Key takeaways
- Wikis and SOPs go stale because updating them is a separate task that loses to billable work every time.
- The four things worth capturing per client are: account logic, test verdicts, client decisions, and active constraints with reasons.
- Staleness is a bigger problem than capture: a KPI changed at the last QBR will not appear in a static doc until someone remembers to update it.
- When a senior AM leaves, they take the relationship. A context system that captures decisions as they happen means they no longer take the account.
- Automatic capture from meetings, Slack, and ad platforms is the only approach that survives a busy agency without degrading.
Why agency wikis die
Most agency knowledge management fails for a specific, predictable reason: the system that captures information is separate from the system that does the work. Someone runs a test, gets a result, then has to decide whether to open Notion and write it down. Usually they do not. The test result lives in a Slack thread for two weeks, then disappears.
This is not a discipline problem. It is a design problem. When capturing knowledge requires a separate action, it will be skipped under pressure, and agencies are always under pressure.
The result is what most performance marketing teams quietly know: their Notion workspace is a graveyard. SOPs written during onboarding that no longer reflect how anything works. A "testing log" last updated in Q3. A brand guidelines page that predates the rebrand. The documents exist, so leadership feels the knowledge is documented. The people doing the work know better.
The cost shows up in three places. New hires take three to four weeks to become useful on an account because they are asking questions that were answered six months ago. Senior account managers become irreplaceable, not because of skill alone, but because they are the only ones holding the actual account context. And AI tools produce generic, sometimes wrong, output because nobody typed in the background before asking the question.
What most agencies try to capture (and why it goes stale)
The instinct when building a knowledge system is to document process: how to structure a campaign, how to name ad sets, how to run a creative test. SOPs. Checklists. Runbooks.
Process documentation has real value for onboarding and compliance. The problem is that it ages slowly and goes stale quickly. A Meta campaign structure SOP written in early 2023 may be technically accurate but miss the practical judgment the team has developed since. More importantly, it does not capture the thing that actually differentiates your agency: what you know about this specific client.
Generic process can be found anywhere. The client is paying for what you know about their business, their category, and their account history. That knowledge is almost never written down.
What to capture instead
Four categories of information are worth capturing per client. They are more specific than SOPs, they compound over time, and they are the things an AI tool or a new team member actually needs.
Tests and verdicts
Not just what was tested, but what the verdict was and why. "We tested broad match in Q1. CPA was 18 percent higher than phrase match at equivalent spend. Killed it." That single sentence saves the next person from proposing the same test. It also saves you from an AI tool confidently recommending you try broad match.
The verdict matters as much as the result. A test that failed on creative is different from a test that failed on targeting. The distinction tells you whether the underlying hypothesis is worth revisiting with a different execution.
Constraints and the reasons behind them
Every account has constraints that are not obvious from looking at the account structure. Legal killed a certain claim. The client's CFO has a hard floor on target CPA that does not appear in any brief. A specific product line is excluded from prospecting because the margin does not support it.
Without the reasons, constraints look arbitrary. New team members work around them. AI tools ignore them entirely. With the reasons, they become part of the account's operating logic.
Client decisions and the context around them
The client shifted budget from brand to performance in October. Why? Because a competitor launched in their core market and the CEO wanted short-term volume. That decision will shape how you read the account's Q4 numbers for the next two years. It will not appear anywhere in the ad platform data.
Decisions made in client calls, QBRs, and Slack threads are the most perishable and most valuable information an agency holds. They are almost never captured systematically.
What was ruled out and why
This category is underrated. Knowing what the team already considered and rejected is often more useful than knowing what they chose. It prevents the same ground from being covered in every quarterly strategy conversation. It also prevents AI tools from recycling ideas that were dismissed for good reasons.
How to structure this per client
The structure does not need to be complex. For each client, a usable knowledge base has four sections.
Account logic. Why the account is structured the way it is. Campaign architecture decisions, attribution setup, bidding strategy rationale. This is the context a new account manager needs in week one.
Test log. A running record of every structured test: hypothesis, execution, result, verdict. Date-stamped. Linked to the relevant campaign or ad set where possible. Not a spreadsheet nobody opens, but a narrative record that is actually read.
Decision log. Client decisions that affect strategy, budget, or creative direction. Each entry should include the decision, the date, and the business reason. One paragraph per entry is enough.
