Zeekend

Sponsored product placements for AI apps. When someone tells your assistant their pan is scratched, Zeekend can put the right pan in front of them, labeled, at the moment it helps.

• Zero perceived latency. The auction fires on the user's question and runs while your model streams. The slot is decided before your answer finishes.
• Nothing fills most of the time. Roughly 5 to 10% of turns. That is the design, not a bug.
• No identifiers, no profiles. No cookies, no device or advertising ids, nothing stored on the device, no profile built across apps. Conversation text does reach us — the last question, optionally the answer — because that is what the matching runs on. See What we receive.

---

Quickstart

The one-line way — installs the package AND finds your chat's message list and drops the component in for you, no copy/paste:

npx @zeekend/sdk init

It edits your source, so here is exactly what it touches before you run it. Two writes, both in bin/init.cjs and both worth reading first: it inserts the component into the one file that renders your message list, and appends your keys to .env. Nothing else, and no network call beyond the install.

It does not show you the edit before making it. Run it on a clean git tree so git diff tells you everything it did — or skip it entirely and use the manual snippet below, which produces the same result. Preferring not to let a script edit your code is a reasonable position and costs you nothing here.

Works when your chat UI renders messages as {messages.map(...)} (true for most React chat apps, including useChat() from the Vercel AI SDK). If it can't find that pattern with confidence, it says so and prints the manual snippet below instead of guessing.

Or by hand:

npm i @zeekend/sdk

import { ZeekendSlot } from '@zeekend/sdk/react'

<ZeekendSlot publisherKey="pub_test" messages={messages} />

Drop it under your message list. That is the whole integration.

Using an AI coding assistant instead? Paste this: Read https://exchange.zeekend.com/skill.md and install Zeekend in this app. (or /skill-dashboard.md for the version that also wires up publisher.zeekend.com reporting.)

pub_test is the sandbox key. No signup, no waiting. It fills on almost every turn so you can see it working in your own app in about two minutes, and it bills nobody. When you have seen it work, get a live key at zeekend.com and change that one string.

Sandbox fill rates are not real. Live fill is closer to 5 to 10%, because most turns genuinely deserve nothing.

What messages needs to be

Your existing array. [{ role, content }], and the common variants ({role, text}, {from: 'user'|'bot'}, Anthropic-style content blocks). If you use useChat() from the Vercel AI SDK, pass its messages straight in.

Everything else is derived from it: which turn this is, what the question was, whether the assistant is still streaming. There is no turn id to keep stable and no streaming flag to thread through.

---

Options

Only publisherKey and messages are required.

<ZeekendSlot
  publisherKey="pub_live_..."
  messages={messages}
  conversationId={thread.id}
  relevance={0.6}                     // optional floor 0-1; unset, each campaign's own applies
  dimensions={{ maxWidth: 640 }}      // leave one axis unbounded
  blockCategories={['gambling', 'crypto']}
  onNoFill={reason => showBackupNetwork()}
/>

relevance is optional. Leave it out and each campaign's own cutoff applies, which is what the exchange tunes. Set it, here or from your dashboard, only to be stricter than that: higher means fewer, better matches and lower fill. Move it on your own numbers, not ours.

Pacing is set from your dashboard, not here: the first turn a placement may appear on, turns between placements, and placements per session. Unset, the first placement can appear on turn 2. Passing minTurns, turnGap or maxPerSession here still works, as a local ceiling: whichever of yours and the dashboard's is stricter wins.

includePrevious (default false) also sends the user's previous message, clipped to 500 characters, so a follow-up like "something classic" can be matched to the watch it is about. It sends a turn the SDK otherwise never does, so turn it on only once your privacy notice covers it.

keepPinned (default true) keeps the conversation at the bottom when the ad appears. The ad is drawn after your app has scrolled to the end of the answer, so without this it lands just below the fold. It only scrolls when the reader was already at the bottom; someone who scrolled up to reread is left where they are. Set false if your app scrolls to new content itself.

Sponsored lines

A placement arrives as a card, a catalogue, or a sponsored line: a short labelled sentence after the answer, naming one product. The advertiser chooses which, per campaign. serve() renders all of them.

A sponsored line has three parts, and who owns which one is the point:

the lead-in	your words, in your assistant's voice
the claim	the advertiser's words, all of them inside the link
the label	"Sponsored", after the claim rather than before it

The lead-in defaults to "You might also want to look at". It is the one string here worth writing yourself, because it is the only part in your voice — and an advertiser buying the words that introduce their own ad is exactly what the split exists to prevent.

zk.serve({ mount, messages, inlineLeadIn: 'Worth a look:' })

<script src="https://exchange.zeekend.com/z.js"
        data-key="pub_live_..."
        data-lead-in="Worth a look:"></script>

Rendering slots yourself? A sponsored line arrives as slot.format === 'inline' with the sentence in slot.inline. headline and body still ride along, so you can render a card instead without a second request.

