What Changed in Digital Marketing: ChatGPT Runs Ads Now

Key Takeaways

OpenAI announced visual ads inside ChatGPT on 5 October, with image-based formats for product inspiration and experiences, 20+ third-party measurement partners already signed up, and US testing beginning later this month.
Google has deployed SAFE, a four-agent AI system designed to catch spam that violates the spirit of its policies, not just the letter. Its timing aligns closely with the September spam update, which may still be running as you read this.
Google Gemini started appending UTM parameters (utm_source=gemini) to outbound links around 1 October. AI-referred traffic that has been landing as direct in GA4 now has a traceable source.
Google removed language targeting from Search campaigns in late September. Ads now reach any user Google deems relevant based on creative and landing page language, with no advertiser override.
Meta’s AI creative tools have reached 1 million advertisers per month, with 15 million ads generated in a single recent month. Meta reports up to a 22% ROAS lift in early testing, though that figure is their own and has not been independently verified.

Five things happened in the past two weeks. The most significant came on 5 October: OpenAI announced visual ads inside ChatGPT, backed by 20+ measurement partners and a US test beginning later this month. Google, meanwhile, deployed an AI spam catcher called SAFE, removed language targeting from Search campaigns, and Gemini started tagging its own links in your analytics. None of these are isolated product announcements. Together they describe an industry where AI is now the delivery mechanism, the spam referee, and the traffic source all at once.

ChatGPT Is Now Running Ads

On 5 October, OpenAI announced it is testing a visual ad format inside ChatGPT. The format is image-based: product inspiration, product usage in context, or brand experiences, appearing during the image generation flow. It is not a text ad at the top of a results page. It shows up alongside the thing a user asked for.

OpenAI announced more than 20 measurement partners already in place: AppsFlyer, Triple Whale, Adjust, Northbeam, Fospha, Measured, Haus, and others covering mobile attribution, multi-touch modelling, and geo-based incrementality. The breadth of that list is deliberate. Building the measurement infrastructure before scaling the product says something about how seriously OpenAI is approaching this. They are not asking advertisers to trust them on faith. They are wiring up the plumbing first.

Two things OpenAI was explicit about: ads carry clear labels, and they have no influence over ChatGPT’s responses. The separation between the ad layer and the answer layer is structural, not just a policy statement.

Why it matters before you can access it

Testing is US-only, invite-only, beginning later in October. NZ advertisers will not access this directly for months, and no pricing has been announced. That is not the reason to pay attention now.

ChatGPT has roughly 200 million weekly active users, a figure OpenAI reported publicly in mid-2024. The current number has not been officially updated, but the direction is upward. A significant proportion of those users are asking questions with commercial intent. When someone asks ChatGPT to help them choose a contractor, a product, or a service, they are expressing purchase intent in plain language at the exact moment of the decision. That is a different place to show an ad than a keyword match on a search results page.

The measurement partner list is the more immediate signal for marketers. Because Fospha, Triple Whale, and Northbeam are listed, agencies and brands already working with those platforms will be able to connect ChatGPT ad spend to downstream revenue in the same reporting dashboard as their Meta and Google spend. When that data starts flowing, cost per acquisition becomes a concrete comparison. That number will determine how fast money moves toward the platform.

For NZ service businesses, the opportunity is not in this month’s beta. It is in watching what the first six months of performance data shows and being positioned to move when the format opens up more broadly.

Google Built an AI Spam Catcher Called SAFE

Around 25 September, Search Engine Journal reported that Google has deployed a new system called SAFE: Scaled Abuse Forensics Examiner. Google published a research paper describing it, and the timing aligned closely enough with the September spam update that the connection was widely noted.

SAFE is a four-agent system. A Root Agent coordinates the investigation. A Content Understanding Agent, built on a large language model, evaluates whether content violates the spirit of Google’s spam policies, not just the specific rules. A Behaviour Understanding Agent examines patterns: timing, posting frequency, coordination across sites. A Channel Cluster Understanding Agent maps relationships between sites using graph analysis to identify coordinated networks operating at scale.

The key phrase is “spirit of policy.” Standard rule-based spam detection looks for defined signals: keyword stuffing, hidden text, known paid link patterns. SAFE looks for content that is technically compliant with specific rules but still manipulative or low-value. There is no single rule to route around when the evaluator is a language model asked to assess intent.

Google says SAFE significantly reduces the time required to identify novel synthetic threats compared to human-in-the-loop investigation. It is designed to catch new forms of abuse, not just replay detection of historical patterns.

What this changes for site owners

The September spam update, the fourth of 2026, started 24 September and can take up to two weeks. We published a full breakdown of what to check when it began. The guidance stands: if your traffic is steady, do nothing; if it dropped, start with Manual Actions and the Pages report before making changes.

