AI Ads Explained: Ads Made With AI and Ads Inside AI Answers

Quick quiz

How should you use AI in your marketing?

Question 1 of 425%

How do you use AI today?

Hardly at allA little, for textsA lot, every day

Two meanings of "AI ads", and why the difference matters

Search results for this phrase mix two topics. Some pages describe ads whose creative comes out of a generative model: a banner rendered from a prompt, a short video assembled from generated stills, a headline written by a model. Other pages describe a new place to buy ads: the sponsored unit that appears next to an answer in ChatGPT, or the managed inventory that vendors now resell on AI assistants. Read a source closely before you act on it, because the budget, the risks and the rules are different in each case.

The table below puts the two side by side.

	AI-made ads	Ads inside AI answers
What is AI	The creative: images, video, copy, voice	The placement: a sponsored unit shown in the assistant
Who sees it	Any audience your campaign targets, on the channels you buy	Logged-in adults on the test tiers, in the US so far
Who controls it	Your team, inside the platform's generation tools	The AI platform's ad system and its policies
Main question	Is it true, labelled and on brand?	Does it pay back, and what does it do to trust?

Most of the confusion comes from mixing the two. A team that wants cheaper creative needs a generator and a review step. A team that wants to appear next to assistant answers needs a placement budget and a measurement plan. You can do both, but decide which one you are testing first, and measure it on its own.

How ads made with AI are produced today

The most documented case is Google's. Google Ads lets you generate images, text and video inside its campaign tools. The Asset studio help page describes the flow: you give a prompt, the tool generates a batch of assets (images, text and videos), and the batch is compiled into one project. You then refine the assets in that workspace and save the ones you want before you deploy them into an existing or new campaign. The help page names the Nano Banana Pro image model.

That flow has a useful property: generation and deployment are separate steps. The help page describes deployment as something you choose after you finalize the assets, so the batch does not go live on its own. Use that gap. Put a person between the batch and the campaign.

Two practical points follow. First, a batch is cheap to produce, so the bottleneck moves to review. Ten variants are only useful if someone can say which one makes a claim you can prove. Second, the output is only as good as the brief. Give the tool the product name, the audience, the offer and the claims you can back with a page on your own site. Vague prompts produce generic ads that are hard to defend, and generic ads are the ones that disappear in a feed.

The third point is that generated assets carry a machine-readable mark. We come back to that in the labels section below. Plan for it before the first campaign, not after the first complaint.

The sequence below is the one we would use for a first batch. It is generic enough to apply to any AI generator, not only Google's.

Ads inside AI answers: ChatGPT as the test case

The clearest example of the second meaning is ChatGPT. According to TechCrunch's report of 9 February 2026, OpenAI started testing ads in the United States for logged-in adults on the Free and Go tiers. Go costs $8 per month in the US. Plus, Pro, Business, Enterprise and Education subscribers do not see ads, and users under 18 are excluded.

OpenAI sets a few rules that matter for a budget. Ads are labelled as sponsored and kept separate from the organic content. OpenAI says ads do not influence the answers ChatGPT gives. Ads are matched to the subject of the conversation, to past chats and to earlier ad interactions. They are kept away from sensitive or regulated topics such as health, politics and mental health. Advertisers do not see the conversations themselves. They get aggregate performance data, such as views and clicks.

Two consequences follow. You buy a placement next to an answer, not a position inside it. And the data you receive is thinner than what Google or Meta report, so your own tracking has to show whether a visit turned into revenue. The ad formats and the Ads Manager steps are covered in our guide to ChatGPT Ads and the OpenAI Ads Manager. This article does not repeat them. It looks at what the placement means for the rest of your AI budget.

Vendors have noticed the opening. Some now sell managed access to ChatGPT sponsored inventory and other AI ad surfaces, and they build retargeting audiences from signals such as agent calls or abandoned cart sessions. One example is Asva AI, whose "AI Ads and Retargeting" page claims a 7.62x ROAS across active campaigns and a 72-hour window from cart abandonment to ChatGPT retargeting. We could not verify those figures from an independent source, and the page does not explain how ROAS is attributed. Ask any vendor for its attribution method in writing before you sign.

The timeline below shows the two dates that matter most for planning.

