The Modern Google Ads Structure Is Evolving - Are You?

Sometime in September, Google will change the settings on campaigns nobody on your team has touched in months.

Anything still running automatically created assets or the campaign-level broad match setting gets upgraded to AI Max, asked for or not. Dynamic Search Ads were meant to go the same way in September, but Google pushed that sunset out to February 2027 after enough advertisers said the timeline wasn't realistic. The reprieve is real. The direction of travel isn't changing.

So if there's still a debate running in your team about consolidation versus segmentation, how many campaigns is too many, whether to keep the exact match layer alive, it has been settled for you. Not by evidence. By product roadmap.

Which is worth sitting with, because it changes what's actually worth your attention.

‍

The bigger change was a help page nobody read

Around May, Google updated a Google Ads help page to clarify that the search terms you see in reporting for AI-powered Search experiences may not be what the user typed. In some cases they represent Google's interpretation of the intent behind the query. It applies to AI Mode, AI Overviews, Lens and autocomplete.

Read that again if you manage accounts for a living. The search terms report is the instrument most of us have used for fifteen years to answer the two questions that matter: what did we pay for, and was it worth it. It is now, in part, a summary rather than a record.

For a lot of accounts this is a rounding error today. For anyone in a regulated sector, or anyone whose negative keyword work is the difference between profit and not, it isn't. If you're using query data as evidence in a compliance review, or as the audit trail for a client who wants to see exactly where their money went, you need to know that some rows in that table are Google's paraphrase.

The practical consequence is subtle and annoying. Negative keyword mining doesn't stop working, but it degrades. You are now filtering against a description of demand rather than demand itself. The very careful, very manual query hygiene that used to be a genuine differentiator between a good account manager and a mediocre one returns less than it did, and returns less every quarter.

‍

What structure is still for

Structure hasn't become worthless. It has become narrower.

A campaign boundary now buys you four things and only four things: a separate budget, a separate bid target, a separate primary conversion action, and a reporting line you can defend to a finance team. If a proposed split doesn't give you at least one of those, you're building an org chart, not a campaign structure.

That's a smaller job than it used to be, and it's worth being honest about why. The whole case for consolidation was that Smart Bidding needs volume to learn. Everyone knew this. What's happened in 2026 is that Google stopped waiting for advertisers to act on it and started consolidating on their behalf. Arguing about whether to do it is now roughly as useful as arguing about whether to accept enhanced CPC.

The corollary is that the hours you used to spend on structure have to go somewhere. Most teams I talk to haven't consciously decided where.

‍

The one lever the model still has to obey

Conversion definition.

You can steer AI Max with brand controls, location controls and text guidelines. Performance Max now takes campaign-level negatives and shows you channel-level performance, which is a genuine improvement on the black box we all complained about for three years. These are useful. They are also guardrails rather than levers. They shape the edges of what the model does. They don't tell it what good looks like.

The conversion action set does, and it's still the one input Google's systems have no choice but to take literally. Which makes it strange that so many accounts are quietly optimising towards something nobody would defend out loud. Revenue instead of margin, when your product mix has a 40 point gross margin spread across it. Form fills instead of qualified opportunities, when half the forms are students and competitors. A single conversion action carrying five different business outcomes because that's how it was set up in 2021 and nobody wanted to reset the learning.

If you fix one thing this quarter, fix that. Send back what actually happened, not what happened on the website. Offline conversion imports and Data Manager are unglamorous plumbing and they matter more than any campaign setting you could change.

‍

Give the automation its due

It would be easy to write all this as a complaint, and that would be dishonest.

Google's own number on AI Max is 7% more conversions or conversion value at comparable CPA or ROAS when you run the full suite, search term matching plus text customisation plus final URL expansion, against search term matching alone. It's Google's number, from Google's internal data, for non-retail advertisers, so treat it as directional rather than gospel. But the underlying claim holds up in accounts I've seen. Broad matching with a well-fed bidding model does find demand that a keyword list built by a human does not. It finds it in phrasings you'd never have written down, at a volume manual work can't match.

The frustration isn't that the automation doesn't work. It's that it works while making itself progressively harder to audit, and the two things are arriving together.

‍

Measurement is the part worth investing in now

Here's the read I'd offer on Google Marketing Live in May, once you strip away the Gemini branding.

Google announced Ask Advisor, an agent that spans Ads, Analytics, Merchant Center and Marketing Platform. It announced a set of new ad formats built for AI Mode. And it announced that Meridian, its open source marketing mix model, is being built into Analytics 360, alongside a predictive metric called Qualified Future Conversions.

That last one is the tell. A company shipping better marketing mix modelling and predictive conversion metrics is a company acknowledging that click-path attribution can no longer carry the weight being put on it. Google is telling you, fairly clearly, that the in-platform numbers need a second opinion.

So get one. Run a geo holdout. Turn something off in a region for four weeks and see what happens to revenue you can count in your own system. It's slow, it's uncomfortable to sell internally, and it's now the only category of evidence that isn't downstream of a model with an interest in the answer. Search and other advertising revenue was $63.3 billion in Alphabet's second quarter, up 17% year on year, a year into the AI Overviews era. Demand is not the constraint. Knowing what you actually caused is.

‍

If you want a shortlist

Four things worth doing before the end of the month.

Pull a list of every campaign in your accounts still using automatically created assets or campaign-level broad match, and decide deliberately how they should be configured under AI Max rather than inheriting whatever the auto-upgrade gives you. Doing it yourself takes an afternoon. Undoing a surprise takes a quarter.

Open your conversion actions and ask whether the thing you're optimising towards is the thing you'd report to your board. If those two answers differ, everything downstream is optimising for the wrong outcome very efficiently.

Change how you read the query report. Treat it as a signal about intent patterns, not a receipt. Keep mining it, but stop treating a clean search terms report as proof that spend was clean.

And pick one campaign, one region, one four-week window, and run a proper holdout. One real test will tell you more than a quarter of dashboard reading.

‍

The uncomfortable summary

The skills that made someone good at Google Ads for the last decade were mostly skills of observation and control. See the query, adjust the match type, exclude the placement, tighten the segment.

Observation is getting fuzzier and control is being handed back to us in the form of guardrails rather than dials. What's left is the harder, less satisfying work: defining the outcome properly, feeding the machine honest data about your business, and proving cause with tests rather than reports.

That's a less fun job in some ways. It's also considerably harder to replace.

‍