The first 90 days: how we approach a foundation rebuild.

Every new engagement arrives with the same unspoken pressure: show us something by Friday. It is a reasonable instinct from a founder who has just moved a large monthly budget to strangers, and it is also the instinct that produces the worst ninety days available. Tactics applied to an account nobody has diagnosed yet do not compound. They rearrange the noise, generate a chart that points upward for three weeks, and leave the actual constraint untouched underneath.

So we spend the first ninety days on foundations, and we say so before signing anything. What follows is the sequence we actually run, week by week, with the deliverable at each stage and the decision gate that has to be cleared before the next stage starts. It is published in this much detail deliberately: an agency that cannot tell you what it will do in week three is telling you something.

The shape is three phases and three gates. Weeks one to four make the account legible. Weeks five to eight repair it. Weeks nine to twelve start building on it. The gates are the part that matters, because they are what stops an engagement drifting into month four still cleaning up.

Why the order is fixed rather than negotiated

The sequence is not a preference. It follows from a dependency: you cannot evaluate a change you cannot measure, and you cannot measure anything in an account whose conversion definitions are wrong. Every hour spent optimizing before the measurement layer is trustworthy is an hour spent teaching a bidding algorithm something false, and the algorithm does not forget on request. It has to be retrained, which costs another learning period.

There is a commercial version of the same argument. The changes that look impressive in week one are cheap to reverse and small in effect. The changes that move the business are structural, and structural changes made on bad diagnosis are expensive to unwind. Doing the cheap visible work first feels like momentum and is usually just interest accruing.

Weeks 1 and 2: make the account legible

Weeks one and two are for reading, not writing. We take access, score the account against a written standard, and separate the two things every blended dashboard hides: brand from non-brand demand, and reported conversions from business outcomes. Nothing structural changes in this window, which occasionally makes clients nervous, so we show the working as it accumulates rather than saving it for a reveal.

• →Account scored line by line against our published account standard, with the failing criteria named rather than summarized
• →Brand and non-brand separated and quantified, so the real cost of acquiring a new customer is visible for the first time in most accounts
• →Every conversion action inventoried: what it counts, when it counts it, which are primary, which are duplicated, which nobody can explain
• →Reported revenue reconciled against the commerce platform or CRM for the trailing ninety days, and the gap between them written down as a number
• →Change history read end to end, because the last six months of edits usually explain the current shape better than any theory we could construct

Weeks 3 and 4: quantify the waste and the gaps

With the account legible, weeks three and four put numbers against the problems. Not a list of issues, which any audit produces, but each issue with the monthly spend attached to it, because that is what converts a finding into a decision. Waste with a dollar figure gets fixed. Waste described as a best practice violation gets discussed and survives.

• →Indefensible spend identified and totaled by cause: dead geographies, expired audiences, duplicate keyword coverage, campaigns nobody has opened in a year, search terms buying traffic that cannot convert
• →Measurement gaps specified as work items rather than observations, including whether server-side tracking exists and whether it deduplicates correctly
• →Structure mapped against the actual business: margin tiers, product priorities, geographic reality, and where the current architecture cuts across all three
• →A written repair plan in priority order, each item carrying its expected effect, its blast radius, and how we will know whether it worked
• →The plan reviewed with the client before execution, because a structural change nobody agreed to is how trust dies in month two

Gate one, at day 30

The first gate is a factual test rather than a status meeting. Three things must be true before anything structural is touched, and if any of them is false the diagnosis phase extends rather than the repair phase starting early.

1. We can state, without hedging, what every conversion action counts and which ones the bidding is allowed to optimize toward
2. We can reconcile platform-reported revenue against the business source of truth, and explain the remaining gap rather than merely noting it
3. We have a written, prioritized repair plan with monthly spend attached to each item, and the client has read it

Weeks 5 to 8: repair, in dependency order

The repair phase runs measurement first, then waste, then structure, and the order is again a dependency rather than a preference. Fixing structure before measurement means you cannot evaluate the restructure. Fixing waste before measurement means you cannot prove the saving was a saving.

