跳到主内容
@wquguru
精选72Trends.vc(Dru Riley · RSS)独立开发与小生意

合成受众市场:玩家、风险与独立开发者机会

Synthetic Audiences: Verified Humans, Panel Fraud, Consent Gap

原文
发到 X

“Public opinion is no more than this, What people think that other people think.” – Alfred Austin

“舆论不过是:人们认为别人所想的。”——阿尔弗雷德·奥斯汀

Get Full Access to Trends Pro

获取 Trends Pro 完整访问权限

What You’ll Learn

你将学到什么

  • How much of a bought panel is fraud?
  • Which data behind a persona predicts human answers?
  • Who is quietly becoming the supplier to the vendors selling simulated people?
  • Which seat in this market is still empty?
  • What must be true before a board trusts a simulated answer?
  • 购买的样本中有多少是欺诈?
  • 人格画像背后的哪些数据能预测人类的回答?
  • 谁正在悄然成为向销售模拟人群者供货的供应商?
  • 这个市场中哪个席位仍然空缺?
  • 在董事会信任模拟答案之前,必须满足什么条件?

Why It Matters

为什么这很重要

A simulated person now takes the first pass at the survey a real one used to be paid for.

如今,模拟人开始接手原本由真人付费完成的调查初稿工作。

Problem

问题

You need to know how strangers react before you spend the budget.

在你投入预算之前,你需要知道陌生人会作何反应。

So you buy a panel and wait 3 months.

于是你购买一个样本库并等待三个月。

Solution

解决方案

Research firms and startups build AI personas grounded in real human data and survey them like a panel.

研究机构和初创公司构建基于真实人类数据的 AI 人格画像,并将其作为样本库进行调查。

Players

玩家

Synthetic Audience Tools

合成受众工具

  • Simile • The Stanford lab whose method every rival argues against, now selling it.
  • Evidenza • Bootstrapped to 7 figures in year one with no venture round.
  • Artificial Societies • The one tool simulating a social graph, so a message travels a crowd.
  • Simile • 斯坦福实验室,其方法遭到所有竞争对手的反对,如今却开始出售该技术。
  • Evidenza • 第一年实现七位数收入,未进行风险投资融资。
  • Artificial Societies • 唯一一款模拟社交图谱的工具,使信息能够在人群中传播。

AI Interview Tools

AI 访谈工具

  • Listen Labs • Holds the largest interview corpus and says simulation is next.
  • Outset • Ships Digital Twins beside its interviews, so the crossover already happened.
  • Quantilope • The panel purist that converted, launching twins from client tracking data.
  • Listen Labs • 拥有最大的访谈语料库,并表示模拟将是下一步。
  • Outset • 在其访谈服务旁推出数字孪生,跨界融合已然发生。
  • Quantilope • 坚持纯样本库理念的公司,现已转型,从客户追踪数据中推出数字孪生。

Market Research Firms

市场研究公司

  • Qualtrics • Sells synthetic panels inside the survey tool enterprises already own.
  • YouGov • The only vendor shipping validation as a product step.
  • Gallup • Sells no synthetic product and says so publicly, the most useful entry here.
  • Qualtrics • 在企业已拥有的调查工具内部销售合成面板。
  • YouGov • 唯一一家将验证作为产品步骤交付的供应商。
  • Gallup • 不销售任何合成产品,并公开声明这一点,是此处最有用的条目。

Ad Testing Tools

广告测试工具

  • Kantar • The enterprise default, scoring ads from human data alone.
  • System1 • Captures real emotional reaction on camera, marketed against simulation.
  • Zappi • Puts synthetic respondents at the front of the funnel only.
  • Kantar • 企业默认选择,仅基于人类数据对广告进行评分。
  • System1 • 通过摄像机捕捉真实的情感反应,主打对抗模拟。
  • Zappi • 仅在漏斗前端放置合成受访者。

Predictions

预测

  • AI interview firms will become the main suppliers of synthetic audiences.
  • Listen Labs has run over 1 million interviews and points that corpus at simulation.
  • Quantilope launched twins from client tracking after years selling panels.
  • Persona tools under $100 a month will lose buyers to assistants.
  • Anyone can prompt a frontier model to react as 30 suburban parents, free.
  • GWI ships an MCP connector into ChatGPT, treating the assistant as the channel.
  • AI 访谈公司将成为合成受众的主要供应商。
  • Listen Labs 已运行超过 100 万次访谈,并将该语料库用于模拟。
  • Quantilope 在多年销售面板后,推出了源自客户追踪的双生体。
  • 每月售价低于 100 美元的用户画像工具将被助手类应用抢走买家。
  • 任何人都可以免费提示前沿模型,使其以 30 位郊区父母的身份做出反应。
  • GWI 向 ChatGPT 推送 MCP 连接器,将助手视为渠道。

