OpenAI's published position on ChatGPT advertising is that advertisers do not see conversations. A 404 Media report from 15 October 2026 says that is not the whole picture: in its own tests, advertisers could aim ads at specific people using an email address, a phone number or a phone's unique identifier, and could also target using what users type into ChatGPT. Both statements can be true at once, because they describe different systems. This article separates them, explains why uploading a hashed list does not make it anonymous, and sets out what the settings OpenAI has published do and do not reach. Our reading is limited by the sources: 404 Media's article is paywalled and we could read only its introduction.

What was reported, and what we could confirm

According to the introduction of the 404 Media piece, its tests showed that advertisers can place ads in ChatGPT aimed at specific people using their email address, phone number or a unique phone identifier. In some cases, it says, the targeting uses personal data that the advertiser uploaded to OpenAI, which many users may not realise. It calls the approach novel because targeting can also draw on the prompts users enter, and says prompts may be reviewed by humans, citing an earlier 404 Media investigation. It cites one billion monthly active users, attributed to a June 2026 Reuters report. We could not read the test details, the targeting options listed, or any response from OpenAI, so we do not characterise them.

The human-review point deserves care. The article attributes it to prior reporting and, in the part we could read, offers no new evidence. We therefore treat it as a claim made by 404 Media, not as something we have established.

System one: audience lists

The audience-list feature is documented independently. PPC Land reported on 14 May 2026, based on screenshots shared on LinkedIn by Juozas Kaziukenas, that OpenAI was preparing custom audiences for ChatGPT ads in a gated rollout. Advertisers upload a CSV or TXT file of up to 512 MB containing email addresses or phone numbers, either raw or pre-hashed with SHA-256, and OpenAI normalises the identifiers during processing. In include mode, ads are limited to users in the uploaded audience; in exclude mode, they are kept from reaching them. Lookalike audiences were not available at that point. The same report says OpenAI opened its self-serve Ads Manager beta to all US businesses on 5 May 2026, and that ads were then running in the US for Free and Go users.

Two things in that report are not specified and matter for privacy: how quickly matching occurs and what match rates look like, and what retention windows apply to uploaded lists. Those are the kind of detail an advertiser would want and a user would want to know. Neither is available in the sources we read.

Why hashing is matching, not anonymity

A hash function turns an input into a fixed-length string; SHA-256 is a standard one. It is one-way, meaning you cannot compute the email back from the string. That sounds like protection, and it is the reason advertisers can say they never send raw addresses. But a hash is deterministic: the same email always produces the same string. If the platform holds a user's email, it can hash it the same way and compare strings. A match means the platform now knows that the person on the advertiser's list is the same person as this account.

This is why hashing protects the data in transit and in storage but does not hide the person from the party doing the matching. The platform already knows its own users' emails and phone numbers. In effect, the advertiser says: show this ad to everyone whose identifier appears on my list. The platform does it, and the advertiser learns, at most, aggregate delivery results. The user's identity is never disclosed to the advertiser, but the user's membership in the advertiser's list is used to decide what they see. Whether that counts as private depends on whether one's concern is disclosure to the advertiser or use of the advertiser's data about you.

There is also a limit on how well this works. An email is a weak key if the user signed up with an address the advertiser does not have, and a phone number depends on what was provided. Match rates are therefore not 100%, and we do not have OpenAI's. That cuts both ways: it means list targeting is not total, and it means an advertiser cannot be sure who exactly it is reaching.

System two: conversation context

The second system is different in kind. Instead of using who the user is, it uses what the user is doing. OpenAI's help page says that with personalisation turned off, you still see ads based on your current chat, but not on other chats, memories or ad history. That wording implies the current conversation is used as a signal even in the minimal setting, and that the personalised setting uses more. This is contextual targeting, the same principle as showing a travel ad beside a page about travel, applied to a conversation.

What makes ChatGPT unusual is how much a conversation reveals. A search query is a few words; a prompt can be paragraphs about someone's job, finances, relationships or plans. OpenAI's page restricts ads from appearing near personal health, mental health or political topics, and says political advertising is not currently allowed. Some health and financial-services advertisers may qualify under strict criteria, but their ads are still not placed near sensitive user content. How that classification works in practice, and how reliably, is not described in the page.

What OpenAI says, and what it leaves out

The help page says Free and Go users may see ads, while Plus, Pro, Business, Enterprise and Edu accounts do not. Advertisers receive only aggregated, non-identifying performance data such as total views or clicks, and the page states that conversations are not shared with advertisers. Ads are not shown to accounts identified as under 18, and Temporary Chats show none. Under Settings and Ad Controls, users can turn off personalisation, dismiss or report ads, see why an ad was shown and clear ads data. Personalised ads were not initially available in the EEA or Switzerland.

The page does not mention customer-list matching. That is the gap. Every statement on it can be accurate and still leave the list system undescribed: advertisers do not see chats, advertisers receive only aggregates, and yet an advertiser can choose who sees an ad by uploading a list. The first two statements are about what flows to the advertiser; the third is about what the advertiser can tell the platform to do. Readers who take the help page as a complete description of how ads work will miss it.

We also cannot say whether the Ad Controls personalisation switch affects list-based targeting. The page describes turning off personalisation in terms of chats, memories and ad history. Whether an advertiser's own uploaded list continues to apply when personalisation is off is not stated in the sources we read, and 404 Media's paywalled article may address it.

The scale question

One billion monthly active users is the figure 404 Media cites, but ads reach only the Free and Go tiers. We have no figure for how many of the billion are on those tiers, so the share of users exposed to these targeting systems cannot be calculated from the sources. That is worth stating because the headline number suggests a vast audience while the actual exposed population is unknown. It is also the case that the paid tiers are excluded, so a user who pays for Plus or above is not shown ads at all.

The counter-argument

OpenAI and advertising defenders would say this is conventional. Customer-list targeting has been standard on major ad platforms for years; advertisers upload lists of their own customers, and the platform matches them. On that view, ChatGPT is behaving like any other ad platform, with the added protections that conversations are not shared and sensitive topics are excluded. Users who dislike personalisation can switch it off, and those who dislike ads can pay.

The response is about context, not mechanism. The same technique means something different in a product where people confide in the assistant. The reasonable expectation attached to a chatbot is closer to a conversation than a browser, and the combination of identity matching and conversational context is what 404 Media calls novel. Whether that is a meaningful difference or an unfamiliar form of a familiar practice is a judgement, and we do not have the evidence to settle it.

What a user can do

  • Check Settings and Ad Controls and turn off personalisation if you do not want other chats, memories or ad history used.
  • Use Temporary Chats for conversations where you do not want ads shown, since OpenAI says they show none.
  • Remember that turning off personalisation, per the help page, still leaves ads based on the current chat.
  • Treat the phone number and email you attach to the account as the keys an advertiser's list can match against.