What SAFE adds to that picture is a longer-term implication. Google’s spam detection is shifting toward intent-based evaluation. The practical consequence is that tactics that worked twelve months ago, specifically using AI to generate large volumes of content that technically answers questions without providing genuine value, are being assessed by a system designed to evaluate intent and quality rather than the presence or absence of specific technical signals.

The straightforward path is the one it has always been: publish content that actually helps the person reading it, on a site built for real users, with links earned by being worth linking to. SAFE is Google investing seriously in the infrastructure to enforce what its policies have always said.

Gemini Now Tags Itself in Your Analytics

Around 1 October, Search Engine Journal reported that Google Gemini has started appending UTM parameters to outbound links. The parameter appearing in the wild is utm_source=gemini.

Before this change, traffic from Gemini responses arrived in GA4 with no source attribution in most cases. On mobile, and particularly through Android’s Google Assistant integrations, the referrer header was stripped in transit. That traffic landed as direct, making it impossible to separate from visitors who typed your URL directly, and making any estimate of Gemini’s contribution to site traffic unreliable.

A UTM parameter sidesteps the referrer problem entirely. Instead of relying on a header that may or may not survive the journey from Gemini to your site, the source is encoded in the URL itself. GA4 reads it correctly regardless of the browser or device.

Google has not formally documented the change. John Mueller previously acknowledged the referrer limitation when asked, but no official announcement accompanied the rollout.

The practical impact

If your site receives meaningful traffic from AI-generated responses, and you have been watching your direct channel in GA4 wondering whether it contains more than it appears to, this is the moment to check.

Open GA4, go to Acquisition, and filter sessions by source matching “gemini” for the past two to four weeks. If the volume looks low relative to your overall traffic, it may simply mean Gemini does not link to your site frequently. If it is completely absent, the UTM tags may not yet have appeared in links pointing to your specific URLs. Check again in two to three weeks.

What this makes possible over time is a real picture of which pages Gemini cites, how that traffic behaves after arriving, and whether it converts differently from standard Google Search referrals. We wrote separately about what gets you cited in AI search if you want the structural side of that question.

One note on scope: this appears to apply to Gemini’s conversational interface. Whether the same tagging applies to AI Overviews in standard Google Search is not confirmed.

Google Removed Language Targeting from Search Campaigns

In late September, Search Engine Journal reported that Google has removed language targeting from Search campaigns. The campaign-level setting that allowed advertisers to restrict ads to users of a specific language no longer exists.

Under the replacement system, Google determines which language a search is in, infers which languages the user understands from their behaviour, and decides whether to show the ad based on whether the creative and landing page are relevant to that user. Advertisers do not control this directly. The change applies to Performance Max as well: language targeting is removed from the Search inventory within PMax, though it remains available for non-Search channels.

Who this actually affects

For most New Zealand service businesses running English-language campaigns to a domestic audience, the practical impact is low. Google was already making these determinations algorithmically in most cases, and an English-language campaign targeting New Zealand was not meaningfully at risk of appearing to non-English speakers regardless.

The accounts that need to review this are those that used language targeting with specific intent. International campaigns that separated English from other language audiences. Regulated categories where showing an ad in the wrong language created a legal or compliance exposure. Campaigns where language exclusion was a deliberate segmentation tool rather than a default setting from the original build.

For those cases, the language layer is gone. The architecture that replaces it is separate campaigns with separate creative: one campaign with English creative and an English landing page, a separate campaign with the other language, with segmentation handled by geographic targeting and campaign structure rather than a language filter.

If you manage campaigns in this category and have not reviewed whether the language setting was doing real work, this is the right moment to check.

Meta AI Creative Is Running at Scale

Meta’s AI creative tools have been building throughout 2026, but a set of figures from late September clarifies how far they have spread. Search Engine Land reported that more than 1 million advertisers are using Meta’s AI ad creation tools each month, with 15 million ads generated in a single recent month.

The tools now include several distinct capabilities. UGC-style AI video: short clips of a person discussing a product, built from AI rather than real footage, in a format that resembles organic customer content. Brand automation: logos, colours and fonts applied automatically across ad variations. AI voiceovers and translations. Scene extraction: key moments pulled from existing video and assembled into short formats for specific placements.

Meta reports up to a 22% boost in return on ad spend in early testing of the AI creative tools. That is Meta’s own figure, published in their marketing materials, and it has not been independently verified. It describes what they observed in a controlled test. Treat it as a directional signal.

The practical question for NZ advertisers

Whether these tools are worth using depends entirely on what you are comparing them to.

