Most ad systems explain the past. ClickSage predicts the next decision — which click to buy, which offer to route, which install is fraud, when to scale — before the spend happens, and acts on it automatically. 大多数广告系统只会解释过去。ClickSage 预测下一个决策——买哪条 click、路由哪个 offer、哪条 install 是欺诈、何时放量——在花钱之前给出答案,并自动执行。
Privacy-first attribution broke the feedback loop. By the time a dashboard tells you a cohort was fraud or a campaign was unprofitable, the money is already gone. Three structural problems define the era: 隐私优先的归因打断了反馈闭环。当报表告诉你某批量级是欺诈、某条 campaign 不赚钱时,钱早已花出去了。这个时代有三个结构性问题:
Delayed attribution means decisions are corrected days late — after spend.归因延迟,决策晚几天才被修正——钱已花出。
User-level identifiers are gone; rule-based systems fly blind.用户级标识消失,基于规则的系统在盲飞。
A wrong bid or a fraudulent scale-up compounds before anyone sees it.一次错误出价或欺诈放量,在被发现前就已复利放大。
Predict before bidding, before allocation, before scaling.在出价前、分配前、放量前就预测。
A score per decision, not a static if/else table that ages.为每个决策给出概率分,而非会过时的静态规则表。
Dashboards explain; ClickSage acts — bid, throttle, drop, scale.大盘只解释;ClickSage 会行动——出价、限流、丢弃、放量。
ClickSage is not a slideware product. It runs inside a live programmatic network that decides 8 billion clicks a day. We run it in our own production first; the decision logs that fall out become the training data — and the product we open to partners. ClickSage 不是 PPT 产品。它运行在一个每天决策 80 亿 click 的真实程序化网络里。我们先在自有生产环境中长期打磨;沉淀下来的决策日志,既是训练数据,也是面向合作伙伴的产品。
Guiding principle: deploy in-house first, prove the flywheel in production, then productize it for partners. 指导原则:先自有部署、在生产中跑通飞轮,再面向合作伙伴产品化。
What device requested what, when, from where — the upstream truth, server-to-server.什么设备、何时、何地发起了什么请求——server-to-server 的上游真相。
Industry-wide, multi-platform install & reject-rate signal at scale.全行业、多平台的 install 与拒收率信号,规模量级。
Confirmed conversions reconciled across multiple sources & channels.跨多个来源与渠道对账的确认转化信息。
Multi-platform measurement players hold only the middle stream; point anti-fraud tools hold only fragments. Holding all three lets us compute a signal others can't — calibration drift: how far our reject rate diverges from the industry baseline, per cell. 多平台度量方只握有中间那一份,单点反作弊工具只有片段。同时握有三份,让我们能算出别人算不出的信号——校准漂移:每个 cell 上,我方拒收率与行业基准的偏离程度。
| Example cell示例 cell | Industry reject行业拒收 | Our reject我方拒收 | Drift | Reading解读 |
|---|---|---|---|---|
| BR · Banking · Samsung | 91% | 69% | +22pt | We're looser than the market → fraud leaking through比行业松 → 欺诈正在漏过 |
| US · App · Pixel | 5% | 45% | −40pt | We're stricter than the market → killing real users / lost revenue比行业严 → 误杀真用户 / 损失收入 |
~500K cells scored continuously. This per-cell drift is the core USP behind the products below.~50 万 cells 持续打分。这个 per-cell 漂移信号是下方所有产品的核心卖点。
Each module emits a probability; the orchestrator merges them with risk guardrails into a single executable action — bid, route, throttle or scale.每个模块输出一个概率;编排器结合风险护栏,把它们合成一个可执行动作——出价、路由、限流或放量。
Predict bid value before spend.花钱前预测出价价值。
Pick the best offer before routing traffic.路由流量前选出最佳 offer。
Detect fraud risk before scale & exposure.放量曝光前识别欺诈风险。
Auto cap, throttle & smooth during delivery.投放过程中自动调帽、限流、平滑 QPS。
Decide if scaling is safe before it happens.放量前判断是否安全。
Continuous learning from conversion, attribution & settlement signals.从转化、归因与结算信号中持续学习。
| Module | Endpoint | Consumer |
|---|---|---|
| Pre-Bid | /prebid/predict | DSP / RTB |
| Pre-Allocation | /preallocation/decide | ADX / Smartlink |
| Pre-Fraud | /fraud/score | Risk Engine |
| In-Flight | /inflight/control | Campaign Manager |
| Pre-Scale | /prescale/evaluate | Optimizer / AM |
Emitting only the top-scored half lifts effective conversion from 31% to ~50% while cutting emit cost in half. Modeled on current production data (CQM model AUC 0.96).只发分数前一半,有效转化率从 31% 升到 ~50%,同时 emit 成本减半。基于当前生产数据测算(CQM 模型 AUC 0.96)。
