Inside the live demo site: what each node actually does

A demo node is one working piece of the system you sell, running live on a page your prospect can open in their own browser. Lumina Sales gives you six of them: a voice receptionist they can call, a chat agent, a website audit that fetches their real domain, a missed-call recovery sequence, an ROI calculator seeded with their numbers, and a follow-up sequence that plays out on screen. You pick which ones go on the page, in what order, and the prospect works through them without an account, an install, or a scheduled call with you.

That is the whole product idea. Most AI sales calls die because the prospect has to take your word for it. A node is the part where they stop taking your word for it.

What counts as a node, and what does not

A node is interactive. That is the line. A screenshot of an AI receptionist is not a node. A Loom of a receptionist answering someone else's phone is not a node. A phone number they can dial, in their trade's language, with their business name on it, is.

The distinction matters more than it sounds. Consensus, which sells demo automation software and analyzed more than 6 million anonymized interactions on its own platform for its 2026 B2B Buyer Behavior Report, found the average demo view lasts 5 minutes and 14 seconds against a total runtime averaging 20.32 minutes. Buyers self-select roughly a quarter of what you built and leave. The same report found most buyers will not click through more than four steps of a product tour before bouncing.

Read that as a design constraint, not a tragedy. You have about five minutes of attention and four clicks. Nodes are how you spend them.

The six nodes, and what each one proves

Voice receptionist. The prospect starts a call in the browser and talks to an agent trained on their trade. It answers pricing questions inside their price bands, triages emergency against routine, and books. Any single call is capped at five minutes. This is the node that changes the room, because a plumber who has spent twenty years answering his own phone knows within three sentences whether the thing is any good.

Chat agent. Same agent, on a website widget instead of a phone. It answers what the practice manager was going to ask you at 11pm. Cheaper to run than voice and useful when your prospect is somewhere they cannot talk.

Website audit. It fetches their actual domain and reports findings against a site they already know by heart. This is the most underrated node on the list. It is the only one where the prospect cannot suspect you staged the input, because the input is theirs.

Missed-call recovery. An unanswered call, a text back, a reply, a booking. Thirty seconds, start to finish, watched rather than promised. The reason this lands is arithmetic they already believe: they know what a missed call costs them, they just have never seen anyone close the loop automatically.

ROI calculator. Seeded with their average ticket and their close rate, so the number on screen is a sum they can follow rather than a figure you asserted. Do not inflate the inputs. If they catch you padding the ticket size, every other node on the page becomes suspect.

Follow-up sequence. The nurture plays out in front of them instead of sitting in a slide. Useful late in the page for the prospect who is sold on the receptionist and wants to know what happens to leads that do not convert on the first touch.

You build the page by picking the trade, typing the business name, and arranging the nodes. It publishes on your subdomain with your logo, your colors, and your booking link. There are packs for 100+ verticals, each carrying its own services, price bands, and objection language, plus a generic pack written to read as industry-neutral rather than as a template with the trade left blank.

Why the demo has to run on someone's infrastructure

Nothing about a live voice node is free. OpenAI's Realtime API guide describes the architecture behind most of these agents: a speech-to-speech model that works directly with audio, keeps conversation state, and calls tools mid-call, connected over WebRTC from the browser. The guide is candid that the interesting engineering is barge-in, first-audio latency, and natural turn taking, which is to say the parts a caller notices immediately when they go wrong.

That engineering costs money per minute. OpenAI prices audio input for GPT-Realtime-2.1 at $32.00 per million tokens and audio output at $64.00, and latency work is its own line item on top. Telnyx published a latency comparison in September 2026 that is vendor content and reads like it, but the useful part is its insistence that component latency and full-turn latency are different measurements that cannot share a ranking column. It cites a Cekura test holding the model stack constant across platforms where full-turn median latency ran from 1.73 to 3.16 seconds depending on the platform. A second and a half of dead air is the difference between an agent that sounds like a receptionist and one that sounds like a hold queue.

