In June, a Miami travel agency had its highest-spend month in eighteen months and the most clicks in the history of its main campaign. The Google Ads dashboard looked spectacular: cost per click was down, volume was up, conversions had multiplied by the thousands.
The following month, the bookings its sales team closed fell 59%.
It wasn’t a month of less activity. It was a month of the wrong activity. And the cause wasn’t in the bids, or the keywords, or the competition. It was that the signal Google used to decide who to buy had turned false.
Google Ads doesn’t know what a sale is. It knows what you tell it
Smart Bidding doesn’t understand your business. It doesn’t know what a booking, a call or a customer is. All it knows is: “of the things you call conversions, get me more.”
If you define what matters well, it is extraordinarily good at finding it. If you define what matters badly, it is extraordinarily good at finding what doesn’t. The same power, pointed the other way.
In this case the campaign ran call-only ads, where the click is the call: whoever taps, dials. It is the cleanest format there is to measure, because nothing sits between the ad and the result.
For months it worked like this:
| Month | Clicks that ended in a call |
|---|---|
| September | 64.5% |
| November | 74.9% |
| January | 70.0% |
| February | 71.4% |
| May | 54.6% |
| June | 18.1% |
| July | 17.7% |
| August | 9.8% |
Two out of every three clicks were a call. Then, one in five.
What broke, exactly
A conversion tag that lived on another domain shared an identifier with this account’s analytics property. The result: an event that fired when someone loaded a page began flowing into the ad account as if it were a conversion for the business.
That event had been firing for months. And for those months it did no harm, because it wasn’t marked as counting toward bidding. It was noise in the reports, not an instruction.
In mid-June, the key event that made it count was created.
From one month to the next, the conversions feeding the algorithm multiplied by more than five thousand. In the call campaign, 84% of the signal it bid on became a page-load event.
And here is the mechanism that turns a measurement error into lost money: the false event fires when someone arrives on the website. Call ads don’t send anyone to the website. Text ads do.
So the algorithm, doing exactly what it was asked, learned that text ads “converted” and call ads didn’t. The day after a change to the ad group settings, delivery flipped entirely: the campaign went from making 16 calls on 31 clicks to making 75 clicks and zero calls, with cost per click down to a third.
Cheaper. More volume. No business.
Why the dashboard fools you precisely when you should be watching it most
This is the cruel part. While this was happening, the indicators most people check were improving:
- Cost per click went down.
- Click volume rose to all-time highs.
- Conversions multiplied by the thousands.
- Impression share lost to rank collapsed.
Anyone looking at that dashboard would have concluded the account was having its best month. And technically it was meeting its goals to perfection. The problem is that the goals no longer described the business.
A contaminated signal doesn’t look like an error. It looks like a success.
How long does it take to find out? In this case, eleven weeks from the flip until someone dated the origin. Along the way, campaigns were paused without knowing why they were underperforming, new ones were launched on the same poisoned signal, and the highest budget in eighteen months was spent. The cost per useful call multiplied by more than three.
The data that rules out the comfortable excuses
When sales fall, there is always a comfortable explanation at hand: the market, seasonality, the competition, the team. Here there was a way to check, and it is the most useful part of the whole case.
The phone system logs every inbound call to the sales line, wherever it comes from. Google Ads only attributes its own. Comparing the two series:
| Month | Total calls | Those that came from Ads |
|---|---|---|
| May | 873 | 679 (78%) |
| June | 421 | 251 |
| July | 323 | 66 |
The calls that didn’t come from ads stayed steady throughout the period, between 170 and 260 a month. They didn’t move. What disappeared was, almost exactly, the volume Google Ads had been delivering.
If the market had cooled, both series would have fallen. Only one did.
That is the kind of check worth being able to run before you need it: a source of truth independent of the platform you are auditing. If your only measure of whether the phone rings is the dashboard of the company selling you the clicks, you don’t have a measure.
Damage never comes alone
There is a pattern that repeats, and it’s worth naming. When the numbers on the dashboard turn strange, the natural reaction is to do more things. And that multiplies the problem, because every new thing is built on the same poisoned signal.
