Nine ways a confident-sounding ad account recommendation turns out to be wrong

Every ad account recommendation arrives dressed the same way: a chart, a number that moved, a sentence explaining why, and a suggested action. The confidence of the delivery has nothing to do with whether the underlying reasoning holds up. Most bad recommendations aren't the result of bad intentions or bad tools. They're the result of a plausible-sounding explanation that nobody stress-tested before it got acted on.

This is the internal discipline we run before touching a live account: nine specific ways a metric, a trend, or an explanation can look solid and be wrong. It's especially useful if you manage an agency and don't touch the ad platforms yourself. You can't verify a recommendation by re-running the analysis, but you can ask the eight or nine questions below, and a good analyst will have ready answers. A vague answer, or a defensive one, tells you something too.

A cost-per-result number from March and a cost-per-result number from June can be computed from two different definitions of "result." Conversion actions get redefined: a platform switches from counting every purchase to counting only primary ones, a pixel gets replaced, someone changes the attribution window from 7-day-click to 1-day-click. The metric name stays identical on the dashboard. The thing being counted underneath it does not.

The check is boring and mandatory: before comparing two periods, confirm the conversion action definition and attribution window were identical across both. If they weren't, the comparison isn't wrong exactly, it's answering a different question than the one being implied. Ask whoever hands you a before/after number: "was the conversion definition and attribution window identical in both periods?" If they haven't checked, the comparison is not yet meaningful.

Ad account performance is noisy at the campaign and ad-group level, more so the smaller the audience and the lower the daily conversion count. A single strong week sits well within the normal range of random variation for a low-volume campaign. It's also exactly the kind of result that gets written up as "this creative is working" or "this audience is converting better now," because a week of data feels like enough to a person looking for a story.

The check is to ask what the baseline range looks like, not just the point estimate. A campaign that normally produces cost-per-result anywhere between $18 and $34 across a normal week didn't necessarily improve because it landed at $19 last week. It's still inside its usual spread. A real pattern needs either more volume than one good week provides, or a large enough deviation that it clearly sits outside the account's normal noise band.

Percentages and shares are ratios, and a ratio can move because the numerator moved, the denominator moved, or both. Rising impression share reads as a win, until you notice the auction itself shrank: fewer total impressions were up for grabs, so a smaller absolute number of impressions won produced a bigger share of a smaller pie. Click-through rate can rise while total clicks fall. Conversion rate can rise while total conversions fall.

The check is simple and non-negotiable: never evaluate a percentage metric without pulling up the absolute numbers underneath it in the same view. If impression share went up, ask what happened to total available impressions and total impressions won. If a rate improved, ask what happened to the volume it's a rate of. A ratio without its numerator and denominator next to it is a number you can't actually interpret.

Bid strategy changes, budget increases past a certain threshold, and audience or campaign structure changes all trigger a relearning or adjustment period on both Google and Meta, during which delivery is deliberately unstable while the system re-explores the auction. Performance during that window is not a preview of steady-state performance, it's the algorithm testing broader and often worse-performing pockets of the possible audience before it converges back down.

Reversing a change, or worse, layering a second change on top, during this adjustment window is one of the most common ways an account gets churned into permanent instability, because each new change resets the clock on the last one. The check: know the platform's typical adjustment window for the change type in question (commonly cited as several days to roughly two weeks depending on the change and the platform), and don't call a verdict on a change until that window has actually closed.

Both Google and Meta surface auto-generated explanations for account changes: "performance dropped due to increased competition," "your reach expanded due to broader targeting." These explanations are generated to be plausible and reassuring, not to be independently verified against your account's actual data. They're a reasonable starting hypothesis, not a finding.

The check is to treat the platform's explanation as a claim to test, not a conclusion to repeat. If the stated cause is "increased competition," that should be visible in independent auction-level data, not just asserted. If the stated cause is audience-related, that should be checkable against the account's own targeting and delivery history. A recommendation that leans entirely on the platform's own explanation, with no independent corroboration, is leaning on a source that has no obligation to be right about your specific account.

Ad accounts rarely get one change at a time. A new creative goes live the same week a budget increases. A bid strategy switch happens alongside an audience expansion. When performance moves after that, there's no clean way to know which change caused it, or whether the two changes interacted in some way neither would alone. A recommendation that confidently attributes the result to one of the two changes is doing that with less certainty than it sounds like.

The check is to ask, for any "this change caused that result" claim: what else changed in the same window? If the honest answer is "two or three things," the attribution is a guess dressed as a finding. The fix going forward is sequencing changes with enough gap between them to isolate effects, or at minimum flagging the confound explicitly rather than picking the more convenient explanation.

Conversions don't all get recorded on the day the click happened. There's a lag between the ad interaction and the conversion event being reported back, sometimes hours, sometimes days, longer for high-consideration purchases with longer decision cycles. A number pulled for "the last 3 days" is a number that will keep rising after you pull it, because conversions still in the lag window haven't been counted yet.

This makes very recent windows systematically look worse than they'll turn out to be, which is a dangerous bias when the recommendation on the table is to pause or cut something based on how it's performing "right now." The check: know the account's typical conversion lag profile, and treat any window shorter than that lag as provisional, not final. A campaign that looks weak in the last 2 days of data may simply not have finished reporting in yet.

This compounds failure mode #1 across a longer timeframe. An attribution window update, a conversion action redefinition, or a tracking implementation change doesn't just make two specific periods incomparable, it creates a structural break in the whole time series. A chart showing 6 months of "declining ROAS" that actually contains an attribution window shortening two months in isn't showing decline, it's showing two different measurement regimes stitched into one line and mistaken for a trend.

The check is to mark every known regime-change date on the account (attribution window changes, pixel or conversion API changes, conversion action redefinitions) and treat any trend line that crosses one of those dates as two separate series, not one. A trendline is only meaningful within a single measurement regime. If someone shows you a multi-month trend, ask whether anything about how conversions are measured changed partway through it.

A single blended ROAS or average CPA across an entire account, or an entire product catalog, can sit at a perfectly healthy number while masking two very different underlying stories: one product line or one segment performing well above target, another performing well below it, averaging out to something that looks fine and hides the part that needs attention. The blended number isn't false, it's just answering a coarser question than the one that matters for the decision being made.

The check is to ask for the same metric broken out by the dimension that's most likely to differ, product line, campaign objective, audience segment, before accepting a blended figure as the basis for a decision. If a recommendation is built entirely on an account-level or campaign-level average with no segment breakdown offered, ask for it. If a segment breakdown reveals a wide spread underneath a stable blend, that spread is usually the more important finding, not the average that hid it.

None of these nine require platform access to check. They require asking the person or team giving you the recommendation to show their work: the conversion definitions on both sides of a comparison, the volume behind a percentage change, the absolute numbers under a ratio, what else changed in the same window, and whether a trend line crosses a known regime change. A recommendation that survives all nine questions is one worth acting on. One that can't answer them, or gets defensive when asked, is telling you something too.