Your media dashboard still treats the landing-page visit as the win. OpenAI just moved the click somewhere else.
On September 14, 2026, Digiday’s Krystal Scanlon reported that OpenAI has put a new ChatGPT ad format in front of select clients: the click does not open the brand’s website. It opens a branded conversation with the brand’s AI agent inside ChatGPT. Digiday viewed and verified video of the experience. The call-to-action is a line like “Chat with us.” The destination is a chat, not a URL.
If your scorecard still ends at CTR to site, you are grading an exit that the pilot is designed to skip.
What actually changed
Until this pilot, ChatGPT ads mostly looked like a familiar object sitting in a new surface. Digiday notes the current placements live under responses for Free and Go users. OpenAI’s Help Center (Ads in ChatGPT) says the same: ads can appear below the end of a response, labeled sponsored, on Free and Go; Plus, Pro, Business, Enterprise, and Edu stay ad-free. That is still a sponsored card under an answer. Useful for attention. Not endemic to conversation.
The Sponsored Agent pilot changes what the click is for. Digiday says advertisers can attach a branded business agent to an ad. When someone hits the CTA, they enter a branded chat inside ChatGPT and talk to that agent. WeRSM (September 14) and XBorder Insights (September 14) both corroborate Digiday’s core facts: select-client access, “Chat with us” style CTA, branded conversation in-app, OpenAI silent on the record.
Wayfair is in it. Digiday quotes Morgan Brown, director of paid media: Wayfair is participating in OpenAI’s Sponsored Agent pilot “to explore how conversational commerce can help customers make more informed, confident decisions,” at limited scale, with guardrails on product accuracy, transparency, and service handoffs. That quote is the operator tell. This is not a brand stunt. It is a paid-media test with accuracy and handoff rules written into the brief.
The corporate language trail was already pointing here. At Goldman Sachs’ Communacopia and Technology Conference on September 8, OpenAI CFO Sarah Friar called ads under responses the “basic starting point” and said the company was seeing “early indications” of a format that feels “truly endemic to AI,” Digiday reports. OpenAI’s January 16, 2026 advertising principles post floated the same direction months earlier: conversational interfaces where someone might see an ad and ask the questions they need to decide. XBorder Insights notes OpenAI’s August 31 update again talked about more native ways for businesses to interact with shoppers inside ChatGPT. Digiday says OpenAI did not comment on the Sponsored Agent story itself. Treat the pilot as verified by Digiday’s video review and Wayfair’s confirmation, not as a generally available product with a public rate card.
The mechanism: the ad click becomes a session, not a referral
Strip the AI-native branding and the pipe still has stages. Someone asks ChatGPT something. An eligible Free or Go session may get a sponsored unit under the answer. Under the basic format, a click still behaves like digital advertising you already know: leave the chat, land somewhere owned. Under Sponsored Agent, the click starts a second conversation inside the same product, with the brand’s agent as the other party.
That is why Farhad Divecha, group CEO of Accuracast, told Digiday this looks like Meta ads that send people to WhatsApp: open a line of communication, qualify who needs a human or a purchase path, then route. The landing page does not vanish. WeRSM is careful on that point. The branded agent becomes the layer between interest and action. The site may still get the sale. It may not get the first click.
OpenAI’s own Help Center draws a privacy line that media and legal both need in the brief. Advertisers do not get private ChatGPT chats, histories, memories, or personal details. They get aggregated performance such as views and clicks. Separately: if a user chooses to message an advertiser through an ad, the advertiser can see the messages sent directly to them. XBorder Insights stresses the same FAQ distinction, and adds the “Ask ChatGPT” path (share an ad with ChatGPT and ask about it) where OpenAI says advertisers do not influence those responses. Two chats can live in one app. One is private assistance. One is a vendor conversation. Ops that treat them as the same event will mis-report intent and mis-state what data left the room.
Measurement is the other half of the mechanism, and Digiday does not publish pricing or detailed reporting for the pilot. XBorder Insights notes OpenAI’s conversion docs still describe Pixel, Conversions API, and click references on landing-page URLs for website campaigns. Those docs do not establish how Sponsored Agent scores the chat that happens before a possible site visit. So the funnel now has at least three events that used to collapse into one: conversation start, site reach, purchase or lead. A chat can answer a useful question and never produce a visit. Counting chat openings as “conversions” without the middle and end is how you invent ROI.
eMarketer’s Asa Hiken (September 14) frames the commercial itch: advertisers have been frustrated by limited visibility into what users were asking when ads showed. Sponsored agents could surface intent through the conversation itself, but OpenAI has not spelled out which insights advertisers will get from the pilot. Hiken’s advice is experiment-mode, not pipeline-mode. That matches Digiday’s broader read of ChatGPT ads seven months in: promising, still early, mostly incremental budgets until returns firm up.
What operators get wrong
Mistake one: keeping landing-page CTR as the north-star KPI for a click-to-chat unit. If the CTA is “Chat with us,” a successful click is a conversation open, not a session on your domain. Score chat starts, qualified intents, handoffs, and downstream purchases. Stop celebrating a site bounce rate that the format was built to avoid.
Mistake two: shipping a brand agent with creative QA and no product QA. Digiday and WeRSM both land on the same risk. A static ad makes a claim and sends someone elsewhere to check it. A branded agent keeps answering. Accuracy becomes part of the ad experience. Wayfair’s guardrails on product accuracy, transparency, and service handoffs are not optional theater. They are the minimum operating system for conversational commerce under a paid click.