If your alternative is running the same static image or the same video creative for another three months because producing new material is expensive and slow, then AI-generated variations of existing assets are almost certainly better than doing nothing. Ad fatigue is real and measurable. The same ad shown to the same person seven times in a fortnight stops producing results. Anything that generates enough variation to extend a performing concept has practical value in that context.

If your alternative is well-produced creative built around real customers, real outcomes, and a specific offer with genuine evidence behind it, the AI tools will not match that. They can produce volume and variation. They cannot produce the credibility that comes from a real result or a real person.

The UGC-style video format deserves specific attention. Meta can generate footage that looks like customer testimonial content but is entirely AI-produced. The Advertising Standards Authority in New Zealand applies a straightforward test: advertising must not be misleading. Synthetic footage used in a context that implies real customer endorsement would likely fail that test. The format is legitimate for product demonstration, brand storytelling, or scenarios where the AI origin is clear. Using it as a substitute for genuine customer testimonials carries real regulatory risk.

What Connects All Five

The thread across this fortnight is one that has been building for two years, and is now difficult to ignore.

AI is the layer underneath everything. Google is using it to police spam. Google is using it to decide which language your ad serves. OpenAI is using it to match advertisers to users at the moment of a commercial decision. Meta is using it to generate the creative itself. And Gemini is now using URL parameters to report exactly when it referred someone to your site.

The advertiser and the site owner are still present. But the number of decisions they directly control has been decreasing systematically, update by update. What remains in human hands are the decisions that require genuine judgement: what offer to make, what creative direction to take, what pages are actually worth building, and what conversion data to feed back into the system.

That is the part automation cannot replace at the moment. It is also where the gap between a well-run account and a poorly-run one is most visible, because automated systems do not correct for the wrong strategy. They execute it faster.

What To Do This Week

Check GA4 for utm_source=gemini traffic. Open Acquisition, filter by session source containing “gemini,” and look at the past two to four weeks. What you see now is a baseline. Track it.

Review any Search campaigns where language targeting was doing real work. If you had international campaigns or regulated categories where language exclusion mattered, the setting is gone. Check whether creative language and campaign structure now handle the segmentation, or whether you need to rebuild.

Do not use Meta’s UGC-style AI video as a customer testimonial substitute without explicit disclosure. The Advertising Standards rules in NZ are based on accuracy. Use the format for demonstration and brand storytelling, not as a proxy for real customer endorsement.

Watch the ChatGPT ad announcements for performance data from the US beta. The first CPA figures from the October test will be the evidence that determines whether the format is worth budgeting. That data does not exist yet. When it does, it will be worth reading carefully rather than acting on the announcement alone.

If your traffic dropped during the September spam update, wait until around 8 October before drawing conclusions. That is two weeks from the rollout start. Making changes before the update finishes running makes it impossible to separate cause from coincidence.

We watch these changes each fortnight and apply them directly to the accounts we manage. If you would rather have someone else running that process on yours, get in touch.

Frequently Asked Questions

Can NZ businesses advertise on ChatGPT now?

Not yet. OpenAI’s visual ad format is in early testing with a select group of US advertisers starting later in October 2026. No timeline for international access has been announced. The measurement infrastructure being built now suggests a broader rollout is planned, but no dates are public.

What is Google’s SAFE spam detection system?

SAFE stands for Scaled Abuse Forensics Examiner. It is a four-agent AI system Google has deployed to identify spam that violates the intent of its policies, not just specific rules. One agent reads and evaluates content using a large language model, one analyses timing and behaviour patterns, and one maps coordinated relationships between sites using graph analysis. Google says it significantly reduces investigation time compared to human reviewers.

Will removing language targeting from Search campaigns affect my NZ campaigns?

For most NZ businesses running English ads to a local audience, the practical impact is minimal. Google was already handling this algorithmically. Accounts affected are those that used language targeting deliberately for international campaigns, compliance requirements, or audience segmentation. For those, separate campaigns with separate creative is now the correct structure.

How do I see Gemini referral traffic in GA4?

Go to Reports, then Acquisition, then Traffic Acquisition. Filter by session source containing “gemini.” You can also create a custom segment using utm_source exactly matching gemini. If the row is absent, the UTM tags may not yet be appearing on links pointing to your specific URLs. Check again in two to three weeks.

Are Meta’s AI-generated video ads compliant with NZ advertising standards?

Meta’s disclosure policy on synthetic creative is still evolving. The Advertising Standards Authority in New Zealand requires advertising not to be misleading. AI-generated footage used in a context that implies real customer endorsement would likely breach that standard. Use AI video for product demonstration, brand storytelling, or clearly branded scenarios rather than as a substitute for actual customer testimonials.