Audit validates product-market fit fastest; the Risk Score API is the core product shape; the dashboard is an upsell; predictive bidding has the highest barrier but the highest long-term ROI.审计最快验证 PMF;Risk Score API 是核心产品形态;大盘是增值;预测出价壁垒最高但长期 ROI 最大。
One-time audit of an advertiser's attribution & conversion data → a report: your reject rate vs industry, top fraud cells, fixes. Consulting; not scalable, but proves PMF with a case study.一次性审计广告主的归因与转化数据 → 出报告:你的拒收率 vs 行业、top fraud cells、改进建议。咨询性质,不可规模化,但能用 case study 验证 PMF。
In: (geo, appid, device, source, ip_asn…). Out: {fraud_probability, confidence, market_reject_rate, our_reject_rate}. Per-call or revenue-share on saved fraud budget.入:(geo, appid, device, source, ip_asn…)。出:{fraud_probability, confidence, market/our reject_rate}。按调用计费或按节省的 fraud 预算分成。
Per-cell install trend, reject-rate history, lifecycle stage. For advertisers (see where an app is bleeding) and anti-fraud vendors (buy the industry baseline).每 cell 的 install 趋势、拒收率历史、生命周期。卖给广告主(看 app 在哪掉量)和反作弊厂商(买行业基准)。
Give upstream SSPs a real-time "our bid willingness + expected CR" per ad request. They price better; we buy fewer bad clicks. Needs online inference + SSP integration.给上游 SSP 实时返回每个 ad request 的"我方愿出价 + 预期 CR"。他们定价更准,我们也少买烂 click。需在线推理 + SSP 对接。
Packaging — Basic: Pre-Bid + Pre-Fraud · Pro: + Pre-Allocation + In-Flight · Enterprise: + Pre-Scale + custom Decision Orchestrator.套餐 — Basic:Pre-Bid + Pre-Fraud · Pro:+ Pre-Allocation + In-Flight · Enterprise:+ Pre-Scale + 定制决策编排器。
| Player厂商 | Type类型 | Our angle我方角度 |
|---|---|---|
| Multi-platform measurement多平台度量方 | Measurement度量 | They measure; they don't decide clicks. We optimize on top of their data.他们做度量,不做 click 决策。我们在其数据之上做优化。 |
| Point anti-fraud vendors单点反作弊厂商 | Anti-fraud反作弊 | They see fragments (pre-bid OR post-attribution). We see the full chain.他们只看片段(pre-bid 或 post-attribution)。我们看全链路。 |
| Brand DSPs品牌 DSP | DSP | Brand-oriented, mature markets. We're perf-first, emerging + gray.品牌导向、成熟市场。我们效果优先、新兴+灰色。 |
| Walled-garden platforms封闭平台自建 | Platform平台自建 | They only see their own traffic. We see cross-platform, industry-wide.他们只看自家流量。我们看跨平台、行业级。 |
Thin team — near-zero sales / CS / frontend / compliance. → Self-serve API + light onboarding, no big-account sales.团队薄——销售/CS/前端/合规几乎为零。→ 自助 API + 轻量 onboarding,不做大客户全包。
Zero brand — B2B buyers haven't heard of us. → Win one Banking-fraud vertical, publish the case study.品牌零——B2B 客户没听过我们。→ 先打透一个 Banking 反作弊 vertical,发 case study。
Compliance gaps — GDPR / CCPA / LGPD barely studied. → Build EU/US baseline before entering those markets.合规弱——GDPR/CCPA/LGPD 几乎没研究。→ 进入欧美前先建基线。
Smaller scale — 1–2 orders below the largest measurement platforms in volume. → But sufficient per-cell; we compete on the join, not the size.规模较小——量级比头部度量平台低 1–2 个数量级。→ 但按 cell 维度足够;我们拼的是数据拼接,不是体量。
Engine integration; first real decision logs; model v5 on true negatives; Anti-Fraud Audit PoC #1 + case study.引擎接入;首批真实决策日志;模型 v5 用真负样本;反作弊审计 PoC #1 + case study。
Multi-task model (P_effective + P_fraud + P_repeat); Risk Score API internal beta with 3 free customers.多任务模型(P_effective + P_fraud + P_repeat);Risk Score API 内测,3 个免费客户。
Full market-signal suite live; 5% → 100% rollout with counterfactual monitoring; Risk Score API public + first paying customers.市场信号全套上线;5% → 100% 灰度,反事实监控;Risk Score API 公开 + 首批付费客户。
EU/US compliance baseline; Cell Health Dashboard MVP launches.欧美合规基线;Cell Health 大盘 MVP 上线。
First recurring revenue from data products; multi-tenant API at 99.9% SLA.数据产品 首笔经常性收入;多租户 API,99.9% SLA。
Partnering on anti-fraud, predictive bidding, or a data-product pilot?想在反作弊、预测出价或数据产品上合作试点?
[email protected]