Mistake three: treating owned-site data capture as the only reason to buy the channel. Shamsul Chowdhury, SVP paid media at Zeno Group, told Digiday retailers want user data on owned sites, and also know the journey can happen anywhere. Keeping the shopper inside ChatGPT can cut friction and give the platform more control of the path. That is a trade, not a free upgrade. Decide which data you need on first contact versus which outcome you need on first contact.
Mistake four: stuffing Sponsored Agent into the same reporting taxonomy as under-response cards. Digiday is explicit that this format moves past the basic unit under answers. If your trafficking sheet still says “ChatGPT display” and your optimizer still maximizes site CTR, you will buy the wrong auction behavior and read the wrong post-click story.
Mistake five: assuming general availability and holiday-scale inventory from a select-client pilot. Digiday cannot say how many advertisers have access. OpenAI did not comment. Chowdhury notes OpenAI could become an alternative channel into Q4 if the agent experience is seamless, and calls that the big “if.” Plan a limited test with handoff rules. Do not reforecast Meta and Google savings on a format that still lacks public pricing and chat-to-purchase plumbing.
Mistake six: ignoring disclosure and synthetic-creative pressure in the same week. Search Engine Land (Anu Adegbola, September 14) reports Microsoft Advertising published guidance requiring advertisers to disclose AI-generated or materially altered content where law requires it, preserve provenance signals, and keep disclosures near the creative. That is not the Sponsored Agent thesis. It is same-week proof that platforms are tightening the line between AI-made ads and misleading ads while OpenAI is inventing conversation-native inventory. If your agent can invent a fit, a price, or a ship date, you have a compliance problem with or without a Microsoft label.
Mistake seven: confusing “Ask ChatGPT about this ad” with “chat with the brand agent.” OpenAI’s Help Center and XBorder Insights separate those paths. One is the user interrogating the offer with ChatGPT. The other is messaging the advertiser’s agent. Creative, legal, and analytics need different labels, or you will attribute the wrong conversation to the wrong owner.
Second-order effects
Media plans will grow a new row between impression and site. Agencies that only sell landing-page CRO packages will lose share to teams that can design agent briefs, product fact packs, escalation trees, and chat-to-CRM maps. Accuracast’s WhatsApp analogy is the staffing hint: someone has to own qualification rules the way a WhatsApp reply team owns response SLAs.
Retail and DTC brands will fight a quiet ownership war. Chowdhury’s Digiday quote already names it: eyeballs and dollars versus first-party capture. Expect some brands to keep Sponsored Agent as upper-funnel conversation and force a hard handoff to site for checkout. Others will lean into in-chat commerce if OpenAI ever exposes it. Digiday frames the platform history correctly. Google’s text link, Facebook’s feed ad, TikTok’s in-feed video each defined an era. OpenAI is hunting for the conversation-native equivalent. Early pilots shape the metrics the industry will later pretend were always obvious.
Privacy and brand safety teams get a new diagram. Private ChatGPT context stays off-limits to advertisers per OpenAI’s FAQ. Messages deliberately sent to the advertiser do not. That boundary only works if the UI makes the switch unmistakable. If users think they are still in a private assistant thread while talking to a vendor agent, you inherit consent theater and support tickets. Build the disclosure copy before you scale spend.
Measurement vendors and in-house analytics will invent “conversation quality” scores before they invent trustworthy chat-to-purchase attribution. Resist vanity. eMarketer’s caution is right for Q4 planning: treat this as a personalized advertising experiment, not an immediate sales engine, until you can see intent signals and commercial outcomes without inventing them.
What to do this week
- Rewrite the ChatGPT ads scorecard. Split conversation start, qualified intent, human or site handoff, and purchase or lead. Kill any dashboard that only shows CTR to landing page for click-to-chat units.
- Build an agent fact pack before creative. SKUs, dimensions, ship windows, returns, and “we do not know” rules. Wayfair-style accuracy and transparency guards belong in the media brief, not in a postmortem.
- Map handoffs like WhatsApp ops. Who answers when the agent fails. What triggers a human. What URL or cart state the agent is allowed to send. Divecha’s Digiday comparison only works if someone owns the queue.
- Align legal on the two-chat model. Cite OpenAI Help Center language in the brief: no private chat access for advertisers; messages sent to the advertiser are visible to them. Spell which CTA opens which path.
- Keep holiday tests limited-scale. Digiday’s advertiser consensus on ChatGPT ads is still early ROI. Chowdhury’s seamless-experience “if” is your kill switch. Do not move core Q4 budget until chat-to-outcome reporting exists in your stack.
- Add a same-week creative compliance pass. Use Microsoft Advertising’s September 14 disclosure guidance as a checklist for any AI-assisted assets in parallel channels, and apply the same honesty bar to agent claims inside ChatGPT.
- Brief the board in operator language. Not “we bought AI ads.” Say: we are testing whether a paid click that opens a branded agent conversation beats a paid click that opens a page, on qualified intent and downstream revenue, at limited scale.
Sharp close
Platforms do not change advertising by inventing prettier banners. They change it by moving the click.
OpenAI’s Sponsored Agent pilot, as Digiday verified on September 14 with Wayfair on the record, moves the click from your website to a branded agent chat inside ChatGPT. Media teams that keep optimizing landing-page CTR are polishing the old exit while the new entrance is a conversation. Rebuild the funnel step, the agent brief, and the handoff rules, or you will keep winning a metric the format no longer needs.
The click is the conversation. Score it that way, or stop buying it.
