Weekly AI account review: one long chat with saved rules, or a fresh brief each time?

Picture a small agency running the same weekly review across a dozen client accounts in Claude. Some people keep one long project with a saved instruction block and reuse it every Monday. Others open a fresh chat each time and paste a short brief. For anything that touches budgets, I'd go with the fresh chat. A long thread picks up old assumptions, like a target CPA from two months ago or a campaign that was already paused, and the model keeps answering from that stale context without flagging it. A self-contained brief with the date range, the current targets and the list of changes since last week is less elegant, but it's much harder to get wrong. Saved instructions still make sense for the stable parts, like report layout and tone. Has anyone caught drift in a long thread only after it gave a bad recommendation?

Honestly I don't think a fresh chat gets you out of stale assumptions. It mostly just moves them into the brief, and a brief typed from memory on Monday is often out of date too. Nobody remembers that the CPA target changed or that a campaign got paused on Thursday.

What I'd try instead is asking the model to open each review by listing every target, date range and paused item it's assuming, so you can correct that list before any numbers get read. That works in a long thread or a new one.

Agreed on the list idea, but it's only as good as where it comes from. If the model builds it from what someone remembers, you just get a tidy list of stale memories. In Shopping and PMax the stuff that drifts fastest is what nobody writes down: a tROAS target moved from 350% to 400% on a Thursday, a product group excluded after its margin dropped, a feed item disapproved and then fixed. All of that is in the account change history with timestamps.

So for a shop with a few hundred SKUs, I'd start the review by pulling seven days of change history plus the current target settings and have the model work from that. Then the assumptions list becomes a quick check instead of a reconstruction.

Also, any target change from the last two weeks may still be inside the learning window, so before and after numbers aren't really comparable yet.

I'd push back on the two-week learning window as a fixed rule. Changing a tROAS target in Google Smart Bidding usually doesn't reset learning the way a new campaign or a big budget jump does, so before and after can be more comparable than that cutoff suggests. Mostly it comes down to conversion volume. An account with 300 conversions a month settles within days, while one with 15 might need several weeks, and then two weeks is too short. Meta is different, since significant edits to budget or targeting do restart learning. So look at the change type and the volume before picking a window. The change history helps here because it shows what kind of change it was, not just that something moved.

Agreed on change type, but I'd narrow the volume point to one number: conversions at the campaign level, not the account. An account can look healthy at 300 conversions a month while the Shopping campaign you changed sees five or six. Smart Bidding learns per campaign, so 300 spread across eight campaigns can leave the one you touched still guessing.

So before picking a window, pull the last 30 days of conversions for that specific campaign. If it's comfortably above the level where the bidder settles in days, a two-week before-and-after works fine. If it's thin, the comparison is mostly noise, and you want a longer window or a holdout instead of a calendar split.

It also changes how you read the change history. A target move on a campaign with steady volume tells you something real. The same move on a campaign with a handful of conversions mostly tells you the bidder hasn't finished learning yet.

Count only the conversions tied to that campaign's own conversion actions, not the account total. If Shopping and a remarketing campaign both claim the same purchase, or a phone-call action is mixed in, the number can look healthy while the bidder is working from something different. Google's rough guidance for tCPA and tROAS has long been around 30 conversions in 30 days, so five or six a month is well short of that. Lag matters too. Purchases from the last few days are still coming in, so a 14-day after window can look flat just because the late conversions haven't been reported yet.

The lag point matters more than people expect, and the reporting view makes it worse. In Google Ads, the same two weeks can look very different depending on whether you segment by conversion date or click date. In conversion-date view, the last few days of the after period are always under-reported, so a target change can look like it hurt when nothing really moved. Check the conversion lag report for that campaign first, then either drop the most recent days from the comparison or wait until they've matured.