Rendering several slots, or want config in one place:

<ZeekendProvider publisherKey="pub_live_..." relevance={0.5}>
  <App />
</ZeekendProvider>

---

Checking it works

import { useZeekend } from '@zeekend/sdk/react'
const zk = useZeekend()
zk.stats()   // { requests, fills, fillRate, errors, timeouts, errorRate, p50, p95 }

Watch errorRate, not just fillRate. A broken integration and a quiet one look identical from outside, which is exactly why they are separate counters.

What you see	What it means
errorRate above 0	Connection or auth. The console warning names the fix.
errorRate 0, fillRate 0	Working. Nothing matched. Lower relevance.
requests 0	The slot is not mounted where anything renders.

The SDK warns in the console, without debug, when it cannot reach the exchange, when your key is rejected, and when it loads but never fires. Silence is never a valid state.

Your earnings

curl https://exchange.zeekend.com/v1/publisher/stats \
  -H "x-publisher-key: pub_live_..."

Impressions, clicks, and what you have earned. No dashboard, no emailing us.

---

What we receive

Sent	Never sent
The user's last message	User id, email, phone
The assistant's reply, second pass only	Cookies, device ids
Your publisher key and placement id	Your system prompt
Coarse locale	Earlier turns, unless you set includePrevious (below)

Text is clipped in your app before the request leaves: 2,000 characters of question, 4,000 of answer. Read trimContext() in the source rather than taking our word for it — and note what it does, which is truncate. It does not scan the message for names or emails and remove them. If a user types their own details into a conversation, that text is inside the message and reaches us with it.

Where it goes after us. Deciding how well a product fits a conversation is done by a language model, so the question and answer are processed by Anthropic as our sub-processor. Conversation text leaves our infrastructure for that step. We would rather say so here than leave you to find it.

conversationId is optional and yours. It ties follow-up turns together and we never join it to anything, but we also cannot tell whether the value you pass is a random thread id or something that identifies a person in your own systems. If in doubt, leave it out — nothing depends on it.

Ad safety

Every unit ships with a Report control, no integration work. Reports reach us and the advertiser is pulled from your app.

Generated copy is fenced by each advertiser's own claim rules: required phrases must appear verbatim, banned words are stripped server-side after generation. We do not let a model invent facts about someone's product.

The exchange scores every candidate at zero when a conversation involves health, safety, injury, or money trouble. If your users are having a bad day, you earn nothing on that turn.

Be clear about what that is and is not. It is a rule about monetisation, applied at the exchange, and it runs after the text has arrived. It is not a guarantee that such a message never leaves your app — it does, and it reaches our model provider as part of the matching. If your users discuss anything you would not send to a third party, that is a reason not to install this, and we would rather you decided that now than discover it later.

Billing

• An impression is 50% of the unit visible for one continuous second.
• Rendering is not an impression. Prefetching is never an impression.
• One impression and one click per slot, ever, enforced server-side.
• A click with no prior impression is invalid traffic and bills nothing.

Publishers earn a CPM on viewable impressions. Advertisers pay CPA on confirmed orders. We carry the spread.

---

Custom rendering

If you want the unit to match your app exactly, or you are on React Native where there is no DOM:

import { useZeekendSlot } from '@zeekend/sdk/react'

const { slot } = useZeekendSlot({ publisherKey: 'pub_live_...', messages })

if (slot) {
  // slot.headline, slot.body, slot.price, slot.image, slot.clickUrl
  // call zk.impression(slot) when it is 50% visible for 1s
  // call zk.click(slot) on click
}

Two rules: use slot.clickUrl, not slot.url, and keep slot.disclosure visible. Everything else is yours.

Not using React?

import { Zeekend } from '@zeekend/sdk'
const zk = Zeekend.client({ publisherKey: 'pub_live_...' })

zk.attach({ mount: messageEl, messages })   // call whenever messages change

Repeat calls within a turn are ignored, so wire it into your render loop.

Source and releases

This repository is the SDK's source. Releases are published to npm from here by GitHub Actions with provenance, so the npm page shows the commit each version was built from. The exchange the SDK talks to is a separate, private service; everything the SDK sends it is listed under What we receive, and in SECURITY.md.

Protocol

If you would rather not use the SDK at all:

POST /v1/slot        -> 200 + slot object, or 204 No Content for no fill
POST /v1/event       -> { type: impression | click | report, slotId }
POST /v1/conversion  -> { slotId, orderValue }   from your order webhook
GET  /v1/health

204 means nothing matched. Any 4xx or 5xx means we are broken. Treat them differently.

Base URL https://exchange.zeekend.com/v1.