OpenAI announced that ChatGPT Ads would begin rolling out to eligible users in 31 European markets, including Italy, from 24 August 2026. That does not mean every Italian retailer can open an account and launch a campaign today.
As of 25 August, OpenAI’s country availability page still listed Italy as “Coming Soon” for self-service Ads Manager. The initial buying route is through OpenAI Ads Solutions, partner agencies and technology partners. Retailers without one of those routes can register interest and prepare the test, but access is still a real constraint.
The available audience is narrower than ChatGPT usage
OpenAI says ads can appear to eligible adults on the Free and Go plans, while Plus, Pro and Enterprise subscriptions remain ad-free. In the early U.S. pilot, independent researchers observed sponsored ad units below and visually separate from the model response. OpenAI states that advertising does not influence the answer, but the cited audit did not test that policy claim.
OpenAI’s Advertiser API lets an ad group include free-form context hints, described as audience or placement hints that help guide when an ad appears. The hints can be descriptions or keywords, but the public API documentation does not define search-style match types or promise delivery beside a specific prompt.
There is another European constraint: OpenAI’s custom audience documentation says customer-list audiences are unavailable for campaigns targeting the EEA or Switzerland. The first Italian test therefore needs to stand on contextual relevance, product data and geography rather than CRM retargeting.
Start with one part of the catalogue
Do not expose the whole assortment just because the feed supports it. Choose one or two categories where customers need to compare specifications, use cases or trade-offs. Stock should be stable, product pages useful and contribution economics understood.
OpenAI documents Google-compatible product feeds with titles, descriptions, prices, availability, images and destination URLs. Product sets can filter the feed by attributes such as brand, which gives a multibrand retailer a clean way to limit the test. Product-level reporting can then show which items received impressions and clicks.
This is a useful learning unit: one category, a defined product set, a small group of context themes and a landing experience that answers the constraints a shopper is likely to express. Keep weak-margin, unstable or return-heavy products outside the first test.
Build three measurement layers before launch
The campaign needs three versions of the truth, each with a different job.
Platform reporting should carry the standard ecommerce event, such as order_created, with amount, currency and product IDs. Use it for delivery diagnosis and optimisation.
Analytics reporting should use static UTM parameters on every destination URL. Keep campaign, ad group, category, brand and creative identifiers consistent with the rest of the paid-media taxonomy.
Backend reporting should join orders to net revenue, discounts, fulfilment, cancellations, returns, margin and new versus existing customer status. This is the layer that decides whether the test made commercial sense.
OpenAI supports a browser Measurement Pixel and a server-side Conversions API. Its documentation calls the server integration more reliable than the Pixel alone. When Pixel and CAPI send the same order, reuse one identifier value as the CAPI event id and Pixel event_id, and send both under the same Pixel ID so OpenAI can deduplicate them. Configure the Pixel around the retailer’s approved consent implementation before it fires.
The principle matches the wider attribution approach: platform conversions help the platform optimise, while backend outcomes anchor the budget decision.
Add a control that improves the business readout
Attributed orders do not establish how many orders the ad created. When volume allows, use a matched regional holdout.
Choose comparable Italian regions using a clean pre-period. Match them on category revenue, order volume, conversion rate, promotions, stock and other media spend. Run ChatGPT Ads in the treatment regions and exclude the holdout regions, subject to the locations your partner confirms are targetable. Keep the tested catalogue, prices, promotions and other paid activity as stable as operations allow.
Randomize eligible regions where scale and operations allow. If regions are only matched, treat the result as quasi-experimental. Report pre-trend fit, uncertainty, spillover risks and any concurrent media, pricing, stock or promotion differences before using causal language.
Compare the change in backend orders, net revenue and contribution profit between treatment and holdout. Branded search, direct traffic and new-customer share can help explain the result, but they are secondary signals.
If regional volume is too low, a category or brand holdout is possible, although differences in demand and cross-shopping make it weaker. A time-based on-and-off test is weaker again when promotions, paydays or seasonality move during the test. State that limitation in the readout.
Keep view-through reporting in its own column
OpenAI’s measurement documentation separates click-through conversions from view-through conversions. The main Conversions total is click-based. View-through reporting, where available, uses a fixed one-day window and appears separately at campaign level. It is not included in billing or conversion optimisation.
Keep those numbers separate. Adding both together and calling the result incremental revenue would answer a question the measurement never tested.
The current OpenAI documentation and cited audit provide no public European benchmark for ChatGPT Ads reach, CPM, CPC, conversion rate or ROAS. An August 2026 independent audit found that retail and software dominated its early U.S. sample and that product-related prompts often attracted ads. It measured exposure, not advertiser performance, and its prompts were not representative of European use.
That is enough evidence to justify a narrow retail experiment. It is nowhere near enough evidence to skip the control, open the whole catalogue or move a material budget before backend contribution moves with it.