What the EU AI Act asks from 2 August 2026

If you publish AI-made ads that reach people in the EU, two parts of Article 50 of the AI Act matter. The first applies to the provider of the AI system: its outputs, including synthetic audio, image, video or text, must be marked in a machine-readable format and detectable as artificially generated. The second applies to you as the deployer. If you use an AI system to generate or manipulate image, audio or video content that is a deepfake, you must disclose that the content was artificially generated or manipulated. If you publish AI-generated text to inform the public on matters of public interest, you must disclose that as well.

The article has exceptions. Paragraph 4 excludes law-enforcement uses, artistic or fictional works, and text that a human has reviewed with editorial responsibility. Read the exact paragraph before you rely on an exception, because each one has conditions. This section is a summary, not legal advice, and a lawyer should check your own case.

The stakes are real. For infringements of the Article 50 obligations, Article 99 sets a maximum fine of up to 15 million euros or 3% of total worldwide annual turnover for undertakings, whichever is higher.

Labels and watermarks: what they mean in practice

Google's AI content label help page describes three layers. First, Google embeds SynthID, a digital watermark, into images and videos generated with its tools inside Google Ads. The help page says the watermark allows images to be checked to determine whether they were generated or edited using Google AI. Second, Google may apply an AI label on its own, where Google is legally required to do so, or when it receives signals from other platforms. Third, when you use Google's fully automated features, assets may be labelled as AI created or edited on your behalf, and those labels cannot be overwritten.

In practice, a label can appear on an ad you did not expect to label, and you cannot always remove it. Check the labels before launch, not after a complaint. The Google help page on AI content label settings is the reference, and it is worth reading once for each campaign type you run.

The watermark is not a substitute for review. It helps people trace where an image came from, but it says nothing about whether the claim in the headline is true. A labelled ad with a false claim is still a false claim.

A step-by-step method to test an AI ad budget

You can run this method in a month without a data team. It assumes you already have analytics and a conversion event for a sale or a lead.

Step 1: set one hypothesis

Write one sentence you can prove wrong. For example: "AI-generated product images lower the cost per lead on the same offer." Test one variable at a time. If you change the image, the headline and the audience together, you will not know which change did the work.

Step 2: tag every URL

Give each asset its own tracking so that the analytics tool can separate it from the others. UTM parameters do this. Use one pattern for every link, and keep the names in a shared sheet so that nobody invents a new one:

https://www.example.com/pricing?utm_source=chatgpt-ads&utm_medium=cpc&utm_campaign=ai-ads-test&utm_content=asset-07

The utm_source names the placement, utm_campaign names the test, and utm_content names the asset. If your goal tracking is not set up yet, our guide to conversion tracking covers the setup for GA4 and the ad platforms.

Step 3: hold everything else constant

Use the same landing page, the same offer and the same budget for every variant. Change only the asset under test. If the landing page also changes during the test, any difference in results is noise.

Step 4: read the right numbers

Views and clicks tell you whether the ad was seen and chosen. They do not tell you whether it sold anything. Use the table below to decide which number answers which question.

Metric	What it tells you	Where to read it
Spend and clicks per variant	Whether the budget reaches the test audience	The ads platform report
Conversions per variant, with UTM tags	Whether the visit becomes a sale or a lead	Your analytics, filtered on utm_source and utm_content
Cost per conversion	The real price of one result	Spend divided by conversions, for each variant
Brand mentions in AI answers	Whether the campaign changed organic citations	Your mention tracking, before and after the test
Assets flagged or rejected	Where the review step caught problems	The platform's asset or policy report

Step 5: decide with a minimum sample

Do not declare a winner from a handful of clicks. Set the minimum number of conversions per variant before you start, and wait until each variant reaches it. If the test cannot reach that number within your budget, the honest result is "no decision yet", and the cheaper move is to run a larger, simpler test next month. Our note on PPC metrics explains how to calculate the formulas behind these numbers.

Paid placement and organic citation are two different channels

A sponsored unit does not make an assistant recommend you in its own answer. OpenAI says the ads do not influence the answers. That leaves organic visibility to be earned the same way as before: pages that answer the question clearly, sources that other sites cite, and facts an assistant can lift without rewriting them. Our guide to AI visibility optimization covers those levers in detail.

The practical rule is to track the two channels separately. Measure what your ads deliver, and measure your organic mentions on their own, prompt by prompt. If ad clicks rise while mentions stay flat, you are buying visits that do not build the brand the assistant repeats. An AI visibility audit gives you the baseline to compare against before the test starts.

Common mistakes with AI ads

These five mistakes show up in most first campaigns. For each one, the description tells you what it looks like from the inside, how to spot it, and how to fix it.

