Now that API Evangelist is itself an MCP server, a question I used to think about abstractly has become very concrete and slightly personal: how do you nudge an LLM? For twenty years the discovery game was search engine optimization — you shaped your content so that when a human typed a query into Google, your page came up and your point of view got into their head. That game is not over, but a new one has opened next to it. Increasingly the thing standing between your knowledge and the person who needs it is not a search results page. It is a language model, answering directly, synthesizing from whatever it has absorbed. So the question that follows me around is: when someone asks a model about API design, or governance, or whether they should use GraphQL, how does my point of view end up in that answer? How do you do SEO for an LLM?
Let me be honest about the uncomfortable part first, because the space is already full of people selling snake oil. You cannot reach into a frontier model and edit its weights, and anyone promising to “guarantee your brand in ChatGPT’s answers” is selling you something they cannot deliver. But that does not mean you have no influence — it means the influence works differently than the old keyword game, and pretending otherwise is how you get scammed. There are real, legitimate layers where you can nudge what a model says, and understanding which layer you are working at is the whole discipline. Get the layer wrong and you are either wasting money or getting played.
The first layer is the training-data layer, and it is the slow, honest one. Models learn from the public web, and if you have spent years publishing clear, structured, genuinely useful writing on a topic, that corpus becomes part of what the model absorbed. This is not a trick — it is just publishing, at length, over time, in a way that is legible to a machine as well as a human. A lot of my own influence here, to whatever extent I have any, is simply that I have written an enormous amount of structured material about APIs that models trained on. You cannot shortcut this and you cannot buy it. You can only earn it by being a clear, consistent, machine-readable voice on a subject long enough that the models cannot represent the subject without representing you. That is LLM SEO’s version of “write good content” — unglamorous, uncheatable, and real.
The second layer is the retrieval-and-context layer, and this is the one that actually excites me, because it is where being an MCP server changes the game. When a model answers by reaching out to tools, resources, and discovery surfaces at query time — instead of relying only on what it absorbed in training — then the structured surfaces you expose become a live channel into its answers. If API Evangelist is an MCP server, and an agent reaches for it while reasoning about API governance, my curated prompts and resources are right there, shaping the answer with current, structured, authoritative context rather than a fuzzy training-time memory. This is the retrieval-era successor to SEO: you optimize not for a crawler’s ranking algorithm but for being the clean, discoverable, machine-legible source a model pulls in when it needs to be right. And it connects straight back to why I keep arguing for federated discovery — the whole point of being in a discoverable, well-described index is that you are reachable at the moment of reasoning.
Which is exactly why nudging an LLM turns out to be the same work as everything else in this series, just pointed at a new consumer. Want a model to represent your view on GraphQL well? Publish clear, structured, machine-readable material on GraphQL — like the connective-tissue argument I made a few weeks ago — and expose it as a discoverable resource an agent can pull at query time. The training layer absorbs it slowly; the retrieval layer reaches it immediately. There is no secret injection point, no keyword incantation, no gateway you bribe. There is just the same discipline the whole industry keeps relearning: be clear, be structured, be machine-readable, be discoverable, and be the authoritative source often enough that the model cannot do the topic without you.
So how do you nudge an LLM? You do it the way you were always supposed to do discovery, only now the audience includes a machine that reasons over your work in real time. You publish structured knowledge, at length, honestly. You make it machine-readable and discoverable — increasingly as an MCP server and a federated index, not just as HTML for a crawler. And you accept that there is no shortcut, because the models are, in the end, a reflection of the public knowledge we all put into the world, weighted toward whoever was clearest and most consistent. That is not a loophole to exploit. It is a reason to keep doing the work I have been doing all along, and to do it in a form the machines can reach. LLM SEO is not a new dark art. It is the old discipline of being genuinely useful and genuinely legible, aimed at a new kind of reader — and being an MCP server is how you make sure that reader can actually find you.
