Claude just shipped organic connector discovery first

If software is free, discovery is the moat

Muzamil opens with the question that has been running through Beyond the Bubble for weeks. If software is trivial to make and everyone can ship it, what actually matters? “If there is no mode,” he says, “if you are solving problems for whoever but you’re not getting discovered, what’s the entire point?”

That is the setup for the guest. Elliot co-founded Ghost Team a little over a year ago after working with enterprises like Stellantis, ManpowerGroup and Statista and watching each of them hit the same wall. They would build an MCP server, get it into the major registries, and then run into the harder problem underneath. “Once I build this MCP or build this MCP app, great, I’ve got it in the major registries, it’s in Claude, how do I actually get it discovered?” Elliot says. Ghost Team’s answer is a platform called agentdiscoverability.com, and the thesis is that a new optimization layer is being born inside ChatGPT, Claude and Gemini.

AEO is text, agent discovery is the widget

Elliot draws a distinction early that shapes the rest of the conversation. Answer engine optimization is about appearing inside the text of a chatbot’s response. It borrows from SEO but leans on Reddit and LinkedIn, because that is what the models actually read.

Agent discovery optimization is different. It is about your product surfacing as a usable widget inside the chat, not just being mentioned in a paragraph. “That sort of changes this from these things just being chatbots to sort of moving into more of an agent experience and also having a rich UI as well,” Elliot says.

The example both return to is booking.com. Inside ChatGPT today, you can go to the apps section, search for hotels, connect Booking, and the next time you ask to book a hotel the Booking widget itself appears instead of a list of text. That is the surface Ghost Team is optimizing for.

Claude quietly shipped organic discovery first

The moment most listeners will not have seen coming is Elliot’s claim that Claude is ahead. “Claude has been the first company or platform to actually launch organic discovery,” he says. Prompt Claude today with “book me a hotel for tomorrow night in London” without ever picking a connector, and you get a connector picker: Booking.com, Expedia, and others surfaced directly in the chat.

ChatGPT is still a keyword game inside its app registry. Type “hotels,” get a shortlist of relevant apps, connect one, and it fires on your next matching prompt. Gemini is further behind. Elliot notes that Google announced MCP connectors at I/O with roughly twenty launch partners, but it is not yet clear whether Gemini will adopt the shared MCP app standard or run its own.

Ghost Team has run tens of thousands of prompts through Claude to reverse-engineer how it surfaces connectors. Elliot’s bet is that the ranking playbook being written in Claude now will cross-apply to ChatGPT once it opens organic discovery. “The tactics that you’re learning within Claude, you can sort of, we believe, just cross apply them to ChatGPT.”

Walled gardens, not the open web

Muzamil walks through the pattern. The open web brought SEO and WordPress. Mobile brought gatekept app stores and 30% cuts. Which shape does the agent era take?

Elliot is honest about the tension. “I think it’ll be more like mobile apps, but I do believe there is a potential for it to slightly open up.” The logic is that platforms need developers, developers need distribution and monetization, and the labs are incentivized to build walled gardens where they take a cut. He points to a leaked Claude project called Conway, described as an always-on agent platform, as a signal that Anthropic is shaping up to capture its ecosystem.

He leaves a door open. If registries scale to hundreds of millions of apps, discovery inside them becomes intractable, and platforms may eventually let agents access a long tail of apps through trusted-web pathways alongside their curated registries. Security is the reason not to, though. Registries exist not just to monetize but to keep the trust model intact.

When your ad is never watched by an agent

Muzamil brings up a rant he had read from a site owner whose bandwidth was being consumed by millions of agent crawls with nothing in return. That becomes the entry into a longer thread about what changes when the primary users of the web stop being human.

Elliot’s answer starts with a line that lands. “Your ad is never going to be watched by an agent.” Whole channels break. The response, he says, is what he calls becoming headless: having the right surfaces so agents can understand and use your product without a human in the loop. That means an API, an MCP server, an MCP app and a CLI. But he is careful to frame that as step one. “You then need to actually make sure that they get discovered and used, which is why we’re so focused on the discovery challenge.”

For a dentist or a restaurant he pushes back on over-engineering. Building a CLI so people can book through Claude Code makes no sense. Meet users where they actually are, which today is ChatGPT apps and Google surfaces, and design the funnel so the booking actually completes.

The black box, and Claude’s three-day-old answer

Muzamil raises the observability problem. In the old web, traffic hopped between independent sites and you could measure it. Inside these LLMs, everything is locked in. How does anyone optimize against a system they cannot see?

Elliot’s framing is that this is a data science problem. You set up target prompts, run them daily, measure whether and where you appear, and iterate. The AEO industry, he notes, is already billions in size and operates entirely by reverse-engineering citation behavior.

Then he drops a fresh detail. Three days before the recording, Claude shipped Claude observability. If you have a connector live, you can now see your total rank across all Claude connectors, error rates, latency, tool call volume over 30 days, and where the traffic came from across Anthropic’s surfaces: Claude, Claude Code, or co-work. It is the first real analytics layer inside one of these clients, and Elliot treats it as a signal of where the others are heading.

The chicken and egg every early adopter names

Muzamil names the hesitation directly. Even with a billion weekly chatbot users, is the discoverability really baked in yet? Should he rebuild his digital footprint for a surface that is not fully switched on?

Elliot splits the customer base into two camps. The first are enterprises with innovation budgets who already believe and are shipping. The second are the wait-and-sees. He cites Airbnb publicly deciding not to launch a connector because they have a different view of what the conversational UI should look like. He does not argue against that position, but he makes the case for going early on two grounds. Testing teaches you the ecosystem better than analysis does. And more interesting: “These models learn what they’re trained on. So we’ve got companies right now that we work with that are getting millions of calls on their MCP server in Claude, and every single time that happens the model is learning from using their server.”

Being early is not just about today’s traffic. It is about being embedded in tomorrow’s models.

Two worlds running in parallel

Muzamil closes with the forecast question, one to two years and five to seven. Elliot’s answer is that two ecosystems will run in parallel rather than one winning outright.

The first is the walled-garden path. The frontier labs keep growing, harnesses like Codex, co-work and Claude Code get stickier, and the chatbot becomes a rich, dynamic UI layer. MCP apps evolve from end-user experiences into infrastructure. Agents pick them up autonomously as long as they sit in a trusted registry and the user approves. “Discovery is so, so important,” Elliot says, “particularly when an agent is just deciding autonomously what to select and a user isn’t even in the picture.”

The second is the more open path. Codex-style browser-use agents will use the web the way a human does: discover, authenticate, trust, transact. The web is not built for that. New standards will need to be adopted and companies will need to prepare their sites to be agent-usable, not just agent-readable.

By the end, Muzamil frames the moment plainly. The inevitability is baked in. The debate is timeline and route. Elliot’s closing point is that even inside a single coding model today you can see the split. Sometimes the agent wants a trusted connector from a registry. Sometimes it wants to go out to the open web and figure things out. The discovery layer that decides which app wins that call is the layer Ghost Team is trying to own.