I used to be the person who'd switch a campaign to Target CPA, watch the CPA spike for four days, panic, and switch it back. I did that on three different accounts before I figured out I wasn't actually testing Smart Bidding — I was just interrupting it before it ever got the chance to work.

So I built a rule for myself: never judge a bidding strategy inside its learning window. And once I stopped breaking my own experiments, I finally had clean enough data to answer the question I actually cared about — not "does Smart Bidding work," but "when does it deserve my trust, and when doesn't it."

Here's the framework I ended up using, and what changed once I applied it properly.

I Started by Checking Whether the Account Even Had Enough Data

The first account I ever ran Target CPA on had maybe 18 conversions a month. I set an aggressive target, turned it on, and got almost no traffic for two weeks. At the time I blamed the algorithm. In hindsight, I was the problem — I'd asked a system that learns from patterns to find patterns in a dataset too small to have any.

The number I go by now: 30 conversions a month before I'll trust Target CPA, 50 before I'll trust Target ROAS.

Below that, I don't fight it — I run Maximize Conversions instead, with no target attached, and just let the account build up a real conversion history first. It's slower, but it's the difference between the algorithm learning and the algorithm guessing.

I also stopped touching accounts mid-learning period. High-volume campaigns usually settle within a week. Thin accounts can take three weeks or more, and every time I made a "quick fix" in the middle of that window, I reset the clock and started over. The single highest-leverage thing I did for my own results was simply leaving campaigns alone longer than felt comfortable.

The Change I'm Watching Right Now

I'll admit this part isn't hindsight — it's happening as I write this. Google is rolling out a change on August 17, 2026 that affects budget-limited campaigns running Target CPA or Target ROAS. Up until now, some of these campaigns have been quietly beating their own targets without anyone noticing, because the system was giving budget-constrained accounts more breathing room than the stated target technically allowed. After the 17th, that breathing room goes away — the algorithm will push harder toward whatever number is actually sitting in the target field.

So before the 17th, I'm doing one thing on every budget-capped campaign: pulling up 90 days of actual CPA and ROAS performance and updating the target to reflect what's real today, not what I typed in eleven months ago. If your account has an old target sitting untouched on a budget-limited campaign, this is the week to check it — not after costs move.

Where I Actually Let the Algorithm Drive

Once the data volume was there and tracking was clean, I stopped second-guessing Smart Bidding in a few specific situations:

• Anything with real auction volume. A campaign running hundreds of auctions a day has more bidding signals in an hour than I could process manually in a week. I stopped trying to compete with that.
• Accounts that had plateaued. I tested Smart Bidding Exploration on one stalled ecommerce account, loosening the ROAS tolerance slightly to let the algorithm chase queries it wouldn't normally touch. It found a meaningful chunk of new converting search terms I hadn't seen in months of manual keyword hunting.
• Anywhere conversion tracking was airtight. This became my litmus test. If I trusted the tracking, I trusted the algorithm's decisions built on top of it. If I didn't trust the tracking, nothing downstream mattered.

Where I Took the Wheel Back

I also learned, the expensive way, that automation isn't the right call everywhere:

• Brand campaigns. My branded terms convert at a rate that made Smart Bidding want to bid far more aggressively than necessary. I moved brand to manual with a bid cap and kept the same conversions at a noticeably lower cost.
• Seasonal products. I once watched a Smart Bidding campaign enter its highest-demand month still calibrated to the previous quarter's flat baseline — it simply hadn't caught up yet. Anything with sharp seasonal swings gets either seasonality adjustments or a manual override going into the spike.
• New campaigns with zero history. I stopped starting cold campaigns on Smart Bidding entirely. Manual CPC first, until there's enough real conversion data to hand off.

I Don't Fully Trust the "It's Always Right" Story Either

I've seen the vendor research floating around claiming manual and hybrid bidding models can outperform Smart Bidding by a wide margin in certain cases, and I don't dismiss it outright — some of what it points to (overpaying on branded terms, chasing low-intent clicks) matches things I've personally caught and fixed. But I also don't treat it as gospel; it comes from a company selling a competing product, tested on its own terms. My read: it's less "Smart Bidding is broken" and more a reminder that automation without oversight will happily optimize toward whatever you feed it, mistakes included. That's exactly why I still check in weekly instead of setting a target and walking away.

The Test I Run Before I Trust Any Strategy Change

This is the part that actually changed my results the most, more than any single strategy pick:

1. Baseline first. Run manual, or whatever's currently live, long enough to know your real average CPA.
2. Test in a controlled window. Split traffic if I can, or accept a defined test period if I can't. Minimum three weeks.
3. Don't touch it mid-test. No target changes, no budget swings over 20%, no panic-switching back.
4. Judge on the full window, not week one. If the new strategy lands within 10–15% of target with equal or better volume, I keep it. If it's off by 40%+ with nothing improving, I revert and go fix tracking or targets before trying again.

That process is boring. It's also the reason I stopped losing weeks to my own impatience.