Keyword Architecture From Search Terms, Not Keyword Planner

Keyword Planner is the wrong place to start and I have not started there in years. The list that matters already sits in your account: twelve months of search terms, exported through the Google Ads Data Transfer Service into BigQuery and broken into one, two and three word n-grams ranked by cost, by conversions and by value. In the accounts I open, somewhere between a quarter and a third of search term spend cannot be attributed to any visible query, because those terms fall under Google's disclosure threshold. So the first honest thing this engagement does is tell you how much of your spend you cannot see, before it tells you what to do with the part you can. Where the list comes from: Your own search terms first. Then competitors' visible coverage. Then the query patterns that exist in your category but not yet in your account: specification strings, part numbers, standards references, the phrasing your customers use with your salespeople rather than the phrasing your marketing team uses with each other. In industrial accounts that usually surfaces a whole tier of part-number and tolerance queries nobody bid on, because each looked small on its own and together they turn out to be the highest-intent third of the account. In staffing accounts it means separating candidate intent from employer intent, the most expensive confusion in that vertical, and one that hides inside terms which read identically on the page. Match types are routing, not targeting: Exact has not been exact since close variants widened, and duplicating a keyword across ad groups no longer steers traffic, because Ad Rank decides which one serves. Duplicate-keyword architecture is therefore not a control mechanism and it comes out. Broad earns its place only where the conversion signal is clean enough for the bidder to steer on; where the signal is a form fill nobody qualifies, broad is a fast way to spend money and I will hold it back until the signal is fixed. Phrase carries most mid-funnel discovery. Exact carries the terms whose economics you already know. The negatives are the deliverable: The negative structure is the thing you are actually buying. Shared lists at account level for the permanent exclusions, which for industrial means jobs, salary and do-it-yourself queries, and for games means free, cheat and download queries. Campaign-level lists then hold the borders between your own campaigns. Every list is checked for conflicts against currently converting keywords before it goes live, because a negative that blocks a term you are winning on is worse than having no list. Structure: Ad groups by theme, sized to carry enough conversions for the bidder to model. I do not build single keyword ad groups; they starve bidding and multiply maintenance in exchange for a Quality Score effect that stopped mattering. Brand and non-brand split into separate campaigns for budget control and honest reporting, not because the split changes how anything bids. Not included: Ad copy and asset writing, which is separate work. Landing pages. Research in languages I do not read, which I will scope and then tell you to hand to a native speaker. Ongoing management, because this is a build. The build is handed over with the queries and the scripts that produced it, so your team or your agency can rerun the whole thing next quarter without me. Not for you if: You want a count of keywords as the deliverable. Your account holds under three months of search term history, in which case there is nothing to n-gram and you need a different starting point. You are looking for a second opinion that confirms the structure you already have. Or your conversion tracking counts every form submission as a conversion, in which case fix that first, because otherwise this list will be optimised toward the wrong queries with more precision than before.