Custom Segment Builds, Tested Against a Holdout

Custom segments are cheap to build and easy to be wrong about, so the deliverable here is not a list of segments. It is a decision on which of them are worth continuing to target, supported by a holdout. On the accounts we have tested this way, between a third and half of hand-built custom segments performed no better than the campaign's own delivery once a control was in place. That is not a failure of segment craft. It is what optimised delivery does now, and knowing which half is which is the entire value of the exercise. What arrives: A built and live set of segments in your account, a written note on each saying what evidence it was built from and what it is meant to catch, the test design, and after the test a single page saying keep, narrow or delete. Everything sits in your account, in your name, and survives us leaving. How segments get built: Search terms taken from your own query reports and from the questions your sales or recruiting team actually get asked, not from a keyword tool's related list. URLs of the sites your buyers browse before they arrive, evidenced from analytics referrals and from a short round of customer interviews where you can arrange them. Six conversations beat a topic model here, consistently. App identifiers where the category is app-mediated. For DV360 we build custom affinity from URL sets. For Google Ads, custom segments and data segments, with a written note on which are eligible for which campaign type, because that eligibility has changed twice in two years and stale segment plans are the most common thing we inherit. Consent, and the parts people skip: For anything using your own data we ask for the consent record before we ask for the file. In the EEA, Customer Match and remarketing require Consent Mode v2 signals, and a CMP emitting a valid TCF string is the difference between a segment that serves and one that silently does not. We check the CMP output ourselves rather than take the vendor's word for it, and we will not upload a customer list where the legal basis for advertising use cannot be shown. That refusal has cost us work and we intend to keep it. The test: A user-split experiment where the account supports one, a geo holdout where it does not. Same creative, same bids, same budget shape, with targeting the only difference. Two to four weeks depending on conversion volume, with the minimum detectable effect stated before launch so nobody argues about it afterwards. Result read as conversions per thousand eligible users rather than as CPA, because a lower CPA on a smaller audience is the easiest illusion in this discipline and the one most often presented as a win. Refused: Third-party audience data whose provenance cannot be traced back to a consent chain. Seed-based builds from lists too small to be anything but noise. And a quarterly segment refresh sold as a subscription, when the honest answer is that a good segment set needs revisiting about twice a year and left alone in between. Not for: Accounts with fewer than roughly 200 conversions a month, where no holdout resolves inside a sensible window and the better recommendation is to spend the money on the offer instead. Advertisers who want the list handed over with no test attached. That work exists and other people will take it, and the segments will look excellent in a deck indefinitely.