent names, cites, reads and gets wrong. That gives your pick rate. We're gonna take the factor of keywords, pages, and links, and try to build our own score that's the highest correlating with the reality. The pick rate is that score for agents. Comparing two on a research task, Joe noted: I like seeing how there's a connector here. ChatGPT doesn't seem to use connectors as much. I like ChatGPT in scenarios like this, where you bring the data and the goals, and you can clearly measure if something worked or not.
Build: the agent-readable layer
- Markdown mirrors of your key pages and docs
- An llms.txt file and an agents.md style guide
- Clean schema on product, pricing and docs pages
- A docs MCP server, so agents can search and quote your docs
- Optionally, a ChatGPT app built on the Apps SDK
Joe's note on custom GPTs shows why a docs server matters: Knowledge is a key thing for the GPTs, because you can upload a bunch of files, and it will reference those files first, as I understand. It's a good way to store a bunch of stuff in a folder and then share the GPT around, so it all stays together as one package. A docs MCP server does that for every agent.
An agent reads every detail. SEOs have a bad habit of looking at a high level at all content, and trying to do things at scale. That means we often miss the minor details on a page. An old install command becomes the agent's answer. You can't rank for a keyword without a page. If you're only to think about and worry about one thing, it's a single page and the quality of that page as it's built for a topic or a subject that is valuable to the business. That's kind of like the core unit.
A good agent can also try what you sell. There's a tactic of just start doing the work for people. What is it like to work with you? You're kind of giving them a sense of it.
Run: a monthly re-test and fixes
Each month we re-run the pick test and fix what agents misread. We never mass-produce bot pages. The scaling up junk can work in the short term, but not in the long term. Connectors have gaps, so we test by hand. Joe hit one drafting email through an agent: Through the API connection between Gmail and Claude, it can't insert the signature, so I'm doing it manually, which is okay for now because I kind of want to QA them.
Marketing can own this work now. Joe wrote in 2021: As a marketer it's very frustrating to not know how to code. I know there's an entire world of automation and time saving that I can't access.
What's included
What you get
- The pick test: prompts, answers and your pick rate
- What agents misread, with a fix for each
- Markdown mirrors, llms.txt and an agents.md guide
- A docs MCP server tested in Claude Code, Cursor and ChatGPT
- A monthly re-test with a change log
How it works
- Apply, and Joe books a call.
- Map: we run the first pick test.
- Build: we ship the layer in a repo you own.
- Run: we re-test monthly.
You keep the code and a recorded walkthrough. I think it's good to go through the process of setting it up on your own, so you know and feel how the system works start to finish, so that you can make any edits you need to.
How we charge
Priced per scope after a call. Map is fixed, Build depends on your pages and docs, and Run is monthly, all agreed in writing first.
The limits
What we will not promise
No one can promise that an agent will pick you. We promise the work, the test and a clear number each month. For the near future, we will still be running these validation checklists manually. I would have it be part of the ideation stack or the editing stack, but not the whole stack.
This is an alpha service. Client MCP servers are what the alpha builds first, so we have not shipped one for a client yet. We take a few teams at a time.
What we already do to see how AI reads brands
GFD studies what ChatGPT cites in its ChatGPT citation study, runs an AI-answer monitor, and serves its own service pages from one system with structured data on every page. Joe wrote the principle in 2021: Create a site that's pure signal instead of noise.
Case study · Green Flag Digital
4 AI-built video ads for Green Flag Digital's own Quick Look campaign
Fifteen-second and six-second cuts, tall and wide, with the logo in the first frame and captions burned in.
From our own Quick Look ads


The work, in numbers.
- spec commercial made for a current client, vertical, with voiceover
- 30s
- spec commercial made for a current client, vertical, with voiceoverEvery claim on screen came from the client's approved claims list; the one example figure is labelled Example. The client stays unnamed until they approve it.
Results and coverage from our client reporting. Results vary by company, budget and market.
Joe RobisonJoe Robison
Founder and principal, Green Flag Digital
Joe sets the strategy on every engagement and makes the final call on what counts as evidence. A principal reviews the findings before any client reads them.
