Google Ads AI labels: what you have to declare

You asked an LLM for fifteen headlines, kept four, rewrote two of those, and pasted them into a responsive search ad. Somewhere in the campaign build there is now a control asking whether AI created or edited your assets. What do you tick?

Nothing about using AI to write ads is banned, and Google doesn’t penalise copy for being machine-written. What changed in July 2026 is disclosure. Google added an AI content label setting across Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Ads Editor, alongside a “How this ad was made” panel in My Ad Center. Use Google’s own generative features and the labelling happens for you. Generate assets somewhere else and you declare them yourself. In the EU, India and New York, where local rules require it, the label can appear on the ad itself rather than only behind a click.

That’s the whole change. It is smaller than most of what’s been written about it, and in one respect it’s stricter.

Do you have to label AI-written ad copy?

Start with what Google actually published. The AI content label setting lets you “disclose assets as AI generated or edited”. Assets you designate that way “will then include a disclosure in the ‘How this ad was made’ section” in My Ad Center. Google names the regions driving this plainly: the requirement to disclose applies to campaigns targeting the European Union, India and New York.

Now the part that most write-ups skip. Google’s policy note on AI labelling is specifically about permitting “text or visual labels directly within image and video ad creatives that were generated or modified using AI”, and confirms those labels won’t trip the existing text-overlay and watermark policies. The label setting page talks about “assets” more broadly, and gives an example that is visual: animating still product photos into an optimised product ad.

So the honest answer for plain text is: Google’s documentation builds this around image and video assets and does not spell out AI-written headlines and descriptions the way it spells out generated imagery. If you write RSA copy with an LLM and edit it yourself, there is no line in Google’s help centre that says you must tick the box. If you generate the creative, there is.

The safe default is to declare what you generated and not to agonise over a headline you rewrote. The unsafe default is to assume any of this is settled, because the underlying laws in three jurisdictions are what’s moving, not Google’s UI.

What counts as AI-generated or edited?

Google’s framing covers both ends: assets created entirely by AI, and existing assets meaningfully changed by it. The worked example on the help page is a set of still product photos animated into a product ad. Nobody generated a fake product, and it still counts, because AI materially made the thing the viewer sees.

Three cases are already decided for you:

• Google’s own generative features. When you use Google’s fully-automated creative features, “assets may be labeled as AI created or edited on your behalf”. This includes anything Performance Max generates for you.
• Regions where Google is legally required to label. Google says it “may label your assets in some cases, such as in regions where Google itself is legally required to do so”.
• Both of the above are final. In those cases, “labels cannot be overwritten”. You don’t get to untick it.

That last point is the one worth internalising. If you’re running Performance Max with asset generation switched on, some of your ads carry AI labels already, decided by Google, with no control on your side. That happened whether or not anyone on your team read a policy page.

Where does the label actually show up?

Two different surfaces, and conflating them is where most of the confusion comes from.

After the click, everywhere. The “How this ad was made” panel lives in My Ad Center, which a viewer reaches through the three-dot menu on the ad. Google describes it as available globally, and it only appears for ads that use AI-created or edited assets. Alongside it, Google carries machine-readable provenance as SynthID and C2PA metadata.

On the ad itself, only where required. Google’s announcement puts it as: “based on local requirements, a label may also appear directly on the ad.” For ads you’ve designated as AI-created in the regulated regions, the help centre says those ads “will also include visible overlays”.

The practical difference is large. A post-click disclosure in a panel most people never open is close to invisible. A visible overlay on the creative is a design constraint, and it lands on the ad your art direction assumed was full-bleed.

Which regions does this bite in?

Google names three: the European Union, India and New York. Everything else gets the My Ad Center panel without an on-ad overlay.

If you’re a UK or US advertiser running to a global audience, this is not theoretical. Campaign geotargeting decides it, not where your company is registered. A campaign that includes EU countries is a campaign in scope.

The caveat Google puts on its own product is worth quoting, because it’s the part an agency will not volunteer: using the AI label setting “doesn’t guarantee compliance with specific regulations”, and advertisers should take their own measures to meet local legal obligations. The setting is a mechanism, not an indemnity. Where the money is meaningful, this is a question for someone qualified to answer it, not for a blog.

What the first page of search results gets wrong

Search for this and you’ll get confident, specific, wrong answers. The recurring claim is that from 5 March 2026, every ad using any AI-generated element across Search, Display, YouTube, Shopping and Performance Max must carry a visible “AI Generated” label in the ad unit, with disapprovals, policy strikes, and account suspension after three.

That date and that mandate don’t appear anywhere in Google’s own documentation. What Google published is dated 9 July 2026, describes a labelling control that rolled out gradually across July, ties on-ad labels to three named jurisdictions, and frames the global surface as a post-click panel. The strike-count escalation being quoted is Google’s general policy-violation process, borrowed and attached to a requirement that isn’t written the way it’s being described.

This is a fair preview of what SEO looks like now: a query with real demand, no authoritative independent coverage, and a dozen pages generated to fill the gap, each confidently inventing the specifics. The defence is boring and it works. Open the platform’s own help page and read the dates. If a claim about policy doesn’t have a support.google.com URL under it, treat it as someone’s guess. The same instinct applies to keyword tools quoting AI search volumes that turn out to be counting something else entirely.

What this changes about writing ad copy with AI

Less than you’d think, and it does put one new demand on your process.

Generation is unaffected. Ask for thirty headlines, throw away twenty-six, rewrite the rest. Nothing about that has changed, and nothing about it is penalised. What the ad platforms have always penalised is low-relevance, misleading or policy-breaking ads, and an LLM will produce those enthusiastically if you let it, because “write me a compelling headline about our guarantee” is one prompt away from a claim your legal page doesn’t support.

The new demand is provenance. To answer the label question at all, you have to know which assets were generated rather than written, and by what. That’s easy for four headlines you made this morning and hard for two hundred assets across nine campaigns built over five months by three people and an agent.

A workable habit, in order of how much effort it costs:

1. Decide the rule once, in writing, for your account. Which asset types you declare, and which you don’t.
2. Keep generated image and video assets separately named, so provenance survives the upload.
3. Check what Performance Max has already labelled on your behalf before you write a policy that contradicts it.
4. Re-read Google’s two pages when a campaign expands into the EU, India or New York, because that’s the moment the overlay becomes a creative problem.

Track which assets your agent touched

This gets sharper when the copy isn’t written by a person at all. If an agent drafts headlines, uploads creative and builds the campaign, the question “which of these assets did AI generate” has an exact answer, and it’s in the record of what the agent did.

That’s a good reason to run agent-built campaigns through a surface that logs its own writes. Every FlyWheel tool call is recorded with args, status, latency and actor, so “what did the agent upload to this campaign, and when” is a query rather than an archaeology project. It’s the same reason bidding changes deserve a paper trail: when the number of things touching your ad account goes up, memory stops being an acceptable source of truth.

Declaring AI use is going to keep getting more specific, in more places, on shorter notice. The accounts that will find it easy are the ones that can already answer where an asset came from.

---

If an agent writes your ad copy, you should be able to say which assets it made. FlyWheel gives your AI agent one MCP surface across Reddit, Google Ads, Meta, and X, with every tool call logged with args, status, latency and actor, and new campaigns shipped paused by default. Get started with FlyWheel.