Lead Generation & GTM Engineering

A waterfall chains multiple data providers in order — Provider A, then B, then C, then D — with each one filling gaps the previous couldn't cover. A well-built waterfall typically lifts email and firmographic coverage from around 20% to over 80%.

Clay is the reference orchestrator for this pattern. Explorium and ZoomInfo GTM.AI extend it with server-side batch processing.

The cost logic is important: always put the cheapest provider first and the most accurate (and most expensive) last. You only spend premium credits on records the cheap providers couldn't complete.

The reference implementation here is the aaronkao/clay-outreach-bot pattern, which combines Pinecone (a vector database for retrieval) with Clay. The idea: retrieve something real that the prospect actually published — a post, a talk, a press release — and ground the opening line of your outreach in that specific observation.

Never let the model invent specifics about a person. Ungrounded personalization ('I loved your recent work at {company}') is instantly recognizable as AI-generated and performs worse than a plain, honest cold email. Grounding is what makes AI outreach actually work in 2026.

Before you scale any play, run 50 prospects through it over 2 weeks and log everything: send rate, reply rate, positive-reply rate, meeting-booked rate, disqualification rate, and the reasons people disqualified.

That dataset is your baseline. Every future iteration — new signal, new message, new enrichment step — must beat it with evidence. Skip the pilot and you have no baseline, which means you have no way to tell if a change is an improvement or a regression.

After a lead hits your CRM and gets a disposition (qualified, junk, no-reply), that outcome needs to flow back into three places.

First, the suppression list — so that person is never contacted again. Second, the ad platforms — a retraction upload tells Google and Meta to stop treating the junk-lead conversion as a positive signal and stop finding more people who look like it. Third, your enrichment model — so you can correct the data or rules that let the junk lead through in the first place.

Closing this loop is the structural differentiator almost no agency ships. It's also what stops a broad-targeting campaign from eating its own garbage.