Buying ChatGPT ads just got easier. That is not the same sentence as "ChatGPT ads just got better."
On November 2, multiple outlets reported that OpenAI has signed deals with Magnite and PubMatic so advertisers can set budgets, submit bids and upload creative through those supply-side platforms while OpenAI keeps control of delivery. The same day, Search Engine Land, Search Engine Journal and PPC Land all walked through an October working paper from WU Vienna and the Frankfurt School that puts a hard number under the other half of the story: in U.S. traffic, a paid ChatGPT visit returned about $0.89 in revenue against $2.45 for a paid Google Search visit. That is a 64% shortfall before you even look at the CPC.
The media job this week is not a scale brief because a new door opened in the stack. It is a CPC and ROAS fence, plus a holdout, while you decide whether the channel clears your own parity price.
Two headlines, one operator problem
The Kaiser and Schulze paper, "Can Advertising Fund Free AI?," follows OpenAI's country-by-country ad rollout with shop-side analytics from Grips Intelligence. The observation window runs from late June 2025 through August 22, 2026. Nearly 10 billion sessions across 1,898 websites sit in the organic panel. The paid financial sample is smaller: 102 advertisers with complete sales tracking, 87 of them buying both ChatGPT ads and paid Search in U.S. traffic.
The load-bearing paid numbers, corroborated by Search Engine Land on November 2 and by PPC Land's long read the same day, are blunt:
- Revenue per visit in U.S. traffic: $0.89 from paid ChatGPT versus $2.45 from paid Search.
- Conversion rate: 0.37% versus 1.17%.
- Among the 87 dual-channel advertisers, 97% earned less revenue per ChatGPT visit than per Search visit.
- At Search's rate of return, a ChatGPT click is worth about $0.26. The largest publicly known account, reported via Opascope and cited in the paper, paid $1.72. OpenAI's own ads basics guidance has recommended starting CPC campaigns with a maximum bid of $3 to $5.
Read that again slowly. The parity price is a quarter. The known market price is nearly seven times that. The recommended start bid sits even higher. Unfavorable unit economics do not become favorable because you can now push the same inventory through Magnite instead of only through Ads Manager.
What "easier to buy" actually changes
eMarketer's November 2 briefing, drawing on The Information, describes the Magnite and PubMatic arrangements as a way to fill unused ChatGPT ad space and to let brands work inside tools they already know. Initial transaction volumes are still small. Prior buying and measurement partners already included Amazon Ads, Criteo, LiveRamp and StackAdapt. AdExchanger's November 3 roundup put the same deals in the classic early-platform pattern: open the inventory to intermediaries first, then decide how walled the garden becomes later.
ContentGrip's November 3 write-up is useful for what it refuses to claim. Deal terms, broad availability, formats and partner economics were not disclosed. An OpenAI spokesperson, as reported, framed the point as making ChatGPT ads easier to manage through familiar workflows. OpenAI still decides where the ad lands.
That split matters for a media lead. A new SSP path can shrink the operational friction of getting a line item live. It does not rewrite the conversion rate of the click that arrives. It does not invent prompt-level reporting that several trade outlets say buyers still want. eMarketer notes, citing Adweek, that ChatGPT ads still lack search- and prompt-level reporting; that Adweek piece was not independently opened for this draft, so treat the gap as "widely reported buyer complaint," not as a line you heard from Adweek yourself.
Easier buying is a logistics upgrade. ROAS is still a math problem.
The study is not a vibes piece
Kaiser and Schulze are careful about what they can and cannot prove. The paid comparison is descriptive. Advertisers chose to buy ChatGPT ads; they were not randomly assigned. The data is last-click. It does not show prompts, impression sets or whether the visitor was on Free, Go, Plus or Pro. Campaigns that never earned a click are invisible. Maximilian Kaiser discloses that he is a scientific adviser to Grips Intelligence, which supplied the data free of charge under an agreement that, per the paper, gave the company no approval rights over the analysis.
Those limits do not erase the pattern inside the sample. Seventy-three of the 87 dual-channel advertisers recorded zero paid ChatGPT transactions despite 18,726 paid ChatGPT sessions between them. Fourteen sites carried all 570 transactions. One advertiser accounted for 66% of paid ChatGPT purchases in the financial sample; dropping that advertiser made the shortfall worse, not better. Paid ChatGPT was the only major channel in their comparisons that underperformed direct traffic on conversion, order value and revenue per session.
Organic ChatGPT referrals are a separate ledger. After ads launched in a market, unpaid ChatGPT referrals fell an estimated 5.9% over the first 13 weeks relative to organic Search and to the never-treated EU, with a confidence interval that still includes zero. Once the authors adjust for faster ChatGPT adoption outside Europe using OpenAI's own regional adoption index, the estimated decline moves to roughly 7.2% to 8.4% and the intervals exclude zero. Visitors who still arrived converted about as before. The fact sheet's punch line is that even separated sponsored cards may have moderately reduced the free referral volume shops care about, while the paid cards themselves under-earn.