In five weeks, in this account:
- It went from 2 campaigns to 6, all bidding on conversions over the false signal.
- One was launched without a single biddable conversion goal configured.
- A Performance Max campaign went from 158 calls to 20, with six times as many clicks. The same campaign, the following month.
- A new campaign recorded 277 conversions and zero calls. Not one.
- Google’s autopilot (auto-applied recommendations) applied the “add new keywords” recommendation 505 times.
- Search partners were switched on in a campaign literally named “no networks.”
None of those decisions was absurd in isolation. All of them were reasonable given the information the dashboard showed. It’s the same mistake, amplified six times.
How to catch it in 48 hours instead of eleven weeks
You don’t need a forensic investigation. You need four checks, and none takes ten minutes.
1. Review which conversions are “primary,” one by one. It prevents the most damage and is done the least. This account had twelve actions marked as primary, among them “website visits,” “get directions” and “menu views.” A travel agency has no menu. If it’s marked primary, it’s bidding.
2. Separate what is worth money from what merely signals interest. A WhatsApp click, a scroll, a visit to the pricing page: these are micro-conversions. They help you understand the funnel. They don’t work as a bidding goal. The moment a micro-conversion enters your primary goals, you have asked the algorithm to buy curiosity instead of customers.
3. Watch a ratio, not a total. Totals rise and fall for a thousand reasons. Pick your ratio — click to call, lead to meeting, cart to purchase — and put it where you see it every week. Here the ratio gave a month’s warning: the drop from 70% to 54% was already there before the incident.
4. Check which domain your conversions come from. If you manage several sites, several brands or several clients, verify that each tag measures what it thinks it measures. A tag shared between two domains is a time bomb, and it gives no warning: it simply starts adding up.
The three rules that prevent all this
A primary conversion is something that earns money. If it doesn’t, it’s secondary. No exceptions and no “but it’s useful for seeing interest” — that’s what secondary conversions are for: they get reported, they don’t bid.
Every ad format needs the signal it can produce. Mixing call ads and text ads in the same group under a web-conversion bid guarantees the text ads eat the budget. Not because they work better, but because they are the only ones that can fire the event you measure.
A change in measurement is a change in strategy. Creating an event, marking a conversion as primary or installing a tag are not minor technical chores: they are direct instructions about where your money goes. They deserve the same record and the same review as a budget change.
How it ended
The account was fixed in stages, and the order mattered: first clean the signal — take the phantom event out of the key events, isolate the analytics property, remove the micro-conversions from the primary goals — then hand control back to bidding, and only at the end touch budgets.
The indicator watched today is not the conversions column. It is the number of calls longer than sixty seconds, counted from the call log and not from the dashboard. Alerts propose; they don’t execute. And every change to the account is logged before it is made.
In September the cost per useful call came in below May’s, the good month before the incident. It recovers. But the road back took four months, and the hole in the summer does not recover: those bookings were never made.
The honest part about this case
A note that matters, because this trade sells too much certainty.
The sequence is measured: the measurement broke in mid-June, calls from ads fell 63% in June and 90% in July, both measured against May, total inbound calls were cut in half while those that didn’t come from ads didn’t move, and the team’s bookings fell 59% from June to July.
The full causal story is not proven. The evidence points strongly in that direction, but a third of that company’s sales carry no trace of channel, April to June is high season, and there is no July from the previous year to discount seasonality against. Our own internal review recorded a legitimate objection about how far the contaminated tag, on its own, explains the behavior of that particular campaign.
So this is the reading of a business owner with those numbers in front of them, not a theorem. We say so because a diagnosis that doesn’t separate what it measured from what it inferred isn’t a diagnosis: it’s a story.
What you should really take away
Google Ads will do, with uncomfortable efficiency, exactly what you ask. The risk is not that the algorithm fails. The risk is that it works perfectly, chasing the wrong thing, while the dashboard congratulates you for it.
Tuning your goals and metrics is not the boring chore you do after launch. It is the task everything else depends on.