Ask the internet about AI and carbon and you get one answer: the model. How much energy a prompt burns, how much water a data centre drinks, how many cars' worth of emissions a training run costs. It is a real and well-covered story.
*But what about the ad that now appears inside the answer?*
In 2026, advertising quietly entered the AI assistant. A "Sponsored" card under a ChatGPT reply. A shopping unit inside a Google AI Overview. A product prompt in Amazon's AI shopping chat. Each one is a media impression with its own delivery footprint - and each one is, today, completely unmeasured. Here is why that gap exists, what can honestly be measured now, and what advertisers should do about it.
Why does everyone measure AI's carbon but not the ads inside it?
Because AI-surface ads fall into a gap between two fields that don't talk to each other. The "AI carbon" conversation belongs to machine-learning researchers and climate reporters, and it measures the *model*. "Ad carbon" is a separate discipline - and it is still built for display and video.
Nobody sits at the intersection of ad-tech sustainability and AI surfaces, so the ad inside the answer has no owner. It is also brand new: sponsored units only reached these tools in 2026, and measurement always lags a new surface. The result is a media format that both rulebooks miss.
Which AI tools now sell advertising?
Several of the largest AI products carry bookable ads as of mid-2026:
- **ChatGPT** shows a "Sponsored" card beneath answers, bookable through OpenAI's self-serve ads manager (testing began February 2026).
- **Google AI Overviews and AI Mode** place shopping and search ads inside the AI-generated summary, labelled "Sponsored," bookable through Google Ads.
- **Microsoft Copilot and Bing** run sponsored options inside AI answers via Microsoft Advertising.
- **Amazon's AI shopping assistant** surfaces Sponsored Products and Brands prompts inside the chat.
For contrast, **Perplexity** tested sponsored answers and then withdrew them, and **Anthropic's Claude** has publicly committed to staying ad-free. The direction of travel is clear: the answer box is becoming an ad surface.
Why don't display and video measurement standards fit AI ads?
Because the unit of exposure is different. Ad measurement is built around a fixed creative in a defined slot - the Media Rating Council's viewability standard, for instance, is **50% of pixels in view for one second**. An AI-surface ad has no fixed geometry: it is a card or a line of text synthesised inside a streamed answer, appearing after the response begins to render.
There is no standard impression definition for it, no viewability rule, and no published emission factor. The display and video methodologies the industry relies on simply do not describe it - which is precisely why it goes unmeasured.
What's the difference between an AI model's carbon and an AI ad's carbon?
They are two different things, and conflating them is the core mistake. The **model's carbon** is the compute cost of generating an answer - inference energy, measured per query or per token. The **ad's carbon** is the media-delivery cost of the sponsored unit - the rendering, the network transfer, the device draw, weighted by the electricity grid where it loads.
The model's carbon is heavily studied. The ad's carbon is not. And for an advertiser, the ad's delivery footprint is the part that sits in their Scope 3 advertising emissions - the part they can actually influence by choosing leaner media.
Can the carbon of an AI-surface ad be measured today?
Partly - and honestly, only the part that can be observed. The measurable portion today is the **ad-delivery path**: detecting and counting the sponsored impression, then attributing the energy of rendering it, transferring it over the network, and drawing it on the device - weighted by the live electricity grid at the moment it loads. That is the same delivery-path modelling already used for display and video, applied to a new surface.
What it produces is a **provisional, delivery-path figure** - clearly labelled as such. It is not a made-up per-impression number for AI ads, and it does not borrow the display and video factors as if they applied. It measures what can be seen.
What can't be measured yet - the inference frontier
The open question is how much of the *answer-generation* energy should be attributed to the ad. When a sponsored result is woven into an AI-generated response, some share of the model's inference cost arguably belongs to the ad - but there is no agreed method for splitting it, and no public factor to cite.
That component stays a named research question, not a fabricated number. Publishing a confident AI-ad emission factor today would mean inventing one. The honest position - measure the delivery path now, flag the inference attribution as unresolved - is the credible one, and it is the one the industry will need.
Why can only a browser see an AI ad's footprint?
Because the ad is rendered client-side, inside the user's session. A sponsored card in a ChatGPT answer or a Google AI Overview is assembled and displayed in the browser - it never passes through the server-side ad-tech pipes or the top-down carbon models that estimate advertising emissions from spend. Those approaches structurally cannot see the rendered card, its label, or the click.
A consented, first-party browser vantage can. It observes the real impression as it happens - which is what makes measuring AI-surface ads possible at all, and why the discipline that measures spend or models the supply chain from the outside will keep missing it.
What should advertisers do about AI-surface ads now?
Treat AI-surface ads as a distinct line, not as display or video, and start counting them - the surface is small today and growing fast, and the reporting habit is easier to build early. Then apply the same discipline that already pays off elsewhere: **leaner delivery paths cut carbon and cost at the same time.** The waste that emits - long supply chains, non-viewable impressions, made-for-advertising inventory - is the same waste that burns budget, so cutting it lowers emissions and lifts working media and ROAS together.
The advertisers who measure this surface first will be the ones who can answer the question every sustainability and finance team is starting to ask: *what is the footprint of our advertising - all of it, including the part now running inside AI?*