Metlok sells premium metal roofs to Florida homeowners. It needed landing pages its ads could send people to, a quote request more visitors would finish, and a clear picture of which ads, pages and offers produced real leads.
We built that system and have run it since launch: the pages, the testing, the lead capture and the data behind it.
Live in three days.
Three days after the brief, the first pages were live: four page designs and four audiences—premium buyers, storm damage, homeowners burned by a past contractor, and builders—across Orlando, Tampa and Jacksonville. Each one asks for a quote through a short step-by-step form instead of a long contact form.
The system has grown since. A monthly offer page changes in place without breaking a single ad link. Direct-mail recipients get their own page and tracking number. Panhandle homeowners rebuilding after a hurricane get pages written for them.
Tested properly, not by gut feel.
Every visitor is randomly assigned a page version on our server, and every result is read against rules written down before the test starts: how many visitors each version needs, how long it runs and how sure we have to be before calling a winner. Across the program we have tested four page designs, dozens of headline and page versions, and details like the call button and the first question on the form.
That discipline showed most in what we didn’t ship. Two early “winners” looked dramatic after three days and faded once the full data came in. One test seemed to show the call button driving calls, until we found most of those taps came from outside Florida. We added location to every event, cut the out-of-market traffic and kept the decisions honest.
Catching the leads that almost got away.
Plenty of people start a quote request and stop at the contact step. When someone has typed their details and then leaves, the form keeps what they entered, as Metlok’s privacy policy discloses.
Those visitors are never emailed or texted. The team gets a call-only follow-up instead, flagged for manual dialing, showing whether the visitor ticked the consent box and reminding the caller to check the Do Not Call registry first.
It surfaces roughly one extra person to call for every three completed requests.
Spend moved to what converts.
Because every lead carries its source, the channel picture became clear quickly. Within the first weeks, the strongest channels were turning visitors into requests many times better than the weakest. One ad network sent thousands of visitors and almost no requests, so it was switched off, and out-of-market traffic from broad campaigns was cut as well.
Data piped into the systems they already use.
Leads land in Metlok’s own database and reach the sales inbox within seconds, flagged when they come from outside the service area. Every visit and form step flows to their analytics and ad platforms, tagged with the page version and city, so the ads learn from real requests.
Call tracking numbers separate phone leads by source, direct mail included. Our attribution work joined their CRM records back to each lead’s ad, page and offer, and the team gets regular reports and a call list.
More than twice the requests per visitor.
In the first month of paid traffic, 0.48% of visitors asked for a quote. Over the last 30 days it was 1.11%: 2.3 times as many requests from every thousand visitors. So far this month it is 1.21%.
Completed quote requests divided by unique visitors, bots and test traffic removed. Month 1 is the first month of paid traffic.
No single test produced that. Better pages, monthly offers and a channel mix tuned to what converts all moved together, and the testing kept every change honest.
An agent that makes the call.
During the main testing period, an AI agent ran the program every morning. It checked that the data was flowing and the site was healthy, confirmed visitors were being split evenly, read every test against the locked rules and then decided what to do.
It was allowed to make one change on its own: take a clearly losing page version out of rotation. A version had to lose twice, days apart and with high confidence, before it was cut. Then the agent deployed the change, checked the live site and rolled back automatically if anything looked wrong. Every decision went into a ledger with its evidence, and any cut could be reversed.
We rehearsed that loop with test cuts before letting it run. On real data, no version ever lost clearly enough to cut. That restraint is the point: the pages changed only when the evidence did.
Currently implementing full ServiceTitan automation.
Metlok is moving its sales operation to ServiceTitan, and we are currently implementing full automation between the website and it: every lead pushed straight in with its source, city and page version attached, routed to the right person, and the outcome fed back so reporting moves from requests to booked jobs by channel.