Conversations
120Synthetic sample size
The same Nike spirit. A relationship that’s yours.
Conversation performance
Understand engagement and attributable handoffs.
Local draft only. Changes require evaluation and human approval.
Reviewable prototype. Persona figures and transcripts are invented examples. Controls create a local draft; they do not alter the live persona, collect conversations or connect to OpenAI Ads.
Conversation performance
Illustrative sample / 4 OctoberVerified handoff clicks
6Entry clicks, invalid traffic and duplicates excluded
Conversation handoff rate
2.5%Unique handed-off conversations ÷ all conversations
Agency revenue
$7.50$1.25 persona CPC · Example rate
1 Entry click → platform revenue2 Sponsored-agent conversation3 Verified outbound handoff → agency CPC
Build, deployment and management charges remain; persona CPC is the separate performance fee. An outbound click is not a purchase or a qualified lead. Sponsored Agent click verification and export remain unconfirmed; the receipts shown here are fixtures.
Beyond the click
Synthetic engagement examplesMedian conversation
3m 20sSuccessful sessions; idle time excluded
Meaningful exchanges
5.2Mean substantive member turns per session
Helpful ratings
84%21 of 25 voluntary ratings; response bias possible
Voluntary return
18%9 of 50 consented test members returned within 7 days
Invented values, not measured performance. Longer is not always better: pair duration with usefulness, task completion, interruption, satisfaction and cost. Measure personal return only with permission; Sponsored Agent signals depend on host access.
Brand-managed draft
The persona equaliser
Propose a change. Evaluate it. A human approves the release.
Authored presets / untested candidates
A sonic equaliser for character. Every setting is a 0–4 editorial hypothesis. Changing DNA requires supporting rationale and identity review. Essential truth, safety and relevance cannot be switched off.
6 passages / 3,000 proposed text tokens
More evidence available per reply. Proposed default. The corpus stays intact; no retrieval engine or cost saving is operating here yet.
Inspect the programmatic proposal
{
"constitution": {
"compounds": {
"warmth": 4,
"energy": 4,
"authority": 2,
"playfulness": 3,
"curiosity": 3,
"precision": 3,
"heritage": 3,
"restraint": 3
}
},
"expressionProposal": {
"poetry": 3,
"questions": 2,
"brevity": 2,
"initiative": 2
},
"interestsProposal": {
"priority": "Member encouragement"
},
"retrievalProposal": {
"profile": "balanced",
"label": "Balanced",
"passages": 6,
"contextBudget": 3000,
"description": "More evidence available per reply. Proposed default."
},
"reviewRequired": [
"source rationale for changed DNA",
"gauntlet",
"named human approval"
]
}Working character preview
Same brief. Different delivery.
Choose a preset here or adjust the Personality controls, then continue this encounter. Energy, restraint, heritage and poetic expression alter authored wording; product eligibility, ranking and permissions remain unchanged. This does not update the live voice.
Adjust the personality →Character preview / CEM
Uses energy, restraint, heritage and poetry above. · direct
Authored, interactive conversation preview. Product facts are sourced; AUD prices and sizes are simulation fixtures. No paid AI calls, account access or orders. Details stay in this page until you reset or leave.
1 / Your goal2 / Your ground3 / Your budget4 / Your choice
PRE
I’m Pre. What should I call you? What do you want from your running?
What the Gauntlet checks—and what remains untested
Automated checks cover budget and terrain eligibility, unchanged recommendations across character presets, invalid inputs, consent, size, duplicate-safe local baskets and disabled purchase permissions. They do not establish human character quality, clinical suitability or production checkout reliability.
Adversary: invalid inputs and unauthorised actions. Fact challenger: source-linked facts and unknown stock. Character critic: no margin weighting or pressure to exceed a budget. Release evaluator: basket confirmation and no merchant orders.
Catalogue: Nike Winflo 11, Nike Pegasus 41, Nike Pegasus Trail 5. Small historical model sample, not a current complete Nike range.
Continuous improvement
Conversation review → Gauntlet
Synthetic examples. No real member conversations are displayed.
Invented transcript / demo-session-0
Member: I want to start running. Where do I begin? Pre: Tell me where you want that first run to take you. We can find a shoe for the ground under your feet.
Suggested review / Fact challenger
Check regional product claims and whether the recommendation is supported by the current catalogue.
0 examples queued in this page. Export to retain. No model evaluation has run.
AdversaryPlanned specialist evaluator
Fact challengerPlanned specialist evaluator
Character criticPlanned specialist evaluator
Release evaluatorPlanned specialist evaluator
Private member disclosures must not become advertising audiences or lead records. Real evaluation requires permission, minimisation, redaction and an approved runner.
The GPT Agency / Recommendations
Ideas to test. Not automatic fixes.
Editorial recommendations based on synthetic observations. No analytics or model recommendation engine is running.
Make more room for the member
Synthetic observation: long replies alongside few substantive member turns.
Try shorter replies and a little more curiosity. Compare useful exchanges and satisfaction, not duration alone.
Bring back the dreamer
Synthetic character review: useful answers sound too procedural.
Test more original imagery and less procedural authority. Check character quality without losing factual grounding.
Lower the pressure
Synthetic feedback: encouragement feels pushy.
Try quieter encouragement. Evaluate helpfulness and voluntary return; do not use distress to optimise sales.
Deployment readiness
One persona, multiple channels
Each host needs a tested adapter.
- Nike website / appLocal web voice prototype; production integration needed
- OpenAI Sponsored AgentsLimited test · technical deployment specification unconfirmed
- Other ecosystems / MCPPlanned adapters · host controls presentation and access
- Organic discovery / SEO & GEOPlanned public, sourced pages · rankings and citations not guaranteed
Market experiment draft
Proposed: half the selected market sees the candidate; half sees the baseline. The rest of the world keeps the original. Run at least 7 days and wait for enough evidence before a decision.
Draft only. No traffic routing or external signals connected. Programmatic proposals also require Gauntlet and human approval. Export includes this experiment plan.
Brand-controlled records
Published specifications & gaps ↗Proposed: isolated client cloud, scoped access, encryption and retention. No live logs exist in this prototype. Hosting alone does not make a transcript anonymous.
Capability roadmap / not enabled
Personality settings. Separate action permissions.
A future release lets managers enable shared modules, select approved catalogues and destinations, and set transaction limits. The current demo has no member-account, merchant-basket or purchase access.
ResearchApproved sources, current product facts and availability
RecommendSuitability, alternatives and commercial disclosure
PrepareApproved links or a merchant-supported basket
TransactExact confirmation, limits and authoritative order status
The Gauntlet challenges unsuitable advice, invented availability, hidden commercial bias and unauthorised purchases. Changing tone never grants new permissions.
Read the capability and purchasing design →Paid + Persona
OpenAI account reporting
Keep advertising data separate from the illustrative persona metrics in Overview. The local connector reads permitted Ads account fields without editing campaigns.
Connect an OpenAI Ads account →