Third-Party Segment Vetting, Testing and Selection

Most third-party segments fail one test: nobody can tell you what event put a person into them, or how long ago. Ask a provider for the collection method and the recency window in writing and a large part of any shortlist drops out before a single test runs. What survives gets bought at a data fee layered on top of media, and that fee has to be earned against a holdout. In our categories most segments do not earn it, and the ones that do are rarely the expensive ones. The market changed and the plans did not: Oracle closed its advertising business in September 2024, taking a large body of syndicated segments out of the market with it, and media plans we are handed still name segments sourced from it. In the same period, seller-defined audiences, built by the publisher from its own logged-in data and passed in the bid request against the IAB Tech Lab taxonomy, became the more honest supply in exactly the categories we work in, because the publisher can say where the signal came from and is accountable for it. How we select: Provenance first: declared, inferred or modelled, and from what behaviour. Then recency and refresh cadence, because a segment refreshed quarterly is a different product from one refreshed daily and is often sold at the same price. Then overlap with your own converters, measured rather than assumed, since a segment that mostly contains your existing customers is not evidence that it finds new ones. Then deliverable scale inside your own supply path, which is always smaller than the provider's headline audience because a segment only exists where its data is present in the bid stream you actually buy. Price comes last. Testing: Segments are tested against a holdout, not against each other's click rates. Where the campaign is regional we prefer a geographic holdout; where it is not, a matched control. Each test runs to a read date agreed in advance, and segments that do not clear their data fee get switched off, which is the entire purpose of running the test. Providers we work with regularly include LiveRamp's marketplace, Eyeota, Bombora for business intent and Semasio, alongside publisher-direct audiences. We hold no reseller relationship with any of them and take no rebate. Where third-party data is the wrong answer: An industrial manufacturer selling into a few thousand accounts in Israel has no third-party segment containing enough of those accounts to matter, whatever the audience size estimate says. The working answer there is the trade publisher's own audience plus a first-party list matched through an identity provider. We will say that rather than sell a test we expect to fail. The same is often true in property, where postcode, life stage and behaviour on your own listings outperform any purchased intent segment. What we refuse: Segments whose provider will not disclose collection method and recency in writing. Visitation segments built on panels too small to support the claim being made from them. Business intent data resold with the original source stripped out. Anything priced as a share of media spend. Modelled lookalikes presented as observed behaviour, where the model may well be sound but the label is not. Who this is not for: Teams looking for a segment list to hand to a trading desk without testing it. Campaigns too small to support a holdout, where the honest recommendation is contextual and supply-side targeting instead. What you receive: A shortlist carrying provenance, recency, measured overlap and deliverable scale for each candidate, the test design with its read dates, the results with a keep or drop decision on each, and the surviving taxonomy written into your DSP so the next planner inherits the reasoning rather than a list of names.