After
Real people's ChatGPT, Claude and Gemini conversations. Consented by them. Enriched for you.
Consent and compliance
Consented. Compliant. Real.
AI conversations.From the user.Enriched.
Scraped data is a legal risk you carry. Signal is real consumers’ own ChatGPT, Claude and Gemini conversations, shared by them with explicit consent, paid into their wallet, compliant on day one, and enriched before it reaches you.
Consent
What comes with every record.
record sg_7a3e · verified · consent v2 · US · 25–34
USER Choosing running shoes for a half marathon, $150.
AI [suggests two models]
USER The first hurt my knees. What else?
AI [revises recommendation]
USER Ordering the second. Build me a 10-week plan.
Click the record →
- A real personVerified adult. Not a bot, not a contractor.
- Explicit consentGiven by the user for this data, versioned and timestamped, attached to the record.
- Paid into their walletThe more they consent to share, the more they earn.
- De-identifiedNo PII in what you receive.
- Revocation handledWe manage withdrawals; your dataset stays clean.
- Fully reconciledEvery record ties to one consent, one payment, one export.
How we build it
01
Consent
The user opts in, per data type, and is paid into their wallet.
02
Collect
Their full conversation history, branches and all, then every new conversation.
03
Enrich
De-identified, one schema, labelled, with the person behind it.
04
Deliver
Datasets from 50,000 people a month, as schema’d JSONL.
$ signal consent --member m_82c1
identity verified adult · US · payout history 14 mo
scope ai_conversations consent v2
wallet credited receipt rcpt_5a9e
reconciled consent ↔ payment ↔ export
audit logged 2026-11-04T18:12:09Z
ready to collect
$ signal collect --member m_82c1 --source chatgpt
conversations 1,284 11 months of history
nodes 41,906 branches 2,310 · edits 1,077
tool traces web.run · python voice 212 · images 88
sync 51s continuous from here
collected
$ signal enrich --member m_82c1
schema signal/v2 chatgpt → common tree
de-identify done names 37 · emails 4 · third parties 12
labels intent · correction · outcome 1,284 / 1,284
person 25–34 · US · tenure 14 mo joined at person level
behaviour gpt-5 → claude switch detected, Oct
dataset ready
$ signal deliver --client lab_01 --panel us_25_34
format jsonl schema signal/v2
contributors 50,000 / month refreshed weekly
consent attached per record
revocations applied 0 this cycle
delivered
Enrichment
Raw export vs Signal record.
Same conversation. This is the work we do before it reaches you.
{"title":"Running shoes","create_time":1761590412.33,"mapping":{"c0f1":{"id":"c0f1","parent":null,"children":["9b2e"]},"9b2e":{"id":"9b2e","parent":"c0f1","children":["a7d3","e4c8"],"message":{"author":{"role":"user"},"create_time":1761590412.33,"content":{"content_type":"text","parts":["Hey I'm Marco, choosing running shoes for a half marathon, budget $150"]},"metadata":{}}},"a7d3":{"id":"a7d3","parent":"9b2e","children":["f0a1"],"message":{"author":{"role":"assistant"},"content":{"content_type":"text","parts":["Here are two solid options..."]},"metadata":{"model_slug":"gpt-5","finish_details":{"type":"stop"}}}},"e4c8":{"id":"e4c8","parent":"9b2e","children":[],"message":{"author":{"role":"assistant"},"content":{"content_type":"text","parts":["A few things to consider first..."]},"metadata":{"model_slug":"gpt-5"}}},"f0a1":{"id":"f0a1","parent":"a7d3","children":["b55d"],"message":{"author":{"role":"user"},"content":{"content_type":"text","parts":["the first one hurt my knees last time, what else?"]}}},"b55d":{"id":"b55d","parent":"f0a1","children":["d9e2"],"message":{"author":{"role":"tool"},"content":{"content_type":"tether_browsing_display","result":"..."},"metadata":{"model_slug":"gpt-5"}}},"d9e2":{"id":"d9e2","parent":"b55d","children":["77aa"],"message":{"author":{"role":"assistant"},"content":{"content_type":"text","parts":["Based on current reviews..."]}}},"77aa":{"id":"77aa","parent":"d9e2","children":[],"message":{"author":{"role":"user"},"content":{"content_type":"text","parts":["ok ordering the second one, build me a 10 week plan"]}}}},"moderation_results":[],"current_node":"77aa","conversation_id":"68f2...","is_archived":false}
record sg_7a3e platform chatgpt model gpt-5 started 2026-10-27
person p_41c9 · verified adult · 25–34 · US · 14 mo on panel
consent v2 · ai_conversations · 2026-11-04 paid yes
tree 8 nodes · 1 regeneration (e4c8 abandoned) · 1 tool call (web browse) · linear path resolved
turns user 3 · assistant 3 · tool 1
de-id "Hey I'm Marco" → "Hey I'm [NAME]" · 1 name removed · 0 remaining
intent purchase domain sport / footwear budget $150
correction 1 · rejected first suggestion · reason: knee pain
outcome decision · bought option 2 · follow-on task: training plan
person history 1,284 conversations · 3 purchase intents / 30 d · uses claude since Oct
person p_41c9 · verified adult · 25–34 · US · 14 mo on panel
consent v2 · ai_conversations · 2026-11-04 paid yes
tree 8 nodes · 1 regeneration (e4c8 abandoned) · 1 tool call (web browse) · linear path resolved
turns user 3 · assistant 3 · tool 1
de-id "Hey I'm Marco" → "Hey I'm [NAME]" · 1 name removed · 0 remaining
intent purchase domain sport / footwear budget $150
correction 1 · rejected first suggestion · reason: knee pain
outcome decision · bought option 2 · follow-on task: training plan
person history 1,284 conversations · 3 purchase intents / 30 d · uses claude since Oct
Sample: 100 conversations61 people10,147 nodes~1.1M tokens22 voicetool traces: web, pythonSept–Nov 2026
The panel
Real consumers. Current behaviour.
11 mo
of history per contributor
0
have more than a year
0
active in the last 30 days
2/3
of conversations from 2026
Full
trees, files, voice. Not samples
<1 min
sync, then continuous
50k
people per dataset, per month, to start
Live
since 20 October 2026
Where it goes
One verified person. Every kind of data.
AI conversations first. Then everything else that person consents to share, joined at person level.
ChatGPT liveFull history and every new conversation
Claude nextSame schema, same person
Gemini nextSame schema, same person
Browsing 2027What they searched before and after the conversation
One verified personconsented · paid · de-identified
Shopping 2027Did the purchase intent become a purchase
Email 2027Receipts, bookings, confirmations
Finance 2027Spend categories, outcomes of decisions
Apps & devices 2027Time, context, platform switching
Consent is per data typeJoined at person levelSame wallet, more consent, more earned
Get a sample
Run it through your eval.
100 consented conversations, the schema, the consent documentation. Datasets start at 50,000 people a month.