Which is the honest reason the plan has caps. You get 100 live voice minutes a month, 500 chat messages, and 50 website audits, with no API key to obtain and no phone number to buy. At a five-minute maximum per call, 100 minutes is about forty demo calls. Hit the cap and the live nodes pause until the next cycle while the rest of your page and portal keep working. Your prospect sees a neutral message saying the demo is not accepting calls right now, never a billing notice and never your plan name. Nothing overspends, because there is no metered invoice. The full numbers are on the pricing section.

How to sequence a page you are actually going to send

Order matters more than node count. The instinct is to put everything on and let the prospect explore. Do not.

Lead with the node that proves the thing they are most skeptical about. For an HVAC owner who thinks AI cannot handle a real emergency call, that is the voice receptionist. For a dentist who thinks their site is fine, it is the audit. Put the ROI calculator after at least one live node, never before, because a number means nothing until they believe the system producing it works.

Then send the link and stop talking. Consensus found 45% of demo views happen within the first 60 minutes of the send, and that buyers take an average of five days to open sent content. Both of those are true at once, which means your follow-up window is either the same hour or the next week, with a dead zone in between. Plan the second touch for day five, not day two.

The dashboard tracks opens, calls, and audits, so you know which node they used and for how long before you call back. That is a better opening than "did you get a chance to look at it."

Where this falls down

One published demo at a time. If you are working ten deals in parallel and want ten live pages, the plan does not do that. You republish between calls, which is fine at the volume most one-person agencies actually run and annoying above it. Drafts are unlimited, so the building is not the bottleneck. The publishing slot is.

Every published demo carries a sandbox disclosure. The page states that the AI is a demo and that bookings inside it are simulated. That is deliberate and it is not removable. Some reps hate it because it breaks the illusion. I would argue the opposite: a prospect who discovers on their own that the booking was fake has caught you, and you will not recover that call. A prospect who is told upfront treats it as a proof of concept, which is what it is. We wrote more about what prospects actually see on a white-label demo if you want the full accounting of what is visible and what is not.

The five-minute call cap is real. If your pitch depends on a fifteen-minute open-ended conversation with the agent, this will frustrate you. Most demo calls that go past five minutes are the prospect playing with the toy rather than evaluating it, but not all of them.

And the trade packs are only as good as the trade. The catalog covers 100+ verticals, but if you sell into something genuinely unusual and appointment-driven is a stretch for it, the generic pack will read as generic. Niche packs get written by hand on request, and requests from subscribers get built first, but that is a queue and not an instant.

Finally, none of this helps if you have nobody to send it to. The demo converts attention you already have. It does not create it. Small business AI adoption has been flat for six months among the smallest firms according to Census data, which is an opportunity and a warning at the same time: the market is wide open and almost nobody in it is looking for you.

FAQ

Do I need to configure anything before my first call? No. The voice and chat agents run on our infrastructure and the minutes come with the subscription. There is no API key to obtain, no phone number to buy, and nothing to install. Most of the ten minutes to first demo goes to choosing a subdomain and uploading a logo.

Can I put all six nodes on one page? You can, and you probably should not. Buyers self-select about five minutes of content and rarely click past four steps. Three well-chosen nodes in the right order beat six in a pile.

Does the website audit really fetch their site? Yes. It runs against their live domain, which is why it is hard to dismiss. It is also why you should open the page yourself before you send it, in case their site is down or behind something that blocks the fetch.

What happens to my demos if I cancel? Published demos unpublish so the links stop serving, and everything you built stays as a draft. Nothing is deleted. You keep your login with zero voice, chat, and audits until you resubscribe, then it goes back up unchanged.

Do you contact the prospects who open my demo? Never. We do not email, call, or market to anyone who opens your page, and the addresses captured by the email gate are yours. Lumina Sales is not a lead marketplace.

Next step

Open the live demo on the homepage, switch the trade to whichever one you sell into most, and call the receptionist. Ask it the single hardest pricing question your last prospect asked you. If it handles that, you know what your page will do on your next call, and you can start building yours for $49 a month.