ChatGPT advertisers can target users with uploaded email, phone and Android ID lists

What 404 Media tested

On 15 October 2026, Joseph Cox at 404 Media published How ChatGPT Targets Ads To You Based On Your Prompts, Phone Number, and Email. To see how the system works from the advertiser's side, 404 Media ran two ads inside ChatGPT. Both read "ChatGPT isn't private. OpenAI hires humans to read users' chats. Learn more," and linked to the outlet's earlier reporting on contractors reviewing real prompts.

The report's central point is not that the tools are unusual. Customer-list targeting is standard across the ad industry. Its point is that ChatGPT's roughly one billion monthly users may not realise they can be singled out by personal data an advertiser uploads, or that the prompt they are typing can shape which ad appears. Zach Edwards, co-founder of the ad-research firm DecryptAds, told 404 Media that OpenAI is "boiling the frog" on privacy.

OpenAI started showing ads to Free and Go users in the United States in February and has since expanded to other markets. This month it also began placing ads next to image-generation results, which AI Lookout covered in ChatGPT's visual ads. The ad usually sits at the bottom of a related answer.

Context hints: targeting by what you ask

The first targeting lever is the conversation itself. OpenAI's campaign targeting documentation describes context hints as extra information about an ad group: the products, use cases or needs that the creative and landing page might not fully cover. The docs say hints are not exact-match keywords and do not replace geographic, platform or audience targeting. An ad group can carry up to 2,000 hints.

404 Media used this lever directly. It told OpenAI its ad was relevant to people searching for topics about privacy, the privacy of ChatGPT chats, or current technology and AI. In practice that means a prompt about chatbot privacy could surface an ad about chatbot privacy. The model decides the match; the advertiser only describes the intent.

Outsiders are already mapping that behaviour. 404 Media points to ChatGPT Ad Library, a third-party site that sends realistic prompts to ChatGPT and records the sponsored cards that come back. The site says it has captured at least 43,410 ads across 415,289 placements, and stresses that the figures reflect its own sample, not OpenAI's data.

Custom audiences: targeting by who you are

The second lever is identity. OpenAI's Custom Audiences documentation lets advertisers turn customer or prospect lists into audiences, either in Ads Manager or through the Ads API. A list can be a UTF-8 CSV or TXT file up to 500 MB. Supported identifiers are email addresses, phone numbers in E.164 format, SHA-256 digests of normalised emails or phone numbers, and Google Advertising IDs, the device identifiers on Android phones.

Advertisers can mix identifier types in one CSV, and OpenAI counts each matched user once even if several identifiers point to the same person. Audiences can be created, added to, removed from, replaced wholesale or merged, with 2 to 64 existing audiences combined into a new one. Single-type files can hold up to 5 million identifiers; automatic resolution handles larger files within the same size limit.

The docs also set conditions on the data. Advertisers must upload only first-party data they have the right to use for ads, must not upload broker-sourced data, and should confirm consents, notices and legal bases and get privacy or legal sign-off. Those are contractual rules, enforced through OpenAI's Ad Tools Terms, rather than technical checks the docs describe.

The 25,000-user threshold and what advertisers see

The main safeguard is a size floor. To include an audience in a campaign, or to raise bids for it, an advertiser needs at least 25,000 matched users, and OpenAI warns that uploading 25,000 identifiers does not guarantee 25,000 matches. Exclusion works differently: ready audiences of any size, including empty ones, can be used to keep people from seeing an ad.

Advertisers do not see who matched. The API reports identifier and matched-user counts only as ranges, such as under 25,000 or 1 million to 5 million, and says exact counts and individual membership results are not exposed. Finer ranges are available only for audiences of 100,000 matched users or more. Bid multipliers can run from 0.1x to 10x, and if a user falls into several audiences the highest multiplier applies.

OpenAI repeated the point to 404 Media. A spokesperson said being on an advertiser's list does not by itself cause an ad to be shown, that inclusion and bid adjustments require at least 25,000 matched users, and that advertisers cannot see which individuals matched or pick specific people within an audience.

Raw or hashed, and how long files are kept

One question stayed open. Meta hashes raw emails and phone numbers when advertisers upload them. 404 Media asked OpenAI whether it does the same, and the company did not answer directly, instead repeating that advertisers can upload either raw or hashed data. The public docs say the API trims and lowercases raw emails and hashes Android IDs internally, but they do not say what happens to raw email and phone values after matching.

On retention, an OpenAI spokesperson told 404 Media that uploaded files are deleted after processing, which usually happens within 24 hours. The developer docs tell advertisers not to treat an upload's file ID as permanent storage. They do not state how long matched audience memberships themselves persist.

What the personalisation switch does and does not do

ChatGPT users can turn off ads personalisation. According to the setting text 404 Media quoted, that stops ChatGPT from using past chats, activity and preferences to select ads. It does not stop ads, and the setting itself warns that ads may still be based on the current chat. In other words, the context-hint system keeps working even when personalisation is off.

That distinction matters because prompts often carry more personal context than a search query. The same tension shows up across AI products that read what users are doing, from ChatGPT's Chrome extension that reads open tabs to assistants with memory. Ads are one more consumer of that context.

Where these tools are not available

OpenAI's Custom Audiences page says the feature is not supported for campaigns targeting the European Economic Area or Switzerland, where personalised ads are not yet available. Users there still see ChatGPT ads where ads are offered, but advertisers cannot use uploaded lists to reach them.

Edwards argued in the 404 Media piece that the gap should be closed by law, not by settings: advertisers that share personal data with ad platforms such as OpenAI should have to register as data brokers and give people a clear way to remove their data. No regulator has said it is examining ChatGPT's ad tools, and OpenAI has not published a privacy assessment of them.

What users and advertisers should take from it

For users, the practical lesson is to treat a ChatGPT prompt on the Free or Go plans like a search box that can show ads. If an advertiser holds your email or phone number and has a large enough list, you can be included in an audience; even a tiny list can be used to keep you out. Turning off personalisation narrows the signals but not the current conversation.

For advertisers, the documentation is detailed and familiar to anyone who has used Meta or Google customer lists: first-party data only, a 25,000-match floor for inclusion, ranges instead of exact counts, and no EEA personalisation. What OpenAI has not yet published is a clear answer on hashing raw uploads and a retention policy for matched audiences. Those are the next details to watch.