Google's AI Now Writes the Ads. The Craft Moved Somewhere Duller

Google's AI Mode has crossed a billion monthly users and now carries four Gemini-powered ad formats that generate product copy in real time inside AI responses. The ad itself is no longer the deliverable. What the model builds from, your product data and structured information, now determines what gets said about you.

Gemini-written ad formats are not headlines you wrote being served against keywords you picked. The copy is assembled on the spot, per query, in response to whatever the person actually asked. To access the formats, advertisers need to be running Performance Max, AI Max for Search, or AI Max for Shopping. Testing started in the US.

The artifact stopped being the work

For as long as paid media has existed, the craft was the ad: the headline, the angle, the offer, the test, the argument about which variant to run. That artifact is now generated per person, by a model, in the moment.

Nobody is going to keep paying for forty headline variants when the system writes its own, tailored to a question you never saw.

Where the leverage actually went

It moved from the sentence to the substrate. From the creative that gets approved in a meeting to the attribute completeness that has never once been presented in one.

Google's own guidance is unusually direct about this: advertisers whose assets and product data are clean, detailed, and well structured give Gemini more to work with. That is the whole mechanism. The model can only assemble a compelling, accurate ad from information that is present and unambiguous.

Which produces an awkward situation for a lot of teams. The work that just gained enormous value is the work agencies have historically pushed down, out, or offshore.

Advertising tasks losing value compared with tasks gaining value

Lost value	Gained value
Writing headline variants	Feed hygiene and completeness
Manual A/B copy testing	Attribute coverage and accuracy
Keyword-level ad group sculpting	Structured product data
Presenting creative for approval	Auditing what the model actually generates

Is this the end of creative in paid media?

Not of creative, but of a particular billable version of it. The strategic layer, deciding what the offer is, who it is for, and what claim is worth making, matters more than ever because it determines what goes into the feed.

What disappears is the production layer: turning one brief into fifty competent variations. A model does that in seconds, and does it per user rather than per campaign.

The new quality control problem

There is a genuine risk here that is not being discussed enough. If the model writes the copy from your data, then errors, omissions, and stale attributes in that data become public statements about your business, generated at scale, that nobody approved.

The old workflow had a human sign-off before anything ran. The new one has a human sign-off on the inputs and then trust. That is a different discipline, and most teams have not built it.

1. Audit what the model currently generates for your top products or services.
2. Trace anything generic or wrong back to the missing attribute that caused it.
3. Fix the input, not the output. There is no output to fix.
4. Re-check after each feed change, because the generation is not static.

The businesses that come through this are not the ones defending the craft of the headline. That fight is lost and it was never where the margin was. They are the ones who can demonstrate that their data produces accurate, compelling generated copy rather than generic filler, and who can show the difference.

What "clean data" actually means here

The phrase does a lot of unexamined work in this conversation, so it is worth breaking down into things you can check.

• Completeness. Every attribute the format can use is populated, not just the required ones.
• Accuracy. Nothing stale. Discontinued items, old pricing, and superseded descriptions are actively harmful now, because they get generated into copy.
• Specificity. Attributes that distinguish rather than describe. "Blue" is an attribute. "Fits standard UK door frames" is a differentiator.
• Consistency. The same product described the same way across feed, site, and structured data, so the model is not reconciling three versions.

None of this is new advice. What is new is the consequence. Bad data used to mean a slightly worse match. It now means a model publicly inventing a description of your product from insufficient information.

The approval problem nobody has solved

Every advertising workflow ever built has a human sign-off before anything runs. Generated copy breaks that, because there is no artifact to approve in advance. There are only inputs, and then output at scale, per query, that nobody sees before the customer does.

Real-time generated copy has real implications for regulated sectors and for any business where a careless claim carries consequences. Finance, health, legal, and anything with compliance language attached now needs the constraint expressed in the data and the campaign settings, because it cannot be applied at the copy stage.

Where does that leave agencies?

In a better position than the panic suggests, but only if the offer changes.

The billable production layer, turning one brief into fifty variants, is going away. That work was always the most commoditized part of the service and the first thing clients questioned.

What replaces it is less visible and harder to fake: deciding what claim is worth making, ensuring the inputs support it, auditing what the machine actually produces, and being accountable for the gap. That is closer to consulting than to production, and it prices differently.

The agencies that struggle will be the ones whose retainer was implicitly a headcount of people producing variants. The ones that do well can point at generated output and explain precisely why theirs is accurate and a competitor's is generic.