Joe Robison on LinkedInWhat clients say
“Green Flag consistently delivered high quality PR pieces that got mentions and links from some of the largest media brands on the internet.”
Ryan Carrigan, Owner, moveBuddha · via Clutch
Reviews from our Clutch profile. Results vary by company, budget and market.
Proof from B2B and consumer brands.
The results, and the full write-up behind each one.
B2B · Cybersecurity16earned stories for an upstart cloud security company, 23.3M potential readersRead the case study
B2B · Ed tech SaaS165 to 1,868weekly organic sessions after a three-site relaunchRead the case study
B2B · Business insurance51placements, USA Today, Newsweek and MarketWatch among themRead the case study
Consumer · Moving3,740referring domains from one data-study programRead the case study
Consumer · Health media+35%traffic on refreshed articles within weeksRead the case study
Consumer · Tours+902%organic traffic over four yearsRead the case study
Consumer · Bedding13Knew monthly visits from 22 articlesRead the case study
Consumer · Education15 daysfrom publishing the Ramen Index to a New York Times featureRead the case study
Consumer · Travel275+placements across 26 digital PR campaignsRead the case study
Results from our client reporting. Results vary by company, budget and market.
See where you stand.
Tell us where to look. We come back with where you stand today, who gets named instead of you, and the work we'd take first, then walk you through it in a 20-minute teardown.
Questions buyers ask
What is MCP server development?
MCP server development means building a server that speaks the Model Context Protocol, so agents like Claude Code, Cursor and ChatGPT can call it as a tool. For a brand, the first is usually a docs server.
Do we need a ChatGPT app?
Only when users can finish a real job inside the chat. Most start with the docs server.
Will this make AI agents recommend us?
We cannot promise that. We make your product easy to read and try, and measure the pick rate monthly.
How is this different from generative engine optimization?
Generative engine optimization works on what AI answers cite. This adds what agents use when they act, tracked with AI search analytics.
What does alpha mean here?
A new service, built with a few first clients and priced per scope after a call.
How do we apply?
Use Apply for alpha on this page. It opens a short form, Joe reads every application, and we reply within two business days.
What do you most want from search right now?
Three taps, then your site. About a minute. Public data only, no logins.
·
Sources and method
Methodology
This page describes how Green Flag Digital plans to build agent-ready brands as of 2026-11-15. It draws on Joe Robison's journal entries, voice memos and recorded working sessions from 2021 to 2026. Market figures come from the sources below, each checked on its source page on 2026-11-15. Search volume comes from Ahrefs.
References
- G2, The Answer Economy: G2's 2026 AI Search Insight Report (1,076 B2B software decision makers surveyed in March 2026). learn.g2.com
- OpenAI, Developers can now submit apps to ChatGPT, December 17, 2025. openai.com
- VentureBeat, OpenAI now accepting ChatGPT app submissions from third-party devs, launches App Directory, December 18, 2025. venturebeat.com
- MarTech, OpenAI expands ChatGPT into an app platform and a new channel for marketers, October 7, 2025. martech.org
- Webflow, developer documentation index for AI agents (llms-full.txt), checked 2026-11-15. developers.webflow.com
- CMSWire, Webflow Launches Source Agentic Workspace for Marketers, September 2, 2026. cmswire.com
- Ahrefs Keywords Explorer, US, 2026-11-15.
- Green Flag Digital journal entries, voice memos and working sessions, 2021 to 2026.
History of this topic
- 2021: Joe Robison writes that a site should be "pure signal instead of noise" (reference 8).
- 2024: Joe learns a new database tool in an hour with ChatGPT walking him through it (reference 8).
- October 2025: MarTech reports ChatGPT becoming an app platform for marketers (reference 4).
- December 2025: Joe finds ChatGPT citing a GFD roundup that sits on page three of Google and first in Bing (reference 8).
- December 2025: OpenAI opens app submissions and an app directory in ChatGPT; the Apps SDK is built on MCP (references 2 and 3).
- March 2026: G2's survey finds 51 percent of B2B software buyers start research in a chatbot more often than Google (reference 1).
- September 2026: CMSWire quotes Webflow's CMO saying AI agents read websites the way people do (reference 6).