You do not need OpenAI's 2030 ad-revenue fantasy for the weekly decision. You need your own revenue per visit next to your own CPC.
Why buyers still want in
None of this means the channel is empty of interest. Bank of America survey cuts, as relayed by eMarketer via TheStreet, sketch the tension cleanly enough for a planning meeting: only about 6% of advertisers and agencies said ChatGPT's AI was increasing ROAS fast enough to justify new spending, against 15% for Facebook and 33% for Google Search. Separately, other BofA-related coverage has ChatGPT near the top of "where budgets might grow next." That is the classic early-channel posture. Marketers want a seat. Marketers do not yet believe the seat pays.
OpenAI has said publicly that ChatGPT Ads passed a $1 billion annualized run rate by late August 2026 and that more than 100,000 advertisers had run campaigns as the footprint reached 89 countries. Platform milestones are not merchant unit economics. Search Engine Journal, covering the Kaiser paper on November 2, also noted OpenAI's own FAQ admission that ChatGPT Ads "does not yet have performance benchmarks across advertisers, industries, or campaign types." When the seller will not publish the benchmark table, outside shop data is what you have.
The public success story the paper checks against Opascope's June account (about $1.72 CPC, stronger conversion) does not travel. In the 102-site financial sample, the median website recorded 0.00% conversion and $0.00 revenue per paid ChatGPT visit. One site matched the public story on both conversion and revenue per visit, and it did so on nine paid sessions.
So yes, someone somewhere is making ChatGPT ads work. The distribution says that person is not you by default.
Two ledgers, one fence
Treat ChatGPT like a channel with two books that must stay separate.
Ledger one is organic referral value. If unpaid ChatGPT traffic converts near Search for your category, protect it. Do not let a soft paid test cannibalize the free answers that already send buyers. Watch organic ChatGPT sessions and revenue in the weeks after any paid flight turns on in a market.
Ledger two is paid visit value. Compute a parity CPC the way the paper does, on your own numbers: take ChatGPT revenue per visit, divide by Search revenue per visit, multiply by your Search CPC. That is the price at which tracked last-click returns match. Anything above it needs a reason that shows up in your data, not in a partner deck: incrementality, assisted conversions your last-click model misses, or a creative test with a hard stop date.
The SSP door changes how ledger two gets funded. It does not merge the ledgers. If Magnite or PubMatic makes it one click easier to pour budget in, the fence has to get clearer, not softer.
What to do this week
Write the fence before anyone opens the door. In the shared media doc, put three numbers on one line: your trailing Search CPC, your ChatGPT revenue per visit (or proxy CPA), and the implied parity CPC. If you have no ChatGPT paid data yet, use the paper's U.S. ratio as a planning prior ($0.89 / $2.45) against your Search CPC, then replace it the week real clicks land. No parity number, no scale conversation.
Cap first flights at a holdout, not a "learning" blank check. Keep ChatGPT at a fixed share of search or of total paid, with a hard CPC or target ROAS ceiling tied to parity plus a small experimental premium you can defend. "We got SSP access" is not a premium. "We are measuring lift for two weeks" is.
Demand the measurement you are missing before you raise the cap. Confirm which events OpenAI, Magnite or PubMatic can return, which attribution window you are on, and whether prompt- or query-level reporting exists for your seat. If you only get aggregate outcomes, your fence has to be stricter, because you cannot diagnose losers by intent cluster.
Split organic and paid ChatGPT in analytics this week. Tag paid clearly. Compare organic ChatGPT revenue per session before and after paid turns on. If organic softens while paid under-earns, you are paying twice.
Do not reallocate Search on a hope curve. Digiday-style chatter about moving 10% to 30% of Search into other pots, including ChatGPT, is a rumor climate, not a brief. Move money only after a ChatGPT flight clears your fence on tracked revenue or on a pre-registered holdout. BofA's own split mood (curiosity up, proven ROAS pace down) is the warning label.
Brief creative for the card you actually buy. ChatGPT's sponsored units still sit labeled below the answer for Free and Go users. That separation may be why organic trust held in the study and why paid clicks under-earn. Do not write Search-style headlines for a card that never owned the answer. Write for a second look after advice already landed.
OpenAI is doing what every young ad platform does: open more pipes so unused impressions find budgets. Magnite and PubMatic are doing what SSPs do: stand near the new inventory while the walls are still negotiable. Your job is narrower. Paid ChatGPT visits, on the best outside evidence we have this week, earn far less than Search. Buying got easier on November 2. Pricing did not. Fence the CPC, run the holdout, and scale only after your own ledger says the door was worth walking through.
