Google Keeps Closing the Manual Levers. What's Left of the Agency's Job?

Today, August 14, 2026, a developer notice arrived from the Google Ads API team.

Starting in late September 2026, the campaign-level language targeting setting is being removed from Search campaigns. Matching will follow the language of your ads. The same rule applies to Performance Max on the Search surface. The setting keeps its meaning only for YouTube, Display, Discover and Gmail. The change was announced publicly the same day.

Six weeks until that lever closes.

On its own, a small change. Added to the list of closed levers, less so. I will open that list, and four cases from the past three weeks: three accounts, a July 25 to August 14 window, all live account data. The numbers were re-pulled before publication. Business names are withheld for client confidentiality.

One more thing. In the first case I would assume we were losing measurement, and I would be wrong. Three checks show the cause had nothing to do with measurement.

The levers we lost

I pulled an inventory of the past few years out of my own account archive.

• Expanded text ads are gone. Responsive search ads took their place, and the algorithm writes the copy.
• Ad rotation control is gone.
• "Exact match" stopped being exact. Close variants widened.
• Performance Max arrived. It locked channel breakdown, placement transparency and separate bidding inside one box.
• "Auto-apply recommendations" became a default placement.

Now language.

None of these came back. It is reasonable to assume the trend continues. This concerns me directly, because part of that list is my own daily work.

Let me be honest: the layer that is melting is a layer I use too.

Where did this morning's email hit my own setup engine?

The campaign setup code. Three lines that write a language criterion when opening a new Search campaign. It never touched the reading and diagnosis side. That is no coincidence, and I picked the four cases below to show which layer is not melting.

What Google's algorithm optimizes, and what it does not

By automation I mean Google's smart bidding algorithm. It maximizes the definition of "success" you hand it. It enters every auction for you. It prices to that definition. When the definition is right, it is excellent. I use smart bidding on nearly every account now. The problem is not the bid. It is the definition.

• Who defined the conversion?
• Is that signal measured correctly?
• Does what the ad promises exist in the business?
• Which contact channel goes entirely unmeasured?

None of that is inside the objective function. The gap grows as automation gets stronger. Put stronger optimization on top of bad measurement and you reach the wrong place faster.

Automation optimizes the objective. It does not audit whether the objective is right.

Three of the four cases below are about measurement accuracy, one about whether the ad matches business reality. Causal discipline runs through all four.

1. Measurement accuracy: the false cause of a widening gap

At a diving center I run in Kaş two numbers drifted apart in early August. Google Ads said "this many conversions". GA4 saw more.

My first suspicion was the cookie consent screen we put on the site on August 7. A visitor who never consents leaves no trail, so we lose the conversion. The date seemed to fit.

It was wrong.

If measurement is not lost, what opened the gap?

No conversion had disappeared. Some of them had started landing in a different column. The problem was not measurement, it was where the counting got filed.

The main "Conversions" column in Google Ads does not show every conversion. It counts only the ones you marked as "primary". On August 3 those marks changed, and one campaign's phone and WhatsApp conversions stopped being primary.

The next day they vanished from the main column. They had not gone anywhere. They kept counting in the "all conversions" column. The money in the till had not shrunk, one item was being written in a different ledger.

Three checks showed it.

• Contacts in GA4 had not fallen. They went up, from 17.7 to 24.2 per day. So the tags work.
• The ratio from the "all conversions" column (all_conversions in the API) did not fall the same way.
• The gap opened three days before the consent screen: August 4, the day after the marks changed.

The 17-day total of the Display campaign shows the difference nakedly.

68 conversions show in the main column, 129 in "all conversions". The ones in between were never deleted, they only left the main column.

The campaign breakdown lands in the same place. The drop is in the account total, not in the campaign itself.

The real trap starts here. The campaigns no longer count the same thing.

Search campaigns count calls and WhatsApp, the Display campaign counts qualified contact.

Divide every campaign's spend by that single column and the average cost you get is nobody's real cost. It blends two different definitions into one number.

