Google Ads Performance Dropped? Diagnose It in Order

Performance drops feel like emergencies and get treated like mysteries — an hour of tab-hopping, a meeting, a theory somebody defends because it was theirs. They are neither. Drops have a diagnostic order: measurement, change history, auction, demand, constraints. The order runs cheapest-to-check and most-often-guilty first, and each layer either names the culprit or hands a cleaner question to the next. An AI with account access walks all five in one session, evidence attached. The order matters more than the tooling — it is what stops the investigation from starting at the loudest theory in the room.

First: is it real, or is it measurement?

The most common Google Ads performance drop is a reporting artefact, which is why measurement goes first — it is also the fastest layer to clear.

Two suspects. Conversion lag: conversions attach to their click date but report days later, so the recent week always looks worse than it will finish. If your typical lag is several days, a "drop" covering exactly the last several days is probably arithmetic, not adversity. Tracking breaks: a site migration, a tag manager change, a consent banner update, a payment flow rework — any of these can cut conversion recording while clicks sail on untouched.

Before anything else: show conversion counts by conversion action for
the last 21 days against the prior period. Is any action recording
zero or sharply less? What does my historical conversion lag look
like, and how much of the recent dip falls inside the lag window?

Clicks flat, conversions down, one action flatlined on a specific date — that is a tracking break, and the date is your lead. Everything degrading proportionally inside the lag window — wait before diagnosing further. One more artefact worth naming while you are here: attribution and counting changes. A switched attribution model or a conversion action redefined mid-month moves numbers without moving reality, which is why the prompt asks per action rather than in total.

Second: what changed? Reading change history like a detective

Accounts rarely degrade spontaneously; they degrade after edits. Change history is the closest thing to a confession, provided you read it wide enough — automated changes included, because auto-applied recommendations, rules and scripts count as changes no human will remember making.

Read change history from [drop date minus 2-3 weeks] to today, all
sources including automated. List changes to budgets, bids, targets,
keywords, ads, audiences and settings. Rank by plausibility as the
cause of a [symptom] starting [date], and say why for each.

The lookback window matters: lag means a change lands in the numbers a week or more after it lands in the account, so the culprit is often older than the symptom. The classic finds: a bid-target tightened, a broad keyword added by auto-apply, a campaign's location options widened, an ad swap that tanked relevance. Two habits sharpen the read. Ask for changes correlated with the symptom's exact start date — a budget edit the same day is a different animal from one three weeks prior. And treat absence as information: a clean history moves suspicion outward to the auction and demand layers, which is the order working as designed. If a plausible change surfaces, you still finish the remaining layers — but now with a hypothesis to confirm instead of a mystery to mourn.

Third: auction pressure — impression share and competitor moves

Nothing changed inside the account still leaves everyone else. Competitors raise bids, enter, leave, run promotions; your ad's effective position and price move without a single edit of yours.

The evidence lives in impression share and auction insights: search impression share falling while lost-to-rank rises means you are being outbid or out-scored; a new name climbing the auction insights table dates the pressure. CPC creeping up at stable positions tells the same story from the cost side.

For the affected campaigns, chart impression share, share lost to
rank, share lost to budget, and average CPC weekly for the last 8
weeks. Compare auction insights for the drop period against the
prior period: who gained overlap or outranking share?

Auction-pressure drops have a distinctive signature — volume and efficiency degrade together while conversion rate holds. Your traffic converts like it always did; there is simply less of it, at a worse price. One caution before responding: pressure discovered in week one often relaxes by week three — promotions end, budgets exhaust, entrants recalibrate. Before re-bidding the whole account, let the timeline say whether you are watching a regime change or somebody's campaign.

Fourth: demand shifts and seasonality

Sometimes fewer people wanted the thing. Search volume moves with seasons, news cycles, paydays and weather, and a demand drop masquerades as a performance drop until you look at the impression base.

The tells: impressions falling with impression share stable — the pie shrank, not your slice — and the same downturn visible in Google Trends for your core terms, or in last year's numbers if you have them. Year-over-year is the strongest single check here and the least used; most operators compare against last week, which cannot distinguish demand from seasonality by construction. This layer is also where honest seasonality reasoning belongs: a drop that recurs every year on the same weeks is a calendar, not a crisis.

Compare impressions and impression share for the drop period against
both the prior period and the same period last year. Does volume
explain the drop while share holds? Is the pattern consistent with
seasonality in this vertical?

Demand drops have unglamorous remedies — ride it out, or broaden coverage deliberately — but misdiagnosing one as an account problem produces the worst outcome: panicked restructuring of an account that was never broken, which turns a demand dip into a self-inflicted learning reset.

Fifth: budget and bid strategy constraint signals

Last, the self-inflicted ceilings. A campaign hitting its budget cap loses its strongest hours. A bid target set too aggressive strangles volume — Smart Bidding obeys the target by buying less. A strategy still in learning after a big change behaves erratically by design.

For the affected campaigns: are any limited by budget, and what is
impression share lost to budget? What are the bid strategy statuses?
Have targets changed recently, and how does actual CPA or ROAS
track against target since?

Constraint drops carry their fingerprints in the status column — "Limited by budget", "Learning" — and in a target-versus-actual gap. The fixes are mechanical: headroom, a loosened target, patience through learning. The one non-mechanical case is a target that was always unrealistic, set from ambition rather than trailing data; the fix is still mechanical — loosen toward what the account has actually achieved — but the conversation that authorises it is not, which is why the evidence table matters. This layer goes last not because it is rare but because its evidence is unambiguous once measurement, changes, auction and demand have been cleared — and because "just raise the budget" is the wrong first move for every other cause on this list.

The single diagnostic prompt that walks all five layers

Symptom-to-layer shortcuts, for orientation:

Symptom	Start at
Conversions down, clicks flat	Measurement
Everything down after a specific date	Change history
Volume and CPC worse, conversion rate stable	Auction
Impressions down, share stable	Demand
"Limited by" statuses, target-actual gap	Constraints

None of the shortcuts replaces the walk — they decide where it starts, and layers upstream of your entry point still deserve a glance on the way past. The full prompt, when you want the whole investigation in one pass:

Account 123-456-7890: [symptom] since [date]. Walk the five layers
in order and report evidence for and against each: (1) measurement —
conversion action health and lag; (2) change history including
automated sources, back 3 weeks; (3) auction — impression share
decomposition and auction insights shifts; (4) demand — impressions
vs share, year-over-year; (5) constraints — budget limits, bid
strategy status, target vs actual. Name the most likely cause, your
confidence, and the single next action that would confirm it.

Minutes later you have what the meeting was supposed to produce: a ranked cause with evidence and a test. The confirming action deserves its own discipline: good confirmations are cheap and reversible — revert the suspicious change and watch, exclude the suspect region for a week, wait out the lag window with a dated re-check — while bad confirmations restructure the account to test a hunch. Asking the model for the smallest test that would distinguish its top two hypotheses keeps the response proportionate, which mid-drop is the hardest thing to keep.

The discipline the order enforces — measurement before blame, changes before theories — is the same discipline a good senior analyst carries; the model just never skips steps when it is told not to. For the account-wide version of this rigour, run a full AI audit once the fire is out.