OpenAI Is Testing an Ad That Replaces Your Landing Page With a Chatbot

Juozas Kaziukėnas found a new campaign type sitting in the ChatGPT Ads Manager and posted a screenshot of it on LinkedIn. It’s called Agent. Selecting it means the click on your ad doesn’t open your website. It opens a ChatGPT conversation with a business agent you configured, and that agent handles the question, recommends the product, books the appointment, or takes the lead.

Search Engine Land reported the find on July 31. The setup runs in three parts. ChatGPT crawls your site and generates a business profile from it, covering common customer questions and support details. You then layer custom instructions on top, plus a product feed. You can also add MCP connections, the plumbing that lets the agent read live data out of your systems, so it answers “is that in stock” instead of guessing. The campaign points at that conversation instead of a URL.

Anu Adegbola, Paid Media News Editor at Search Engine Land, put the consequence in one line:

“the conversation, not the website, becomes the destination”

We’ve covered both halves of this arriving separately. In July, Time started selling ads inside a machine-readable version of its site, aimed at the models reading it. In May, product feeds became the thing agents shop from. This is the two halves closing on each other. Your inventory data and your ad now land in the same place, and the place is a chat window owned by someone else.

Nothing here is generally available yet. It’s live in the ads manager for a limited set of advertisers, which is where every ChatGPT ad product has started. That’s still a short runway. OpenAI dropped the minimum spend from the $200,000 to $250,000 range down to $50,000 when the self-serve ads manager opened in April, which put the platform inside reach of mid-market budgets in a single move.

Debra Aho Williamson, founder and chief analyst at Sonata Insights, read the pace correctly at the time:

“OpenAI needs to move quickly to establish itself as a legitimate player in the ads business, so launching an ad manager early is an important step.”

The mechanism worth understanding is what a landing page has actually been doing all these years. It’s a fixed answer to a question you guessed in advance. You wrote the headline for the objection you thought was most common, put the pricing where you hoped people would look, and measured which version of that guess converted better. An agent doesn’t guess. It gets the question, then produces the answer, then the next one, which means the sequencing and the emphasis are decided per visitor at runtime instead of by you in Figma.

That’s the trade. You get a page that adapts to every visitor and you give up knowing what it said.

The measurement problem lands first and lands hardest. Your analytics stack was built to count sessions, scroll depth, form fills, and the path between them. None of those exist inside a conversation. The events you’d need are different in kind: what got asked, what the agent recommended, where a conversation stalled, which objection came up in 40% of chats. SEOteric’s guidance to advertisers preparing for this is to rebuild tracking around conversational events and wire them into the CRM. That’s the honest version of the work, and it’s a quarter of engineering time for most 20-person teams, not an afternoon in Google Tag Manager.

The brand problem lands second. Everything a company says on its site went through somebody. An agent built from a crawl, a feed, and a set of instructions will say things nobody wrote and nobody reviewed. The customer on the other end reasonably believes they’re talking to you. The blast radius of a bad answer is a support ticket at best. It’s also the failure mode that gets pinned on the advertiser rather than the platform, which is the pattern we’ve watched play out on every automated ad product Google has shipped.

For anyone selling B2B software, there’s a smaller and more immediate read. The demo request form is one of the last places where a marketing team still controls sequencing: you qualify, you route, you decide who gets a calendar link. An agent that qualifies inside the conversation does that job faster and hands sales a transcript instead of five form fields. That’s a better handoff. It also means your qualification logic now lives in a prompt on OpenAI’s infrastructure, and the funnel we described as already broken loses the one stage that was still yours.

The move to make now is cheap and it isn’t waiting on OpenAI. Take the last 200 conversations your sales and support teams had, and write down the 20 questions that keep coming up along with the answer you’d want given every time. That document is the actual asset here. It’s what a business agent gets configured from, and it’s what your site should already say. It’s also the one piece of this that stays yours no matter which platform ships the format first.