Opportunities

机会

  • Run synthetic research done-for-you for teams with no researcher.
  • Evidenza is the template: bootstrapped, no venture round, profitable.
  • The buyer will not learn a research tool. They want an answer.
  • Own one regulated vertical end to end.
  • Seldon states the thesis: regulated industries can’t run fast user research.
  • Clinical protocol design, compliance messaging and public consultation qualify.
  • Launch the pre-post test for content.
  • Artificial Societies simulates networks of personas passing opinions to each other.
  • 2 young companies hold this job, on under $20M in funding.
  • Ship the synthetic-first ad pretest for performance marketers.
  • Kantar’s self-serve ad test starts at $4,500, pricing out anyone running social creative at volume.
  • DAIVID found 70% of Gen AI ads score below average on positive emotion.
  • 为没有研究员的团队提供代运营式合成研究服务。
  • Evidenza 是模板:自力更生,无风险投资轮次,实现盈利。
  • 买家不会去学习研究工具。他们想要的是一个答案。
  • 独占一个受监管垂直领域的端到端业务。
  • Seldon 阐述了这一论点:受监管行业无法快速开展用户研究。
  • 临床试验方案设计、合规性信息传达和公众咨询均符合条件。
  • 推出内容的前后测。
  • Artificial Societies 模拟了相互传递观点的用户画像网络。
  • 两家公司持有该职位,融资额均低于2000万美元。
  • 为绩效营销人员推出以合成数据为先的广告预测试功能。
  • 凯度(Kantar)的自助式广告测试起步价为4500美元,将大规模投放社交创意内容的用户排除在外。
  • DAIVID发现,70%的生成式AI广告在积极情绪评分上低于平均水平。

Risks

风险

  • Under-Dispersion • Expect the average back with the spread flattened, so you lose the minority that decides it.
  • Self-Graded Accuracy • Ask which baseline a percentage was measured against, because 2 vendor figures are not comparable.
  • Consent Gap • The people modeled never agreed and have no standing, the likeliest hard stop.
  • Prompt Sensitivity • Rerun the study on reworded prompts before you believe the finding.
  • Corpus Dependence • Your vendor licenses its interviews from a future competitor.
  • 低离散度 • 预期平均值回归且分布变平,因此你会失去那些决定性的少数群体。
  • 自我评分准确性 • 询问百分比是相对于哪个基线测量的,因为两家供应商的数据不可比。
  • 同意缺口 • 被建模的人从未同意且无法律地位,这很可能是最硬的停止点。
  • 提示词敏感性 • 在相信研究结果之前,先用改写后的提示词重新运行研究。
  • 语料库依赖 • 你的供应商从未来的竞争对手那里授权获取其访谈数据。

Key Lessons

关键教训

  • Ask what the personas are grounded in before you ask the price. You are buying a licensed corpus more than a model, which is why the biggest research firms rent theirs.
  • Sell the answer and keep the tooling invisible. Price against the decision your buyer has to defend, well above the tool underneath you.
  • 在询问价格之前,先问清楚这些人物画像基于什么。你购买的更多是一个授权语料库而非模型,这就是为什么最大的研究机构要租赁它们的原因。
  • 销售答案,让工具隐形。针对买家必须为其辩护的决策进行定价,远高于你下方的工具本身。

Hot Takes

热门观点

  • Every accuracy figure is a self-graded exam. The category is waiting for one adult to mark the papers.
  • 每个准确率数字都是一场自我评分的考试。该领域正等待一位成年人来批改试卷。

Haters

批评者

“You’re asking a model trained on the internet what normal people think. It’ll tell you what the internet thinks.”

“你是在问一个在互联网上训练出来的模型普通人的想法是什么。它会告诉你互联网是怎么想的。”

Fair. A plain model picks up what the internet sounds like. Santurkar and co-authors showed that the gap is real. Vendors try to close it by building personas from real interviews.

公平地说。基础模型捕捉的是互联网的声音。Santurkar及其合著者证明这种差距是真实存在的。供应商试图通过基于真实访谈构建人物画像来缩小这一差距。

“ChatGPT does this for free and doesn’t know it’s competing.”

“ChatGPT免费做到这一点,而且它不知道自己在竞争。”

That is the price floor for any tool with no proprietary data underneath. Ask what corpus a vendor holds that a chatbot cannot reach.

这是任何底层没有专有数据的工具的底价。询问供应商持有什么聊天机器人无法触及的语料库。

“Nobody bets a real budget on fake people.”

“没人会把真正的预算押在虚假人物身上。”

Evidenza reached 7 figures selling synthetic AI market research services to BlackRock. While Gallup will not publish estimates from simulated responses.