Measurement comes first because everything downstream reads from it: conversion definitions corrected, duplicates retired, server-side events implemented or repaired, deduplication verified against the order count rather than assumed. Waste comes second because it is the fastest genuine improvement available and it funds patience for the rest. Structure comes last within the phase, and only where the current architecture actively prevents good decisions, not because we prefer our own layout.

Gate two, at day 60

1. Conversion tracking is trustworthy: definitions correct, duplicates gone, server events deduplicating against a verified order count
2. The waste identified in week four is either eliminated or has a documented reason for surviving
3. Structure supports the business rather than fighting it, and every change made is recorded with its date so the next drop can be diagnosed against it

If gate two is not clear, the build phase waits. This is the single most common place a foundation rebuild goes wrong, and it rarely goes wrong through incompetence. It goes wrong because month three arrives, everyone wants to talk about scaling, and starting the interesting work on a half-repaired foundation is the path of least social resistance.

Weeks 9 to 12: begin building

Only now does the work most people picture as the job actually start, and it starts working immediately, which is the entire argument for the sequence. Bid strategy becomes meaningful because the targets refer to something real. Creative testing produces signal you can read because the conversion counts underneath it are stable. Budget can move toward what earns, because it is finally possible to tell what earns.

• →Bidding rebuilt against corrected values, with targets set from breakeven economics rather than inherited from whatever was there
• →A creative testing cadence established, sized so that a result arrives before the next decision is needed
• →Budget reallocated toward demonstrated contribution, which is where the scaling constraints begin to matter more than tactics
• →A monthly review ritual in place, with the same numbers in the same order, so drift is visible early rather than discovered at a quarterly
• →The first experiment designed, whether that is a structural test or a genuine incrementality holdout

Gate three, at day 90

Day ninety is not a performance target, and any agency promising one is guessing. It is a capability test: what can the business now say that it could not say on day one. That is a fair standard because it is entirely within our control, whereas the quarter’s revenue is not.

1. The true cost of acquiring a new customer is known, separated from brand demand rather than blended with it
2. The reported numbers are trusted well enough to make budget decisions from, with the residual gap to the source of truth understood
3. Every structural change is documented with its date, so a future drop can be diagnosed against the record instead of debated
4. There is a written plan for the next quarter that names the constraint being attacked, rather than a list of activity

What we deliberately do not do in the first 90 days

• →Launch new channels. A second channel on a broken measurement layer produces two sets of numbers nobody can reconcile instead of one
• →Chase a headline performance number in month one. It is achievable and almost always achieved by harvesting brand demand, which flatters the dashboard and moves nothing
• →Rebuild structure for its own sake. If the existing architecture does not prevent good decisions, leaving it alone is the disciplined choice, however much we would have built it differently
• →Run incrementality tests. A test measured through an unrepaired conversion layer inherits the break and spends real budget to produce a confident wrong answer

Where this goes wrong

The failure mode is almost never a bad diagnosis. It is a gate waved through under social pressure, usually gate two, usually in the week when someone senior asks why the numbers still look flat. The honest answer in that moment is that the numbers look flat because they are finally accurate, and that answer requires having predicted the flatness in advance. If it was not predicted in writing, it sounds like an excuse, and the engagement quietly reverts to tactics.

The second failure mode is structural: this sequence needs someone whose job is to hold the line on it, which is a real difference between engagement models rather than a matter of skill. It is worth being clear-eyed about who should run your ads before committing to a ninety-day rebuild, because a rebuild that stalls at day forty-five leaves an account mid-surgery.

None of this is clever, and that is close to the point. The sequence is unglamorous, publishable in advance, and mostly a matter of refusing to do the interesting work early. What it buys is an account that keeps improving in month nine, rather than one that peaked in month two when the easy wins ran out and nobody could tell why.