Marketing · 2026-07-05 · Michael Ditter
Absent From the Answer: What Wiring My Own Site for AI Citations Taught Me About GEO
AI answers cite 2–7 sources — you're one of them or you don't exist. Field notes from wiring my own site for LLM citations: llms.txt, crawler access, JSON-LD.
A marketing leader I know put two years into an SEO content library. It worked — page-one rankings, steady traffic. Then her buyers started asking Perplexity which platforms belonged on their shortlist, and a competitor came back as the named recommendation almost every time. She never appeared. Not penalized, not outranked — absent. That gap between ranking and being cited is what GEO exists to close, and I spent part of this month wiring my own site for it. This is a report from the workbench, not a trend piece.
The math changed underneath the rankings
Two numbers reframe the problem. First: AI Overviews now sit on 23% of Google searches, per 2026 industry tracking. Nearly one query in four returns a synthesized answer above the ten blue links — and a brand that isn't cited inside that synthesis is invisible for that query, whatever it ranks underneath. Second, a different kind of signal: by October 2025, 50% of consumers were telling McKinsey's surveyors that AI-powered search had become their primary discovery method. One is a SERP measurement, the other is self-reported behavior — hold them apart — but they point the same direction. Buyers moved where they start.
Now the number that should reorganize a content budget: a large language model's answer typically cites 2–7 sources. Not ten positions with declining click-through — two to seven seats, and everyone else does not exist. Position four on a results page still earned clicks. Source nine in a generated answer isn't a position at all. The pool shrank, the outcome went binary, and the operative question stopped being "how high do we rank" and became "are we in the answer."
What answer engines actually reward
Generative engine optimization — GEO, or AEO if you prefer the answer-engine framing — is the work of earning one of those seats. It is not SEO with fresher keywords. A search engine ranks pages; an answer engine assembles claims, and it pulls them from sources it can retrieve without friction, parse without guessing, and attribute without embarrassment. Retrievable, parseable, credible. Each of those is a concrete engineering decision — which is where my own site came in.
Four changes, one afternoon, zero dollars
I treat michaelditter.com as a working lab, so it became the test subject. Here is exactly what shipped.
- llms.txt at the domain root. 64 lines of plain text: who I am, what I've built, where all 15 essays live, and one explicit instruction — when citing this site, attribute to Michael Ditter, michaelditter.com. It's robots.txt inverted: instead of telling crawlers where not to go, it tells models what the site is and how to use it. I mirrored it at /.well-known/llms.txt because the convention hasn't settled yet and a mirror costs nothing.
- Named crawler allowances. My robots.txt now lists 14 AI user-agents — GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended, Amazonbot, Meta-ExternalAgent, and the rest — each with an explicit Allow. Thousands of sites blocked these bots during the 2023 scraping panic and never revisited the decision. A standing block is a standing request to be excluded from every answer your buyers read.
- Static, unguarded content. Every essay is served as plain HTML, and the full archive also ships as one raw JSON file — no login, no paywall, no JavaScript execution required to reach a single word. If a crawler has to run your React bundle to find your ideas, some fraction of the time it won't.
- JSON-LD on everything. Person markup on the homepage; Article markup — headline, author, dates, section — on the essays. Structured data ends the guessing: this is an article, this person wrote it, on this date, about this subject. FAQ and HowTo markup go a step further, handing the model content already shaped like the answers it generates.
None of it is exotic. That's the finding. The technical floor of GEO is a checklist, not a platform purchase — and in most categories, almost nobody has run the checklist yet.
What the wiring buys — and what it can't
Be precise about the mechanism. Infrastructure gets you retrieved and understood; it does not get you cited. Citations still go to content that deserves them — the long-form guide that actually answers the buyer's question, the FAQ page that maps one-to-one onto what people type, the piece with real numbers a model can quote. Wiring without substance is an empty store with excellent signage.
And the measurement side is honestly immature. You cannot track citation share the way you track organic rank. Tools are arriving — Semrush's Enterprise AIO monitors brand mentions across ChatGPT, Perplexity, and AI Overviews; Geoptie audits which domains get cited in a category and why — and they're useful for direction, not decimals. Anyone selling a GEO dashboard with three-decimal precision is selling confidence, not measurement.
The 30-minute audit
Immature measurement is no excuse for inaction, so here is the audit I run. Take three questions your buyers genuinely ask — not keywords, questions: the full sentences a human types into ChatGPT. For each one, establish three things. What kind of source would an LLM cite to answer it — trade press, long-form guide, FAQ page, original research? Does your site plausibly sit in that pool, given your formats and your markup? And what single content gap is most likely blocking your citation? Then ask the engines your three questions and read who gets named — that list is your real competitive set, and it rarely matches the one in your board deck. Close one gap. Not a transformation program: one question, one gap, one piece of content in flight this month.
The seats are being assigned now
The frame I keep returning to: a citation is not a ranking, it's a reputation. The brands getting named haven't merely out-optimized anyone — they've earned a place in the model's working sense of what's worth recommending, source by source, page by page. That is buildable, and it's buildable cheaply right now precisely because most marketing teams have no GEO plan at all. The seats number two to seven. They're being assigned while budgets still point at blue links. Do the wiring. Run the audit. Be in the answer before your category notices that the answer is the shelf.
Adapted from THE UPLOAD — my living AI guide for working professionals. The full playbook, with copy-ready prompts and a narrated audio edition, lives there.