A dashboard shows the total, not what the total is a total of.

2. Measurement accuracy: the cause of zero was not in the campaign

I looked at a psychologist account in Kayseri. The table was short, and it invited the classic reflex: blame the campaign, the keywords, the bid.

We tested the tag first. Click the WhatsApp link in a browser and a googleadservices.com/pagead/conversion/ request goes out. The enhanced conversion parameter rides along, the field that carries the user's hashed contact detail. The GA4 event fires as well.

The measurement chain worked end to end.

If the tag fires, where does the zero come from?

The cause was the site's consent configuration. It handed everyone a "denied" regardless of region. Both ad measurement and analytics started in a closed state.

A conversion from a visitor who never consented travels without cookies, and Google reports it through modeling. But modeling has a volume threshold. It needs a minimum crowd before it counts anything. Google states the threshold plainly: 700 ad clicks in a 7-day window, per country and domain group. That floor does not scale down to your account size.

This account was at 122 clicks in 17 days. Roughly fifty a week.

Not even a tenth of the threshold. Modeling never engages, and the dashboard reports this account's conversions as zero.

The quantitative proof came from the analytics side. I read the same 28 days from both directions.

These two numbers do not give an acceptance rate. A click is not a user. The smaller number includes organic visitors too. But they give a ceiling: GA4 was seeing at most 27% of the ad traffic. In reality, probably far less.

The business was getting leads. The system could not see them.

No bidding algorithm solves this. It is a software and legal-design problem: a region-scoped consent configuration and the banner itself. The brief went to the developer. On the campaign side we only cut budget until measurement is reconnected.

3. Business reality: the false promise in ad copy

One of my clients changed their service lineup and stopped offering one certificate program. The ad side lagged behind. Five keywords still promising that program, one headline set and PMax search themes were live.

Worse, the sitelink descriptions had been missed in July's first cleanup. A list of courses that no longer existed sat in the ads until mid-August.

We measured what it cost.

Let me not skip a nuance. All of those impressions happened before July 19, the day we first paused the keywords. After that date there are no impressions at all.

So the August work was cleaning up traces of a cluster that had already gone quiet: sitelink copy, residual keywords. Both readings land in the same place. These terms never produced a conversion.

It was never a revenue problem. It was an accuracy problem: running ads that promise a service you no longer provide.

Automation will never flag this. To Google that keyword is a perfectly good keyword, with volume, a decent relevance score and clicks. The fact that the business dropped the service exists in no data source, and the only way to get it is to talk to the owner.

4. Measurement accuracy: the size of a channel nobody measures

Back to the diving center from the first case, this time to the side nobody measures. The table below is analytics data for July 14 to August 12, all sources included.

WhatsApp is two and a half times phone.

So far, fine.

Then I looked at the ad side. In exactly the same date range the "Call" button on the ads had taken 260 clicks. The source is the call extension report in Ads (campaign_asset, field_type = CALL in the API). None of them count as conversions. Call reporting is switched off on the account.

Roughly 59% of the phone path is outside measurement. And the unmeasured part is the largest piece of the phone channel.

A side finding: the two channels run on different daily rhythms. WhatsApp is morning-heavy, peaking at 10 in the morning. A second wave arrives late in the evening, while phone is afternoon-heavy.

That is shift-planning information, not an ad setting.

But the output that carries the most value for a client is often on that side.

An unmeasured channel has a direct consequence on bidding too. Smart bidding never sees those calls. It learns only from people who reach the site and press a button. As call-prone mobile traffic gains weight, that turns into a cost line.

What melts

All four cases produce the same sentence. Automation optimizes the objective. It does not audit whether the objective is right.

The melting layer is this: keyword-level micro-management, bid tinkering, negative list maintenance, campaign setup itself. These are being automated. What you can charge for them is falling. Today's language change hit exactly this layer.

What gains value

The other side of the ledger has four headings.

