ChatGPT Ads gives French buyers more tools after two months, but context stays hidden

Before bids or measurement even come up, early tests sometimes stall on a more basic question: who at the advertiser should own ChatGPT Ads? "The interest is real, but French advertisers instinctively file ChatGPT Ads under search. In reality the channel cuts across search, media, data and creative," said Alexandre Barré, adtech and partnerships director at media agency Jellyfish.

Who owns the channel internally, and which budget line pays for it? "That question, more than the technology, is often what delays the first tests," Barré said. Opening access isn't enough. ChatGPT still has to fit into org charts and media plans that predate it.

Self-serve buying, available in Europe since Aug. 31, has sharply lowered the barrier to entry. OpenAI's documentation lists a minimum daily budget of €15 for accounts billed in euros. The first technology partners required minimums in the tens of thousands of euros.

The list of partners keeps growing: Criteo, StackAdapt, Adobe, Equativ, Magnite, PubMatic and MiQ, plus Amazon Ads, which was still a U.S. pilot when it was announced. Buyers will pick based on the tool they already use for their other channels, and on which team runs activation.

"Access through a programmatic platform only makes sense if the media teams are running it. If it's search, direct access is often more natural. Both routes coexist in our activations," Barré said.

Centralizing activation, budgets and reporting in one integration can keep ChatGPT from becoming yet another silo. So the right comparison is total cost against the work saved: media, any technology fees, support and management time.

All these routes share one limit: OpenAI still decides which ads run. A partner makes the buyer's job easier. It doesn't hand over the keys to the algorithm.

Buying options have expanded to CPM, CPC and conversion optimization, billed per click or per impression. Buyers can set a maximum bid or let the "Maximize results" strategy adjust bids to get as many results as possible within budget. They can't ask it to hit a target CPA or ROAS.

The $3 to $5 per click that OpenAI has cited is a recommended maximum bid to start with. It isn't a set price or a guaranteed CPC.

Buyers can also now optimize toward a standard event, such as a purchase, a sign-up or a form submission, sent through the pixel or the Conversions API. That gives the algorithm a signal to chase more conversions.

But buyers are still limited to OpenAI's predefined categories. A custom event, such as a lead validated by sales or an order above a set value, can be passed back for measurement. It can't be used as an optimization goal. A brand can tell the platform that a form turned into a qualified lead, but can't ask it to optimize for that event.

That's the main constraint on campaign management: the signal buyers can use isn't always the one they'd want the algorithm to optimize on. Each campaign also runs on a single standard event, and switching means building a new campaign.

"You have to pick the one that best reflects business value," said Barré, who argued the limit needs to go quickly. Picking the event deserves more than a checkbox at launch. A campaign optimized for form volume won't necessarily serve a sales team that wants qualified leads.

OpenAI has announced connections with measurement vendors AppsFlyer, Adjust, Triple Whale, Northbeam and DV Rockerbox, so ChatGPT Ads can slot into existing measurement setups. Data platforms Hightouch, Tealium and LiveRamp handle sending conversion data. That's progress: advertisers can judge the channel in the tools they already use to allocate spend.

Buyers still need to compare apples to apples. Ads Manager offers post-click windows of seven, 14 or 30 days, plus a one-day post-view window that can be switched off. Post-click and post-view conversions can be reported separately. Before putting a ChatGPT CPA next to a search or social CPA, buyers need to align the counting rules.

That leaves the question that matters most: how many of those conversions would have happened without the campaign? OpenAI has incrementality work underway with measurement firms Haus, Measured and WorkMagic, but describes it as preliminary. Early cases don't yet show when ChatGPT drives new customers, at what cost and up to what level of spend.

Agencies don't have to wait for the platform to supply the answer. "At Havas, we have the Causal Lab, used heavily for search but applicable to other digital channels," said Nicolas Brilloux, performance manager at Havas Market, Havas' performance marketing agency. He described comparing pre-test and post-test periods to gauge the effect of turning on ChatGPT.

At Jellyfish, Barré also wants measurement set up before activation. "Our approach: read the first campaigns as test and learn, align expectations before any comparison with search or social, and set a baseline before launch," he said.

