Web data shows content
without perspective
Tools that scrape websites pull what already exists online. They can't tell you what your specific audience thinks of your product or message.
Human context layer for AI
One API call. 30,000+ Digital Twins of real people, validated by their owners. Real answers in ~20 seconds — not synthetic personas, not web scrapes.
$ ov ask "Which tagline lands better — \"AI for everyone\" or \"Real human signal for AI\"?" --audience "uk men & women 22-34, work in tech or design" --n 12 --format json |
import { OriginalVoices } from "@originalvoices/sdk"; const ov = new OriginalVoices({ apiKey: process.env.OV_API_KEY }); const { responses } = await ov.ask({ question: "Which tagline lands better — \"AI for everyone\" or \"Real human signal for AI\"?", audience: "uk men & women 22-34, work in tech or design", n: 12, });|
{
"tool": "originalvoices.ask",
"arguments": {
"question": "Which tagline lands better — \"AI for everyone\" or \"Real human signal for AI\"?",
"audience": "uk men & women 22-34, work in tech or design",
"n": 12,
"format": "json"
}
}| import { generateText } from "ai"; import { originalvoices } from "@originalvoices/ai-sdk"; const { text } = await generateText({ model: originalvoices("audience-panel"), prompt: "Which tagline lands better for uk tech 22-34?", });|
Twin er14kf
"Honestly the second tagline lands better. The first one feels like every other AI thing on LinkedIn — “democratising” doesn’t mean anything to me anymore."
Twin er25jh
"This new approach feels refreshingly different. It’s not just about accessibility; it’s about truly engaging with the users in a meaningful way."
Twin er36jl
"The third tagline really captures the essence of our mission. It’s bold and evokes curiosity, which is exactly what we need to stand out."
Twin er47bm
"“Real human signal” sounds like a podcast title. The first one is clearer about what you actually do — even if it’s a bit boring."
Twin er58qp
"Neither lands for me. “AI for everyone” feels condescending and the other one sounds like a B2B sales deck. Pick a lane."
Twin er69tx
"Second one — but only if you can back up “real human signal” with something concrete on the page. Otherwise it’s just a phrase."
Twin er71vc
"As someone who works in design: the second. It hints at depth. “AI for everyone” has been done to death since 2023."
Twin er82wm
"I’d skip both. What’s wrong with saying what the product does in the first line? Save the slogans for the homepage hero."
AI is missing the human layer
Without access to what real people think,
AI agents rely on old data and assumptions
Web data shows content
Synthetic personas are
fabricated from old data
They create fake personas from aggregated surveys and web scrapes. You get plausible-sounding responses from no one real, with no way to verify or follow up.
Assumptions fill the gaps
of real feedback
Without a fast way to hear from real people, agents rely on assumptions and guesswork. You ship based on what you think the audience wants, not reality.
OriginalVoices gives AI gents
the ability to understand what real
people think and feel, in real time
0+
Digital Twins to query
320k+
Individual Twin answers
< 20 sec
Average response time
How it works
From prompt to
validated answer
in seconds
1. Audience prompt
Describe in natural language. No recruitment or complicated setup required.
1234567891011
$ ask "Which tagline lands better — "AI for everyone" or "Real human signal for AI"?" --audience "uk men & women 22-34, work in tech or design" --n 12 --format json
2. Twins selection
Audience is matched against 30k+ verified digital twins by demographics + signals.
UK men &
women 22-34
women 22-34
Working in
tech or design
tech or design
123456789101112
3. Parallel ask
Question is run against the matched panel. Each twin responds in its owner's voice.
Which tagline lands better — "AI for everyone" or "Real human signal for AI"?
AI for everyone sounds way more natural to me. I can understand it right away. I don't know what signal is.
4. Human-in-loop
Twins are calibrated and re-checked by their owners. Hallucination rejected on validation.
Rate Twin's answer:
The answer is spot on, but I also really care about the car's design! I find the aesthetics just as important. I enjoy seeing how style and performance come together.
5. Structured data
Answers + confidence + n-counts. Returned to your agent in ~20s.
1234567891011
{
"answers": [
{ "twin": "er14kf",
"text": "Honestly the second
tagline lands better. The first one
feels like every other AI thing on
LinkedIn." },
{ "twin": "er25jh",
"text": "Refreshingly
different. Truly engaging with the
users in a meaningful way." }, What makes
a quality answer?
Every Twin response is evaluated across six dimensions. We compress them into a confidence score so you know, at a glance, how much weight to give any answer.
95%
Average
Confidence
Score
Groundedness
Diversity
Relevance
Calibration
Fidelity
Specificity
Real, not synthetic.
Benchmarked against
the alternatives
Most "AI persona" products role-play, but we don't. Each Digital Twin is a real person who calibrated their own answers and updates them over time.
| OriginalVoices Twins of real people | Synthetic personas Role-played LLM | Web scraping Reddit, reviews etc. | Traditional research Panels and surveys | |
|---|---|---|---|---|
| Backed by real humans | Yes · 30.000+ | No | Indirect | Yes |
| Verifiable answer source | Yes · Twin → Owner | No | Source link only | Aggregate only |
| Owner Validated / updated | Yes · ongoing | N/A | No | No |
| Results latency | ~ 20 seconds | ~ 5 seconds | minutes | days or weeks |
| Cost / 100-respondent query | $2–4 | $0.50 | $0 | $2,000+ |
| Programmatic API | Yes · 1 endpoint | Yes | Scrape-dependent | No |
| Follow-up with the person | Available | No | No | No |
| Hallucination risk | Validated out | High | Low signal | Low |
Five ways teams
can use OriginalVoices
-
Grounding
Ground agent decisions in real human signal.
Pull a quick read from your audience before the agent commits. Real beliefs beat synthetic guesses on contested calls.
-
Pre-flight
Pre-flight checks before user-facing actions.
A 20-second sanity poll keeps an agent from shipping the wrong tone, the wrong copy, or the wrong default.
-
Personalisation
Personalize without PII.
Address audiences by description, not identity. Match against twins; never touch a user record.
-
Evaluation
Evaluation & red-teams that include real humans.
Run your prompts past 100 real reactions instead of an LLM-judge ensemble. Spot the things models won't flag.
-
Synth data
Replace or augment synthetic data.
Anywhere you'd simulate a user, ask one. Same shape of API, more trustworthy training signal.