• Measurement accuracy. Whether the signal really fires, who it fails for, which channel never shows up at all.
• Causal discipline. One variable, dates verified against the change history, transition days dropped, CPA split into CPC and CVR. A dashboard's before-and-after does not do this.
• Auditing automation. Including the engine you built yourself, which is a post of its own.
• Alignment with business reality. Which service is still offered, who answers the phone, when messages get replies.

None of these has a lever in the Google Ads interface. They are not tied to a button Google could close, because none of them sits in the dashboard as a setting.

So what will the language change actually do?

Six weeks left before it closes, and it will do two things.

The first is an opportunity. The fence is gone. Your English ads can now match users whose browser language is German or French. The problem of a tourist searching in their own language and never seeing you gets lighter on its own. The second is a risk, because that same fence kept irrelevant-language traffic out.

From here the defense line is three things: negative keywords, the language of the ad copy, and the landing page. If you run Turkish and English campaigns separately in one account, the odds of both entering the same auction go up. Prioritization now falls to Google's ad group selection.

The code side is short.

• Stop creating language criteria on new Search campaigns. After September the operation returns an error, and since mutate is atomic it can bring down the whole setup.
• You do not have to delete existing criteria, they simply will not be applied.
• Do not delete the ones in PMax, they still target on the other channels.

Two items sit on my own calendar. A dry run of the setup code in the third week of September, and a query reading split by language in early October: the volume of new foreign-language traffic and whether campaigns started eating each other.

What to check in the age of automation

1. Is your conversion column homogeneous? Do you know, campaign by campaign, what each one counts?
2. Does the tag really fire? Have you watched the network request in a browser, or are you assuming it because a number appears in the dashboard?
3. Is your consent configuration region-scoped, or does it tell everyone "denied"?
4. Which contact channel goes entirely unmeasured? The call button on ads, arrivals through Maps, the number typed in directly.
5. When you read the effect of a change, where did the date come from? Memory, or the change history? Was the transition day dropped?
6. When CPA moved, was it CPC or CVR? Did you decompose it?
7. Is "recommendations → auto-apply" switched off on your account? Did you verify it through the API?
8. Does the business still provide the services your ads promise? When did you last ask?

None of this list has anything to do with the levers Google closed. All of it is still open and largely unaudited. I am unhappy about the levers closing too. Automation optimizes the objective. It does not audit whether the objective is right. That audit is sitting there unclaimed.

A client does not pay you for pressing buttons in a dashboard. They pay you because you can prove their data is accurate and their advertising tells the truth.

Frequently asked questions

When is language targeting being removed from Google Ads? Starting in late September 2026, the campaign-level language targeting setting is removed for Search campaigns. Matching will be based on the language of your ads. In Performance Max, the language setting will be ignored on Search but will keep working as targeting on YouTube, Display, Discover and Gmail.

What should change in Google Ads API code? Stop creating language criteria (CampaignCriterion.language) on new Search campaigns. After September that operation returns ContextError.OPERATION_NOT_PERMITTED_FOR_CONTEXT, and since mutate is atomic it can bring down the entire campaign setup. Removing criteria from existing campaigns is optional, they simply will not be applied. Do not remove language criteria from PMax campaigns, they still apply on the other channels.

What happens to agency work as automation grows? The melting layer is keyword micro-management, bid tinkering and campaign setup. The layer that gains value is measurement accuracy, causal discipline, auditing the decisions automation makes on its own, and keeping ads aligned with business reality. Automation optimizes Google's objective function. Whether your conversion is defined correctly is not inside it.

Does smart bidding compensate for bad measurement? No, it does the opposite. If the signal is dirty, stronger optimization takes you to the wrong place faster. In the cases above, zero conversions came from a site consent setting rather than a broken tag, and a CPA gap came from a goal lock rather than competition. Neither is fixable with a bid adjustment.

The numbers are real account data, re-pulled and verified through the Google Ads API and the Google Analytics data API before publication. Related posts: who audits the verdict your automation produced and half of our record day was bots.