Personalization engine
Two marketplace companies sell to thousands of venues and merchants a quarter. The facts that make an account worth a seller's time this month exist on public pages that no data vendor sells as a field. The engine scrapes them, forecasts from them, and writes from them.
What was actually hard
Templated outreach at marketplace volume converts like templated outreach. The alternative, a rep researching each account, does not survive the volume. The job was to make research a pipeline step.
In order
- Scrape live activity. each account's social posts and event calendar, on a schedule.
- Forecast. a 12-month GMV estimate from what the scrape shows, which sets priority and angle.
- Render variables. personalization rendered from the evidence, against a spec of what is worth saying and what the fallback is when it is missing.
- Write and send. multichannel sequences per account, at tens of thousands of emails a month.
- Wire the CRM. sends, replies and pipeline land in the client's CRM, so the motion is measured in pipeline, not opens.
Decisions that made it work
- Personalization is a data problem before it is a copy problem.The variable spec decides what gets scraped. Copy that references a fact the scrape cannot reliably produce is a fallback waiting to embarrass a seller.
- Forecast GMV sets the angle, not just the greeting.A venue with a full calendar and a growing audience gets a different pitch and a different priority than one with neither.
- Built to outlive its first model.The engine has run through two years of model changes, vendor changes and seller turnover, and is the basis of outreach for multiple sellers at both companies today.
Outcomes
6x lift in conversion against the previous motion.
Thousands of qualified accounts a quarter, in production since 2024 for two GTM teams with multiple sellers.
Production since 2024, not a demo. Ask for the pipeline numbers on a call.
What stays off this page. The clients, the sources and the scoring stay off this page.