Onboarding the first cohort of publishers

Ads for AI apps that never read the conversation.

Cognatu matches ads to what an AI conversation is about — entirely on the user's device. Our servers never see a word anyone writes. Publishers keep 70% of every impression.

70% to publishers 1 line to integrate 0 conversations transmitted
Quiet — AI journaling matched on-device
I keep waking up at 3am with my mind racing about work.
That pattern — sharp waking with looping thoughts — often tracks with unprocessed stress rather than a sleep disorder. A few things worth trying this week…
SPONSORED

Calm — Sleep stories and a 7-day wind-down course. Try it free →

topic:sleepstress publisher +$0.021
Works with
OpenAIAnthropicGeminiLlamaMistralAny OpenAI-compatible API
70%
Revenue to publishers
on every impression
$20–35
CPM in premium verticals
dating, wellness, career
<4 hrs
To first impression
one function call
0
User profiles built
matching is on-device
01The problem

Decisions moved into AI conversations. Advertising didn't.

a.

Banners don't belong in a conversation

Display units were built for pages. Dropped into chat, they read as an intrusion — and users treat them like one.

b.

The tracking era is ending

Third-party cookies are gone and cross-site identity is collapsing. The networks built on following people around the web are running out of signal.

c.

Reading conversations is a breach of trust

Some AI ad networks monitor prompts server-side to pick ads. In apps where people discuss health, money, and relationships, that isn't a growth strategy — it's a liability.

02The concept

The whole idea, in forty seconds.

One rule makes everything else possible: the conversation never crosses the device boundary. Watch what does.

AI journaling app
Thinking about finally moving to Lisbon… is that crazy?
Not crazy — but worth stress-testing. You've mentioned remote work is going well; the bigger questions are visa path and what you'd be leaving…
SPONSOREDFlatio — Month-to-month furnished rentals in Lisbon. See neighbourhoods →
on-device classifier
relocationtravelconf 0.91
device boundary
what never crosses: the conversation · identity · history
topic: relocation
the auction — bids on the topic, not the person
Flatio$24.10
Remote-work insurance co.$21.40
Language-learning app$18.90
second-price auction · clears at $21.50 CPM
publisher 70%
30%
this impression: publisher earns $0.0151 · clearly labelled · no match → no ad
1 / 5
03Integration

One line changes. Nothing else does.

Your model, prompts, and response format stay exactly as they are. The SDK does the rest.

Step 1

Wrap your LLM call

Swap your OpenAI or Anthropic call for the Cognatu SDK. JavaScript, Python, and Go.

// before const res = await openai.chat(params); // after const res = await cognatu.chat(params);
Step 2

Topics match locally

A compact classifier runs on the device, in parallel with the LLM call — off the critical path.

// computed on-device, never sent topics: ["relocation", "travel"] matched: "flatio_monthly" conf: 0.91
Step 3

Labelled ad, real revenue

The winning ad is appended to the response, clearly marked. No confident match, no ad — ever.

// appended after the response role: "sponsored" content: "Flatio — Lisbon rentals" earned: $0.0151 // your 70%
04See it running

Don't take the claim on faith. Watch it.

The live demo runs the matcher in your browser — open devtools and watch the network tab stay silent.
The demo film walks through the publisher dashboard, auction, and payout flow end to end.
Everything shown is the real product, not a mockup.
05Privacy

Not a policy. An architecture.

Most networks promise not to misuse conversations they collect. We designed a system that never collects them — so there is nothing to misuse, leak, or subpoena.

USER'S DEVICE Conversation text Identity & history On-device classifier topic: "travel" Everything above the pill stays here. Permanently. BOUNDARY × × COGNATU AUCTION Bids on the topic Labelled ad back What we receive: a topic category. What we can't receive: anything else.

Matching runs on the device

A compact classifier reads the conversation locally and outputs a broad category — "career", "travel", "wellness". The transcript is never transmitted, to us or anyone.

The auction sees a topic, not a person

Advertisers bid on moments — "someone is discussing relocation" — never on identities. No email, device ID, or conversation text enters the auction.

Nothing stored, nothing to breach

Our servers never hold a transcript, so there is no conversation database to hack, leak, or hand over. You can't lose what you never had.

Built for trust-sensitive apps

Health, relationships, money, faith — the categories where AI apps thrive are exactly the ones where server-side prompt reading is untenable. That's who we built this for.

06For publishers

Your conversations are worth more than banners ever paid.

Contextual demand in premium verticals, a published revenue split, and an SDK that respects both your users and your latency budget.

70% of every impression — the split is public and it doesn't change.
No match means no ad — nothing generic is ever shown as filler.
No minimums, no exclusivity — monthly payouts, leave anytime.
Dating & relationships$28–35 CPM
Wellness & mental health$22–30 CPM
Career & personal finance$20–28 CPM
Beauty & lifestyle$18–26 CPM
Travel & experiences$16–24 CPM
Your share, every impression70%
current demand ranges across active verticals · second-price auction
Targeting signalAnonymous topic category
PlacementNative, labelled, in-conversation
PricingCPM · CPC · CPA
AuctionSecond-price, real-time
PII in the auctionNone, by construction
Campaign setupSelf-serve, minutes
07For advertisers

Reach people while they're deciding — not weeks after.

Someone asking an AI about moving abroad is telling you exactly what they need next. Cognatu lets you be the relevant answer in that moment, without touching their data.

08Questions

Asked often, answered plainly.

How long does integration actually take?+

Under four hours for most apps. You replace your LLM call with a Cognatu SDK call — the SDK proxies the request, runs topic matching on-device, and injects the ad into the response stream. Your prompt, model, system message, and response format don't change.

SDKs are available for JavaScript/TypeScript, Python, and Go. Any app using an OpenAI-compatible API integrates with one function call.

Does this affect response quality or latency?+

No effect on quality — the ad is appended after the LLM's response, never inserted into your prompt or system message. Topic matching runs locally, in parallel with the LLM call, so it stays off your critical path.

What happens when there's no relevant ad?+

Nothing is shown. Cognatu only injects an ad when topic match confidence exceeds a threshold you configure. There's no fallback to generic or irrelevant ads — an empty result is always better than a mismatch that breaks the conversation.

How can you target ads without reading the conversation?+

A compact model runs entirely on the user's device and classifies the conversation into a broad topic category, locally. Only that category — never the conversation, and never anything that identifies the user — is used to select an ad. We walk publishers through the full architecture during integration.

How is Cognatu different from other AI ad networks?+

Most AI ad networks send conversation text to their servers to pick an ad. Cognatu's matching runs on the user's device, so our servers never receive the conversation — there is nothing to read, store, leak, or subpoena.

We also publish our revenue share: publishers keep 70% of every impression, with no minimums and no exclusivity.

What ad categories are available?+

Dating & relationships, wellness & mental health, beauty & lifestyle, career & professional development, personal finance & fintech, travel & experiences, meditation & mindfulness, fitness & nutrition, education, and subscription products. Reach out if your category isn't listed — we add new mappings quickly.

What does the ad look like to the user?+

A clearly labelled "Sponsored" message appended after the AI response — native to chat, not a banner. It reads as a short, relevant suggestion. Publishers customise the styling to match their app's design system.

Early access

We're onboarding the first cohort by hand.

Every early publisher gets a direct line to the team, integration help, and a say in what we build next. Request access and we'll review it personally.