Active constraints. A short list of things that are off-limits, with the reason for each. This list should be reviewed and updated quarterly.
The structure matters less than the habit of updating it. Which brings the conversation back to the original problem.
Why the habit breaks down
Even teams that build a good structure find it degrades within a quarter. The reason is always the same: updating the knowledge base is a separate task that competes with billable work.
The only durable solution is to make capture automatic, or as close to automatic as possible. This means the system reads the places where knowledge is actually created: meeting recordings, Slack threads, ad platform results, client emails. It extracts the relevant information and adds it to the client's context without requiring anyone to open a doc.
This is technically harder than building a Notion template, but it is the only approach that survives contact with a busy agency. A knowledge system that requires manual updates will be updated when things are slow and ignored when they are not. Things are rarely slow.
The staleness problem is worse than the capture problem
Even if you solve capture, you have a second problem: information goes stale. The brand guidelines from March are no longer current. The KPI the team has been optimizing against was quietly changed at the last QBR. The competitor analysis in the client folder is nine months old.
A static document cannot tell you when it is no longer accurate. A living context system can, because it is reading the same sources where updates happen. When the client mentions a new KPI in a Monday call, the system updates. When a test result comes in that contradicts a previous assumption, the record reflects it.
The gap between what the team knows and what the system knows is the gap that produces bad AI recommendations, wrong onboarding, and the specific frustration of watching a tool confidently propose something you already tried.
Where Almond fits
Almond is built specifically for this problem. It connects to the places where paid media knowledge actually lives: Google Ads and Meta accounts, Slack, Notion, email, meeting recordings, and CRM, with a read-only connection that takes about four minutes to set up. It captures tests, results, and decisions as they happen, keeps the context current without manual re-entry, and ranks information by relevance, recency, and business impact so the CEO's strategy call outweighs a throwaway Slack thread.
For agencies, every client gets a scoped context store. One client's account history cannot surface in another client's recommendations, enforced at the data layer. When a senior account manager leaves, the account context stays. The reasoning behind every decision is versioned and auditable.
The products built on the engine, Canvas for AI-powered dashboards and Workflow Builder for Google and Meta automations, draw on that context layer rather than treating each query as a blank slate. An automation that would scale a campaign the team decided to sunset on Monday will not run, because the system knew about Monday's decision on Monday.
Almond 2.0 early access is open at almond.ai. Every member gets a detailed one-on-one Meta and Google ad account audit with the CEO, recorded. Waitlist requires a work email.
Frequently asked questions
- Why do agency Notion wikis stop being used after a few months?
- Because updating them is a separate task that competes with billable work. When capture requires a manual step, it gets skipped under pressure. The information stays accurate long enough to feel useful, then quietly drifts from reality. Most teams know their docs are stale but lack a practical way to fix it without adding overhead.
- What should a paid media agency actually document per client?
- Four things matter most: the reasoning behind account structure, a running test log with verdicts not just results, a decision log capturing client direction changes and the business reasons behind them, and a short list of active constraints with explanations. Generic SOPs age quickly. Client-specific history compounds in value over time.
- How do you prevent knowledge from leaving when a senior account manager resigns?
- The only reliable method is to capture context in the system rather than in the person. If account history, test verdicts, and client decisions are recorded as they happen, a resignation removes the relationship but not the reasoning. Systems that require manual documentation will always have gaps because senior people are the least likely to stop and write things down.
- Can AI tools help with agency knowledge management for paid media?
- They can, but only if the context they draw on is current and client-specific. A generic AI tool given a stale brief will produce generic or wrong recommendations. The value comes from connecting the AI to the places where knowledge is actually created, so it reads meeting notes, Slack threads, and ad platform results rather than waiting for someone to type in background.
- What is a business context engine and how is it different from a wiki?
- A wiki is a static document store that requires manual updates. A business context engine reads the sources where decisions and results are created, captures information automatically, keeps it current, and ranks it by relevance and recency before an AI ever sees it. The practical difference is that a context engine stays accurate under normal agency workload where a wiki does not.
Almond 2.0 early access
Get a 1:1 Meta and Google audit with the CEO.
Every waitlist member gets a 1:1 audit of their ad accounts. Work email only.