Mistake 1: publishing a generated asset without reading it

What it is: the generator invents a detail that sounds right: a feature you do not have, a price, a result, a certification. A batch makes this worse, because each variant adds one more chance for an invented detail.

How to spot it: read the asset the way a stranger would, then check every noun and number against your own site. If a claim has no page behind it, it is not ready.

How to fix it: give each asset a named reviewer and a short checklist: one claim, one source, one approval. Reject the asset if any line of the checklist fails, even when the image looks right.

Mistake 2: measuring only the platform's numbers

What it is: the platform reports views and clicks, the ad looks busy, and the team concludes the campaign works. Sales are recorded somewhere else, and nobody connects the two.

How to spot it: clicks go up for a month and revenue or leads stay where they were. In ChatGPT, the data you receive is aggregate, so you cannot find the gap inside the platform.

How to fix it: tag every URL, record conversions in your own analytics, and review cost per conversion every week, not only at the end.

Mistake 3: treating the platform label as the whole disclosure

What it is: the team assumes that the label Google or another platform applies covers the legal duty. That assumption fails when you are the one publishing a deepfake or AI-generated text on a public-interest topic.

How to spot it: nobody can say how an asset was made, and the campaign has no internal record of the prompt, the settings or the reviewer.

How to fix it: keep a generation record for each asset, and ask a lawyer which disclosure line your markets require before you launch in the EU.

Mistake 4: ignoring sensitive categories

What it is: you sell in or near a sensitive area, such as health, mental wellbeing or politics, and expect the same reach as in a neutral category. ChatGPT's ads are kept away from those topics, so reach can be much smaller than the budget suggests.

How to spot it: delivery stays low for days, with no error message and no rejected asset to explain it.

How to fix it: read the platform's policy for your category before you spend. Keep the claims honest and plain, and expect less inventory where the topic is sensitive.

Mistake 5: expecting ads to build organic citations

What it is: the team pays for placements and assumes that the assistant will start naming the brand in its answers. The placement and the answer are separate, so the assumption fails.

How to spot it: ad clicks rise, but the brand does not appear in the answers for the prompts your buyers actually use.

How to fix it: track mentions prompt by prompt, and invest in the content and sources that give the assistant something to cite. Use the mention data to choose which prompts to work on first.

Tools and placements compared

The table below compares three ways to work with AI ads. Pick by the question you are testing, not by the vendor's pitch.

Option	What it covers	Disclosure to check	Measurement limit
Google Ads Asset studio	Batch of images, text and video from a prompt, refined in a project	SynthID watermark; labels may be applied automatically and cannot be overwritten in fully automated features	Results only inside Google campaigns
ChatGPT ads	Sponsored units for Free and Go users in the US test	Labelled as sponsored and separated from the answer	Aggregate views and clicks, no conversation data
Managed AI-ad vendors	Resold access to AI ad inventory and retargeting audiences	Ask who writes, labels and approves the creative	Vendor dashboards; ROAS claims need your own tracking

If you already run Google campaigns, the asset tools are the cheapest place to start, because the review and tracking tools are already in place. If you want to test placements inside assistants, start with one small test on the platform's own tools before you add a vendor layer on top.

Where Sorank fits

Sorank is an SEO and GEO platform, so its part of this story is the organic side. Its AI Mention Tracking shows when ChatGPT, Claude or Gemini cite your brand, prompt by prompt, which is the measurement you need to see whether an ad campaign changes organic visibility. Sorank's public feature list does not include ad buying or ad creative, so for those you will need the platforms covered above. Our AI Mention Tracking page describes how the tracking works.

Questions people ask about AI ads

What are AI ads?

AI ads are either ads whose creative is made with generative AI, or ads shown inside AI assistants such as ChatGPT. The phrase covers both, so check which one a source means before you act on its advice.

Does ChatGPT show ads to everyone?

No. In the US test that started in February 2026, ads reach logged-in adults on the Free and Go tiers. Paid plans are excluded, users under 18 are excluded, and ads are kept away from sensitive topics such as health and politics.

Do AI-made ads have to be labelled in the EU?

Some do. From 2 August 2026, Article 50 of the AI Act requires deployers to disclose deepfakes and AI-generated text published to inform the public on matters of public interest, subject to exceptions. Providers must also mark synthetic outputs in a machine-readable way. Check the exact paragraph with a lawyer for your case.