A click used to mean a handoff. User sees ad, user clicks, user lands on your page, your analytics take over. That model has governed digital advertising economics since the late 1990s. Now OpenAI is testing something that breaks the chain entirely: ads that launch a business-specific AI agent instead of opening a landing page.
The destination is no longer your website. The destination is a conversation.
For B2B marketers who have spent years optimizing landing pages, form fields, and nurture sequences, this is not a minor UX tweak. It is a structural change to how pipeline gets built after the click. And the math implications deserve a closer look before the hype cycle buries them.
What OpenAI Actually Built
The new ad format, spotted in ChatGPT Ads Manager by entrepreneur Juozas Kaziukėnas, works in three stages. First, ChatGPT crawls your website to generate a business profile, pulling common customer questions, support information, and general context. Second, you configure a "business agent" using custom instructions, product feeds, MCP tools for live data, and lead-generation forms. Third, you launch campaigns that point users directly into conversations with that agent rather than to a URL.
The underlying technology appears to be the same foundation that powers Custom GPTs. But the commercial wrapper changes everything. Instead of a prospect reading your positioning and filling out a form, they are asking questions, getting answers, and potentially qualifying themselves before your SDR ever sees a name in Salesforce.
The Attribution Problem Nobody Wants to Model
Here is where the CFO conversation gets uncomfortable.
AdExchanger's analysis raised the right question: when ads are assembled by AI, who owns the customer journey? The brand, the platform, or the algorithm? For B2B, this question has direct revenue implications.
Your current attribution model probably relies on UTM parameters, landing page visits, form submissions, and CRM timestamps. A conversational ad that never touches your domain breaks most of that instrumentation. You cannot pixel a ChatGPT conversation. You cannot A/B test headline variants the way you do on a landing page. You cannot see which questions the prospect asked before they submitted their email.
OpenAI's lead capture forms presumably pass data back to advertisers, but the intermediate signals vanish. Did the prospect ask about pricing? Did they mention a competitor? Did they express urgency? That context, which your sales team would normally extract from a discovery call, now lives inside OpenAI's infrastructure.
Early analysis from practitioners puts it bluntly: nobody can prove it works yet. The measurement layer does not exist. For a channel to earn budget in a CFO-safe marketing org, you need to show CAC payback, not just click volume.
What the Early Testers Are Seeing
Kevin McClary, who has been running ChatGPT ads since mid-May, shared operational details that matter for anyone considering a pilot. The platform is buggy. Settings change without notice. Ads get paused for no reason. Targeting uses something called "Context Hints," a blank box where you describe relevant conversations, topics, or keywords. No one knows the best way to use it yet.
Campaign objectives are limited to Reach and Clicks. Conversion tracking is coming but not available. Budget pacing is unreliable: a campaign with a $10,000 monthly budget can spend the entire amount on day one if demand exists. McClary's advice: treat this as an experiment, not a channel.
That matches what OpenAI's own updates suggest. The pilot launched in the U.S. in February 2026, expanded to Canada, Australia, and New Zealand in March, and reached the UK, Mexico, Brazil, Japan, and South Korea in May. OpenAI reports "no impact on consumer trust metrics" and "low dismissal rates," but those are engagement signals, not revenue signals.
The Real Question: Does Conversation Replace Qualification?
The bull case for business agent ads is that they compress the funnel. A prospect who would normally read three blog posts, download a whitepaper, and attend a webinar before booking a demo can instead ask the agent their questions directly and self-qualify in minutes. If the agent is well-configured, it surfaces the right products, handles objections, and captures intent signals that a static form never could.
The bear case is that you are outsourcing qualification to a system you do not control, cannot fully observe, and cannot iterate on with the same velocity as your own website. Your landing page copy can change in an hour. Your agent's behavior depends on how OpenAI's underlying model interprets your business profile and custom instructions.
There is also a competitive intelligence risk. If your agent is built from a crawl of your public website, your competitors can prompt their own ChatGPT sessions to see how your agent responds. Your positioning, your objection handling, your pricing signals become queryable.
A Pilot Framework for the Skeptical
If you are going to test this, structure it like any other experiment with unclear measurement.
- Set a fixed budget you can write off as learning spend. Do not pull dollars from channels with proven CAC payback.
- Run the pilot for 60 days minimum to account for B2B sales cycle lag.
- Build a manual tracking layer: ask every lead that comes through the agent how they found you, what questions they asked, and what convinced them to submit.
- Compare close rates and deal velocity against your baseline form-fill leads.
Document the confounders. If you are running brand campaigns simultaneously, you cannot isolate the agent's contribution. If your sales team treats agent leads differently because they are novel, you have a process variable.
The goal is not to prove the channel works. The goal is to learn whether the signal quality justifies the attribution opacity.
Where This Goes Next
OpenAI is clearly building toward a world where conversational AI is not just a research tool but a commercial surface. The business agent ad format is a first step. Expect tighter CRM integrations, richer analytics, and eventually some form of conversion tracking that makes the CFO conversation easier.
But "eventually" is not a budget justification. Right now, this is a bet on infrastructure that does not yet exist. The operators who win will be the ones who run disciplined pilots, document what they learn, and avoid over-rotating before the measurement layer catches up.
Model or it didn't happen. And right now, the model has gaps.