Evidenza向贝莱德(BlackRock)出售合成AI市场研究服务,达到了七位数销售额。而盖洛普(Gallup)不会发布来自模拟响应的估计值。

Links

链接

  • Synthetic Consumers: A Practical Guide • How to actually run one, plus where it breaks.
  • Fraud in Online Survey Research • Why a bought panel returns 15% to 30% fraud while detection loses ground.
  • Synthetic UGC Ads • Simulated humans on the ad’s output side. This covers the input side.
  • Prediction Markets • Polls against predictions, this accuracy argument in another market.
  • 合成消费者:实用指南 • 如何实际运行一个,以及它在何处失效。
  • 在线调查研究中的欺诈行为 • 为什么购买的样本库会产生15%到30%的欺诈率,而检测手段却节节败退。
  • 合成用户生成内容(UGC)广告 • 广告输出端的模拟人类。这也涵盖了输入端。
  • 预测市场 • 将民意调查与预测进行对比,这是另一个市场中的准确性论证。

What else?

还有什么?

Trends PRO #0180: Synthetic Audiences has more insights.

Trends PRO #0180:合成受众群体包含更多见解。

What you’ll get:

您将得到:

  • 25 Players (108% more)
  • 6 Predictions (200% more)
  • 8 Opportunities (60% more)
  • 7 Risks (40% more)
  • 4 Key Lessons (100% more)
  • 2 Hot Takes (100% more)
  • 8 Links (100% more)
  • 25个玩家(增加108%)
  • 6个预测(增加200%)
  • 8个机会(增加60%)
  • 7个风险(增加40%)
  • 4个关键教训(增加100%)
  • 2个热门观点(增加100%)
  • 8个链接(增加100%)

With Trends Pro you’ll learn:

通过 Trends Pro 您将学到:

  • ( Pro) Which acquisition proved that buyers here pay for 30 million people, never for the technique?
  • ( Pro) Which 15% to 30% fraud range has a verification product no vendor sells to rivals?
  • ( Pro) Which 12 Am Law 100 firms already buy simulated juries instead of mock trials?
  • ( Pro) What are you buying in the 10 points between a 74% demographic prompt and an 83% interview-grounded agent?
  • ( Pro) Which peer-reviewed finding shows the same prompt returning different answers 3 months apart?
  • ( Pro) Why will enterprise buyers require an independent benchmark before they purchase?
  • ( Pro) Why will the middle of this price range empty out while both ends grow?
  • And much more…
  • (Pro) 哪次收购证明了买家在此处是为3000万人付费,而非为技术付费?
  • (Pro) 哪个15%到30%的欺诈区间存在一种验证产品,且没有供应商向竞争对手出售?
  • (Pro) 哪些美国《法律500强》前100名的律所已经购买模拟陪审团,而不是进行模拟审判?
  • (Pro) 在74%的人口统计提示和83%基于访谈的智能体之间那10个点的差距中,您到底在购买什么?
  • (Pro) 哪项经过同行评审的研究显示,相同的提示在相隔3个月时返回了不同的答案?
  • (Pro) 为什么企业买家在购买前会要求独立的基准测试?
  • (Pro) 为什么这个价格区间的中间部分会萎缩,而两端却会增长?
  • 以及更多内容……

Get Weekly Reports

获取每周报告

Join 54,000+ founders and investors

加入54,000多位创始人和投资者

Subscribe

订阅

Unlock Pro Reports, 1:1 Intros and Masterminds

解锁 Pro 报告、1:1 创始人引荐和大师班

Become a Trends Pro Member and join 1,200+ founders enjoying…

成为 Trends Pro 会员,加入 1,200 多位正在享受…的创始人

Founder Mastermind Groups • To share goals, progress and solve problems together, each group is made up of 6 members who meet for 1 hour each Monday.

创始人大师班小组 • 为了共享目标、进展并共同解决问题,每个小组由 6 名成员组成,每周一聚会 1 小时。

160+ Trends Pro Reports • To make sense of new markets, ideas and business models, check out our research reports.

160+ Trends Pro 报告 • 为了理解新市场、想法和商业模式,请查阅我们的研究报告。

1:1 Founder Intros • Make new friends, share lessons and find ways to help each other. Keep life interesting by meeting new founders each week.

1:1 创始人引荐 • 结交新朋友,分享经验教训,寻找互助的方式。通过每周结识新创始人来保持生活的趣味性。

100k+ Startup Discounts • Get access to $100k+ in startup discounts on AWS, Twilio, Webflow, ClickUp and more.

10万+ 初创企业折扣 • 获取 AWS、Twilio、Webflow、ClickUp 等超过 10 万美元的初创企业折扣。

Brought to you by the team behind HeadsUp

由 HeadsUp 背后的团队呈现

The post Synthetic Audiences: Verified Humans, Panel Fraud, Consent Gap appeared first on Trends.vc.

本文《合成受众:验证真人、面板欺诈与同意缺口》首发于 Trends.vc。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

另一事件,读法相近