One sentence in OpenAI's announcement does the real work. It says "People also have control over ad personalization, with ad-free paid plans available for those who prefer not to see ads" [7]. Two clauses, two different objects: the first governs which ads get selected, the second governs whether any are shown at all. The dev.to analysis makes the translation explicit, and it is the correct one. Personalisation is a relevance dial. The only switch that removes ads is the subscription [8].
The footprint is worth counting. Testing started in the US in February, and OpenAI says it has since reached eight further markets [5]. Add the 31 European ones [1] and the total is 40 [12].
What arrives with them is a full demand-side advertising product. OpenAI lists bidding "beyond CPM and CPC" for conversion optimisation, geo-targeting, custom audiences, and measurement through the OpenAI Pixel, a Conversions API and third-party integrations [6]. dev.to calls that the standard furniture of the online-advertising industry, assembled inside a chatbot [15]. Six months of testing does not usually buy a measurement API, so some of this was built before the pilot returned a verdict.
The privacy commitments deserve to be read as carefully as the toggle. Dave Dugan, OpenAI's VP of Ads, framed his in terms of access: advertisers will not have chat histories, and conversation content will not be shared with them [10]. The announcement adds that conversations stay private from advertisers, customer data is not sold, ads are always labelled and separate from the answers, and advertising does not influence the answers [11]. dev.to's own read is that this is not the surveillance-advertising model that hollowed out the rest of the web, and on the quoted evidence that is fair [17]. Every one of those commitments, though, constrains the outbound direction, from OpenAI toward the advertiser. None of the quoted lines say which signals feed the custom-audience builder, how long they persist, or whether switching off personalisation stops those signals being derived or only stops them being used at selection time. That distinction is what a reviewer has to settle before calling the design safe for their own users.
For anyone whose product routes people into consumer ChatGPT, the practical change is that the frame around your instructions is now inventory. Free and Go users get ads near the answers; Plus, Pro and Enterprise users do not [4][14]. Plan tier is a rendering variable that belongs to the user. At the 23 euros a month dev.to cites for Plus, the ad-free surface costs 276 euros per seat per year [9][13], and you cannot buy it on someone else's behalf.
So the subscription pitch in Europe now sells a quieter screen alongside the better model. A quieter feed is a reason to keep paying, and that is exactly why the toggle sits where it does.
