Ad operations · Measurement
You're deciding from memory
Nobody has time to look back. So we decide from what we remember. The problem is that what we remember is wrong more often than we think, and lately it's wrong for reasons that have nothing to do with us.
Your client asks why ROAS dropped last month.
You answer. Confidently. From memory.
You remember pausing a couple of ad sets. You remember the creative going tired somewhere around the middle of the month. You remember shifting budget on a Thursday. So you assemble the story out of those pieces, and it sounds right, and everyone nods, and you move on because there are four other accounts waiting.
Here's what nobody says out loud: almost none of us go back and check whether that story was true.
Not because we're lazy. Because looking back is the only job in this business with no owner. Launching has an owner. Reporting has a deadline. Optimising has a calendar invite. Going back to see whether last month's call was actually correct? That never makes it onto anyone's Monday.
So we decide from memory. And we've mostly made peace with it.
Memory would be fine if the numbers held still
They don't.
The numbers do not hold still
In March 2026 Meta redefined what counts as a click-through conversion. Real link clicks still count. Likes, shares, saves, comments and video views over five seconds moved to a separate bucket called engage-through, with a one day window. The full effect landed in July.
Read that again, because the implication is unpleasant. Your reported click-through conversions went down without a single thing changing in your account. Same spend. Same creative. Same performance. Different number.
If you compared July against February and concluded your creative was fatiguing, you were wrong. And you would never find out, because nothing in the dashboard tells you the definition moved underneath you.
And that one is not alone
That one isn't alone. Two more from the same year:
- The fee you pay and cannot see. Since 1 July 2026, Meta passes European digital services taxes through as a separate “location fee” on the invoice. It is a real cost, you pay it, and it is excluded from your metrics, your analytics and your exports. The number you judge efficiency by is not the number you spent.
- The retention window with two official answers. Custom audience retention: Meta’s developer documentation says 365 days, the Help Centre still says 180. Both are Meta. Both are live right now.
None of this is a conspiracy. Platforms change, documentation lags, definitions get refined. But the effect on you is the same either way. The ground moved and nobody sent a memo. Your memory of "what normal looks like" is now calibrated to a version of reality that no longer exists.
Find out which of your numbers changed meaning while you were not lookingIt isn't just you
The State of PPC 2026 global report surveyed 1,306 practitioners across 50+ countries.
53% say managing paid search is harder than it was two years ago. The share who say it got easier hasn't moved in two years: 16%.
Asked why it got harder, the top answer wasn't competition or budgets. It was the platforms going opaque, at 62%. Second, at 53%, was measurement getting less accurate.
So the two biggest complaints in the industry are both about not being able to see what happened.
And here's the part that should bother you. When the same people listed their priorities for 2026, measurement came in at 24%.
Second biggest pain in the industry. Ninth priority. Three quarters of us have decided to live with it.
I don't think that's denial. I think it's triage. When you're running six accounts and every one of them needs launching, reporting and explaining, the thing you drop is the thing with no deadline attached. Looking back has no deadline. So it goes.
See what your own accounts remember, freeWhat actually helps
None of this requires a tool. It requires deciding that the looking back is part of the job, and then making it cheap enough that you'll actually do it.
- Keep a dated log of what you changed. One line each time: date, account, what you touched, why. It takes ten seconds. The value shows up three months later when you can line a performance shift against a real timeline instead of against your recollection of a Thursday. Don't assume the platform is keeping this for you either. Google Ads change history covers two years in the interface, but only thirty days through the API that every tool reads. And there's no backfill.
- Write down which number is the source of truth for which decision, before anyone's in the room. Most reporting arguments aren't about data quality, they're about nobody having decided in advance which number counts for which question. Two platforms will never agree. Different windows, different attribution models, even a different date for the same sale. The four dials behind that are laid out in marketing attribution. Chasing an exact match is a trap. Picking one number per decision and freezing the definition is not.
- Re-baseline after a definition change, and never compare across one. When a platform changes what a metric means, your history before that date is a different unit. It isn't worse data, it's a different question. Draw a line and start counting again.
- Check the invoice, not just the dashboard. If a cost exists on the bill and not in the interface, the interface is not where your efficiency lives.
- Judge on conversions, not on weeks. "We ran it for two weeks" tells you nothing. The unit is how many conversions you accumulated and whether the difference clears the noise you'd see anyway with zero changes. The arithmetic for how big that noise is, per conversion count, is in what is a good ROAS. If the difference doesn't clear it, that's not an inconclusive test. That's the test telling you there's no effect.
None of these are clever. They're just the things that stop being optional once you accept that your memory is a worse record than you thought.
The part worth keeping
The reading was always the tax. The judgment is the job.
Nobody got into this to reconcile two dashboards that disagree. But deciding from memory isn't the alternative to that work, it's the bill for skipping it, and it comes due quietly, in decisions that felt right at the time.
You don't need to look back at everything. You need to be able to look back at all.
Last updated:
Questions people actually ask
Is this telling me my judgement is bad?
The opposite. Your judgement is the part that still has value and the part a platform cannot take. What is unreliable is the record it stands on, and a good call made on a half-remembered Thursday is still close to a coin flip.
I keep notes. Is that not enough?
It is more than most people do. The question is whether the notes say what you decided and why, or only what you did. The first still means something in a month. The second reads like a changelog you cannot argue with.
Do I need a tool for this?
No. One written line per account per week beats any tool nobody opens. A tool earns its place when the number of accounts makes the writing itself the bottleneck, and not one account before that.
See what your own record actually says: a free read of your Meta and Google accounts →
Related reading. Google Ads change history: what it shows, and what it quietly hides. Marketing attribution: why every tool gives you a different number. Marketing report template.
Who wrote this
I'm Manu. I've been buying media for eight years. I got tired enough of this to start building an open source console that reads my Meta and Google data and changes nothing in the accounts. The looking back was the part I kept skipping. It's Apache-2.0 and there's a runnable demo on synthetic data if you want to poke at it.