Electric Social—a technology-forward restaurant event venue MWWE supports on the marketing site—is starting to see a steady stream of event inquiries from people who first learned about the venue in ChatGPT, not from a Google result or an Instagram post. Same pattern: someone asks an assistant where to host a meetup, a product launch, or a private dinner; the model names the venue; the person follows up on the site or by email.
That is not a trick. It is what happens when a business site is easy for models to read and cite. The habits that help are boring on purpose, and most of them align with the llms.txt convention: a small index file at /llms.txt that points assistants to clean Markdown versions of your important pages.
Electric Social is already happening
Event venues live on specifics: capacity, neighborhood, vibe, what kinds of gatherings you host, how to book. Electric Social is Berlin’s arcade, bar, and restaurant—exactly the sort of place people describe to an assistant in a sentence (“somewhere in Berlin with games and food for a group event”). If those facts live only inside a hero video or a PDF menu, assistants guess—or skip you.
Electric Social invested in a clear, crawlable web presence and the llms.txt pattern: structured summaries, pages that state what the venue is for, and machine-friendly companions alongside the normal HTML site. MWWE’s work on the Electric Social case study is the engineering side of that story—component structure and performance sized for a small team—but the discovery outcome is simpler: when someone asks ChatGPT about event space, the venue shows up with accurate context, and inquiries follow.
You do not need a custom model or an “AI strategy” slide deck for that. You need pages that answer real questions in plain language, plus an index that tells tools where to find them.
What changed for local search
For years, local discovery meant maps, reviews, and whichever SEO agency could keep your title tags fresh. That still matters. But a growing slice of discovery runs through conversational search: ChatGPT, Perplexity, Gemini, Copilot, and the assistants baked into phones and browsers.
Those systems do not browse like a patient human. They retrieve, summarize, and recommend from whatever they can fetch quickly—often favoring text that is already structured, recent, and unambiguous. A beautiful site that hides “we’re a Berlin arcade bar that hosts group events—book here” inside a carousel is a site models will misread or omit.
The practical shift for a café, salon, clinic, or venue: write for citation, not only for scroll depth. Name the city. Name the service. Name the next step (book, call, form). Repeat the facts humans need on the pages assistants are likely to pull—not only in footer fine print.
llms.txt in plain language
llms.txt is a convention, not a law of physics. At your domain root you publish a short file—/llms.txt—that looks like a curated readme for agents:
- A one-line summary of who you are
- Sections (Core, Services, Locations, FAQ—whatever fits)
- Links to Markdown companions for each important page, not just HTML URLs
On MWWE’s own site, npm run build generates /llms.txt, per-page index.html.md companions, and expanded context files from the same frontmatter that drives the public site. robots.txt points crawlers at the index. One source of truth; no duplicate CMS.
You can adopt the same idea on any stack:
/llms.txt— human- and machine-readable table of contents.mdcompanions — same facts as the public page, without nav chrome and script noise- Accurate summaries — title, description, and a short summary on every page that matters for revenue
Assistants are not guaranteed to fetch /llms.txt every time. But sites that publish it, plus clean Markdown siblings, are easier to ingest than sites that force models to scrape minified React and guess from og:image tags.
A short checklist
These steps are sized for a local business without a platform team.
1. Pick the money pages. Home, services, booking/contact, location(s), FAQ, and anything you’d want a stranger to understand in thirty seconds. Skip the 2019 holiday blog post unless it still drives calls.
2. Put facts in the HTML, visibly. Address, hours, offerings, pricing signals (“from €X”, “private events up to N guests”), and the primary CTA. Semantic headings help humans and parsers.
3. Add /llms.txt. Follow the spec: H1, blockquote summary, grouped links. Each link should target a Markdown URL (companion file or /page.md route), not only /page/.
4. Publish Markdown companions for those money pages. Same content as the live page—trimmed boilerplate is fine. Update them when hours or offerings change.
5. Mention the index in robots.txt. One comment line is enough; some crawlers and tools use it as a hint.
6. Keep one source of truth. Generate companions from your CMS or static site metadata if you can, so the HTML and Markdown do not drift.
7. Measure the old-fashioned way. Ask new leads “how did you hear about us?” ChatGPT and Perplexity answers will show up in free text long before any analytics dashboard labels them cleanly.
None of this replaces Google Business Profile, reviews, or word of mouth. It feeds the assistants the same way a good menu board feeds a walk-in customer.
What not to do
- Do not block useful pages in
robots.txtand hope AI will ignore you. If customers should find you, crawlers need the facts. - Do not stuff keywords for “AI SEO.” Models punish incoherence the same way humans do.
- Do not put critical details only in images or PDFs without repeating them in text.
- Do not treat llms.txt as set-and-forget. Stale hours and dead booking links become stale recommendations.
Local businesses do not need to predict which model wins next year. They need a site that states what it does in text, and an index that helps tools find that text quickly. Electric Social is an early, concrete example: event inquiries from people who met the venue in ChatGPT first. The playbook is llms.txt, Markdown companions, and pages honest enough to quote.
If you want help applying the same pattern to a marketing or venue site, contact MWWE.