Audience Research Where Broad Targeting Stops Working

Targeting research earns its fee below a threshold and wastes money above it. Meta's optimizer needs roughly fifty conversions per ad set per week to leave the learning phase. Above that line, broad delivery beats almost any stack you can assemble by hand. Below it, in a single Brazilian state, a LATAM country of twelve million people, a product closing eleven deals a quarter, the optimizer has nothing to pattern-match on and the audience you construct is the only real signal in the auction. Most of our clients live below the line, which is the only reason this engagement exists. Where the research comes from: Not the interest browser. We start with recorded sales calls and closed-won interviews, because the words a buyer uses about his own problem are the only reliable input, and they are already sitting in your CRM as recordings nobody has listened to twice. Then review text about your competitors on Reclame Aqui and Google, which tells you what the market is angry about and what it is willing to switch for. Then your own WhatsApp threads and site search terms. From twenty to thirty hours of that material we return with the three or four demand patterns that repeat, and only then do we open a targeting tool. Building the stacks: Interests are grouped into stacks large enough to exit learning and coherent enough to read separately, usually four to six per market. Lookalike seeds are built on CRM stage, not on form submission: a one percent lookalike of everyone who filled a form is a lookalike of people who fill forms. For industrial accounts we build the company side from CNAE codes and CNPJ size bands, matched into LinkedIn and used on Meta as a validation and exclusion layer rather than as its own buy. Exclusions get as much attention as inclusions, including existing customers, your own staff, the CEP ranges you cannot deliver to, and the segment your sales team quietly refuses to serve. The test plan is the deliverable: An audience document with no test attached is decoration. Every stack ships with a budget, a runtime, the conversion count that makes the reading valid, and the decision we will take at each outcome, all agreed before spend starts. Stacks that lose are written up with the reason, and that record is worth more next year than the winners are. What we refuse: We do not scrape personal data, buy lists, or import contact data whose consent basis you cannot show us. Under LGPD that is expensive, and it is a poor way to open a relationship with a regulator. We do not sell interest research to accounts already clearing the learning-phase threshold on broad delivery. We will say so on the call, and you will have spent nothing. We do not build audiences designed to reach people the platform's own policy protects, whatever the commercial argument for doing it. Not included: No creative production, no landing page work, no media management. This is research and a test plan. Execution is yours, or a separate engagement with its own scope. Who this is not for: Ecommerce accounts at scale, where broad delivery plus a clean catalog plus creative volume wins and audience construction is a distraction from all three. And anyone who wants a persona deck with names and stock photography. We do not make those, and nobody in your pipeline is named Carlos, 42, who values reliability.