Without that groundwork, ChatGPT won't graduate from test budgets. A good attributed CPA is a reason to dig deeper, not to decide which other channel gives up budget.

This is the biggest blind spot. OpenAI pitches ChatGPT Ads on its ability to read intent in a conversation. But buyers can't see the prompts or conversations that trigger their ads. They have no equivalent of a search terms report to tie performance to what users actually asked.

Upfront, they can provide "context hints" about the needs, situations or uses their product addresses, for example running shoes for a beginner training for a first 5K. Those are inputs to the system. They don't show which contexts are available, and they don't guarantee ads will run in those situations.

Campaign structure can still offer a first read. "OpenAI recommends we create one ad group per funnel stage: discover, research, compare, evaluate, decide and purchase," Brilloux said, after a recent exchange with the platform's teams.

Buyers can then separate results for ad groups built for discovery, comparison or purchase, with messaging tailored to each stage. That means they aren't entirely blind to the journey. But strong results from a comparison ad group don't prove every user reached was actually comparing. The intent a campaign targets and the context an ad actually runs in are two different things.

Jellyfish is trying to cut uncertainty before launch. "Since we can't see the conversations that trigger ads downstream, we work upstream," Barré said. With its Share of Model tool, the agency analyzes what AI assistants already say about a brand organically. That covers the intents where it's cited or missing, which competitors show up and which attributes get attached to it.

The agency says the diagnostic helps prioritize intents where the brand lacks visibility, feed the context hints and structure campaigns by attribute, intent or product line, with tailored ads. "The structuring logic comes from paid search, but the signal is the conversation, not the keyword," Barré said. The organic diagnostic helps shape activation hypotheses. It isn't a report on the conversations the ads actually reached.

Negative Phrases and the pilots with ad verification firms Integral Ad Science and DoubleVerify tackle the problem from different angles. Negative Phrases give eligible advertisers extra control. The pilots aim to test how well OpenAI's safeguards hold up.

The pilots run in a controlled test environment, with no access to users' private conversations. IAS and DoubleVerify assess how brand safety rules are applied there. IAS has described a varied, evolving set of scenarios to stress-test those protections. The detailed protocol isn't public.

The distinction matters. Testing the system's safeguards isn't the same as checking the context of every impression bought. The pilots may build confidence in the platform's protections. They aren't presented as a way for agencies to see which conversations their campaigns ran next to, or to tie those contexts to performance.

Sponsored Agents show where OpenAI wants to take its ad business. After clicking an ad, users can keep talking to a sponsored agent, ask questions and dig into an offer without leaving ChatGPT right away. For brands, that could open room for advice and lead qualification between the ad and the site visit. It's far more native than the pretty bland banner ads ChatGPT launched with.

But Sponsored Agents are still being tested with a select group of U.S. advertisers. The same goes for the new visual ad format planned during image generation, quickly nicknamed ChatGPT's "pause ads": an ad that runs while the assistant works. Tests started in the U.S. in October, with no timeline for France.

Targeting has limits too. Custom audiences in principle let advertisers include or exclude customer and prospect lists, as on Google and Meta. But OpenAI's documentation says not to use them in the European Economic Area and Switzerland while ad personalization remains unavailable there. A feature can be in the catalog without being usable by a French buyer.

ChatGPT's total audience also shouldn't be confused with the audience advertisers can reach. Ads appear on the Free and Go tiers. The Plus, Pro, Business, Enterprise and Edu plans remain ad-free. That should temper the audience figures used to sell the channel.

Jellyfish is preparing by doubling down on something e-commerce teams already know well. "We believe the channel will move toward products, with feeds automatically powering campaigns," Barré said. If that happens, product data quality and depth are worth working on now. That's the agency's forecast, not a feature presented as available.

Even as formats expand, one distinction holds. "You can buy visibility around the answer, not the answer itself," Barré said. "ChatGPT Ads puts the brand next to the answer; GEO work gets it inside. They complement each other, and neither replaces the other." Generative engine optimization aims for that organic presence but can't guarantee it. Buying an ad gives no say over the assistant's recommendation.