On June 12, 2026, OpenAI switched on product-feed campaigns in ChatGPT. Advertisers upload a feed, between 1,000 and 2 million products, and the ad's title and description are pulled directly from that feed. No copywriter drafts the line. The model assembles it, surfaces it inside a conversation when someone asks for, say, running shoe recommendations, and routes the click onward. Early beta numbers look strong: one global DTC brand reportedly doubled CTR and halved CPC after moving from standard ChatGPT ads to feed ads. The numbers are easy to celebrate. The harder thing is to sit with what they change about who is actually doing the talking.
Five questions worth not answering too quickly
Socrates did not win arguments by being right. He won by asking the question the other person had skipped. Here are five that the feed-ad shift quietly skips over, and they are worth holding open rather than closing fast.
First: when the ad copy is generated from your feed, whose message is it? You wrote the feed. The model wrote the sentence. The user reads the sentence and forms an expectation. If that expectation is wrong, who authored the mistake, you or the system that paraphrased you?
Second: if doubling CTR is good, is it good because the ad is more honest, or because it is more persuasive? Those are not the same thing, and the metric cannot tell them apart. A click is a click whether it was earned by clarity or by a promise the page will not keep.
Third: you can see the aggregate performance data, not the individual conversations. So how would you know if the model is describing your product in a way you would never have approved? The absence of a complaint is not evidence of agreement. It might just be evidence that nobody who was misled bothered to write in.
Fourth: when you optimized ad copy by hand, you at least knew what you had promised. Now the promise is assembled per conversation, contextually, on the fly. Can you be accountable for a message you did not see before it shipped?
Fifth, and the uncomfortable one: if the feed ad and the landing page disagree, which one is lying? The feed you wrote, the sentence the model built from it, or the page the user finally lands on? They cannot all be the source of truth at once.
The unexamined ad campaign is not worth running
Socrates' line was that the unexamined life is not worth living. The marketing version is less grand but no less true: the unexamined ad campaign is not worth running, because you cannot improve, defend, or even understand a thing you have not looked at.
Feed ads make examination harder in a specific way. They move the moment of authorship away from you. Under the old model, the gap between your ad and your page was at least a gap between two things you had both written. You could read them side by side. You could see the message match or its absence. Now one side of that pair, the ad, is generated fresh from your feed data for each conversation, and the other side, the page, sits where it always sat. The ad-to-page congruence you used to be able to eyeball has become something you have to actively verify, because you no longer wrote both halves.
This is where the philosophy stops being abstract. Knowing your ads and your pages are aligned is worth more than hoping they are, and the feed-ad shift widens the distance between knowing and hoping. The widening is not dramatic on any single day. It is the same quiet process that produces alignment drift in hand-written campaigns, except faster, because the copy now changes per conversation rather than per quarter. Drift used to be something you let happen by neglect. With feed ads it happens by design, as the intended feature, and the only counterweight is a deliberate habit of checking. The model optimizes for the click. It does not optimize for whether the page delivers what the generated copy implied. That second job, the post-click experience, is still yours, and it is now harder to oversee precisely because the first job was taken off your desk.
Consider what the strong beta numbers actually measure. Double the CTR means twice as many people believed the generated promise enough to click. If the page behind it was already misaligned, the feed ad did not fix that. It scaled it. A higher click-through rate on a page that does not match is not a win. It is a more efficient way to pay for disappointment, and the aggregate dashboard will show it to you as success right up until the conversion rate tells the truth.
So the examined approach is not to distrust the automation. It is to verify the thing the automation cannot see. The model can write a thousand contextual ads from your feed. It cannot tell you whether the page they all point to honors what they said. That judgment, the creative governance layer, is the part that stays human, and arguably becomes more human as more of the production gets automated.
What this looks like in practice
The Socratic move here is not to stop using feed ads. It is to refuse to run them unexamined. Concretely, that means treating your feed and your landing pages as a single system that has to agree, and checking that agreement rather than assuming it.
AdAlign scores ad-to-page congruence across visual match, content consistency, and tone, on a 1 to 10 scale. For feed ads, that means you can take the copy the model is generating from your feed, set it against the page it routes to, and see where the two diverge before the divergence costs you. You are not auditing a copywriter anymore. You are auditing the agreement between what your feed says, what the model makes of it, and what your page finally delivers. The failure modes are specific and findable: a feed attribute the model leans on that your page never mentions, a tone the generated ad strikes that your page does not match, a product benefit the model surfaces that the page buries. None of these are exotic. They are the ordinary ways a post-click experience breaks, arriving through a new door. If you have ever seen a high CTR page that refused to convert, you already know the shape of the problem. Feed ads just remove the copywriter who used to catch it.
The honest framing is that automation did not remove the need for examination. It moved the examination downstream, from writing the ad to verifying the match. The teams that thrive on feed ads will be the ones who treat the generated copy as a claim to be checked, not a result to be banked.
Find out in 60 seconds whether your ads and landing pages are actually saying the same thing. Run your free ad alignment audit, three analyses, no card required. For the platform-policy version of this same question, where Google now ties ad serving to brand clarity, read how we'd fix Google's limited ad serving risk in 24 hours. And for the groundwork, see what is ad-to-page congruence.
Frequently asked questions
What are ChatGPT feed ads? ChatGPT feed ads are product-feed campaigns launched by OpenAI on June 12, 2026. Advertisers upload a feed of between 1,000 and 2 million products, and the ad's title and description are generated directly from the feed. The ads surface contextually inside conversations rather than against keywords.
Who writes the copy in a ChatGPT feed ad? The model does. It assembles the ad title and description from your product feed data per conversation, rather than a copywriter drafting fixed lines. You author the feed; the system authors the sentence. That split is why verifying the message against your landing page matters more, not less.
How do I keep feed ads aligned with my landing page? Treat the feed, the generated ad, and the landing page as one system that has to agree. Take the copy the model produces from your feed, set it against the page it routes to, and score the congruence across visual, content, and tone before divergence costs conversions. AdAlign does this on a 1 to 10 scale.
Do higher click-through rates from feed ads mean better performance? Not on their own. A higher CTR means more people believed the generated promise enough to click. If the landing page does not deliver that promise, a higher CTR simply scales the mismatch. Click-through rate measures persuasion, not whether the page honors what the ad implied.