ChatGPT Ads Are Rolling Out in Taiwan, and OpenAI's Health, Financial and Legal Eligibility Table Is Far Stricter Outside the US

本文另有中文版：ChatGPT 廣告 9 月起在台灣陸續開放：資格表只分美國與其他國家，健康、金融、法律兩欄差很大

OpenAI’s eligibility table for health, financial and legal advertising has two country columns, US and All Other Countries. The word Taiwan does not appear anywhere on that page, even though OpenAI announced on September 23, 2026 that ChatGPT Ads were rolling out in Taiwan.

Taiwan is not the US, so the plain reading is that the All Other Countries column applies. OpenAI does not name Taiwan, and these pages do not say whether there are exceptions.

OpenAI has written down a lot about how these ads work: who sees them, how an ad gets chosen, how an advertiser account gets approved, and what the smallest daily budget is.

It has not written down whether Taiwan gets any exception to the table, how many Taiwanese users are on the plans that show ads, what results advertisers get here, or what a click costs. This article sorts the two piles, using OpenAI’s announcement and Help Center pages as they stood on September 24, 2026.

Taiwan is on OpenAI’s rollout list, and OpenAI still calls the program a test

The September 23 announcement says ChatGPT Ads “will begin rolling out” across Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan. It follows earlier launches in Australia, New Zealand, Japan, South Korea and India. OpenAI says the ads are now available in more than 60 countries and that tens of thousands of advertisers are running them.

Rolling out describes a process, not a finished state. The announcement gives no date for when the rollout in Taiwan is complete, and the Help Center goes further by framing the whole program as a test. One page says availability “may continue to evolve as testing expands.” Another notes that during the test ads do not appear in the ChatGPT Atlas browser, and the advertiser dashboard is named Ads Manager Beta.

OpenAI also says it will keep expanding over the next few months, into new markets and with new formats, optimization tools and measurement solutions. By OpenAI’s own account, then, the rules below describe one moment in a moving program.

One figure in the announcement invites misuse. OpenAI says ChatGPT Ads passed a $1 billion annualized revenue run rate less than 200 days after launch, announced at the end of August. That number is OpenAI’s own, it covers the whole world, and it says nothing about Taiwan or about where advertising budgets come from.

On the user side, Free and Go accounts may see ads below answers

OpenAI’s Help Center says ads may appear for users on the Free and Go plans. Plus, Pro, Business, Enterprise and Edu accounts will not have ads. Ads can appear below the end of a response, labeled as sponsored and visually separated from it.

A response can be followed by one or more ad units. OpenAI states that ads are not integrated into responses, that they run on systems separate from the chat model, that advertisers have no ability to shape, rank or alter ChatGPT’s responses, and that ads do not influence the answers.

Some situations stay ad-free: Temporary Chats will not show ads, and OpenAI says accounts identified as belonging to people under 18 do not get ads, judged from account-level age information and age prediction where available. It also says ads are not eligible to appear near personal health, mental health or political topics, and that it currently does not allow political advertising in ChatGPT.

Logged-out use is not an exception, since OpenAI says ads shown in logged-out experiences are appropriate for all audiences and that mature-content ads are limited to eligible logged-in users. It adds that an ad’s language may differ from the conversation’s or the interface’s language.

Users have controls that do different jobs. Turning off personalization changes how ads are selected, but ads still appear, based on the current chat thread, while Ads-Free removes ads at a price: on the Free plan it means lower usage limits and reduced feature access, such as fewer messages and no access to some tools such as image generation or deep research.

Go is a separate case. OpenAI says Ads-Free is a Free-plan option, so a Go user would have to switch back to Free to use it, or move to a plan that is already ad-free, such as Plus or Pro.

For a business, the practical meaning is about audience. Ads may reach people on Free and Go accounts, and the pages say Plus, Pro, Business, Enterprise and Edu accounts will not have them. OpenAI also says ads may appear only in regions where ads are being tested, and that eligible users see clear in-product information. It has not said how many Taiwan users are on Free and Go.

OpenAI’s eligibility table sorts health, financial and legal categories into a US column and an All Other Countries column

The table is called the Ads Policy Eligibility Table for Highly Regulated Areas. It sits at the bottom of OpenAI’s help page on onboarding and policy problems and covers health, financial services and legal services. Its country columns are US and All Other Countries. Taiwan is named in the announcement and on the Ads Manager availability list, but not in this table.

Nine rows show how the two columns can differ.

Source: OpenAI Help Center, “Troubleshooting Common Onboarding and Policy issues,” snapshot of September 24, 2026.

Across the full table the picture is the same. In the All Other Countries column, 15 of the 16 health rows read Not allowed and the exception is Health software and infrastructure, while of the 18 financial rows, 11 read Not allowed, 6 read Restricted to Select Approved Advertisers, and 1, Financial software, reads Allowed. Legal services reads Not allowed, whereas in the US column, legal services is Allowed, on the condition that US advertisers are licensed to practice law in the US when serving US users. Across all 35 rows, the two columns differ in 18 and match in 17, and the US column is not a green light either: 11 rows read Allowed, 12 read Restricted to Select Approved Advertisers and 12 read Not allowed.

Two rows have names that look close and results that are not. Health software and infrastructure is Allowed in both columns. Consumer medical devices, wearables and apps is Allowed in the US column and Not allowed in All Other Countries.

Restricted to Select Approved Advertisers is not a green light either. A note above the table says certain categories are restricted to select approved advertisers, that OpenAI is working with a small group of advertisers, and that it is not accepting applications for approval in certain policy categories. The note does not say which categories, and another passage on the same page adds that a registered or licensed business is “not automatically eligible for every advertising category or market.”

OpenAI’s pages do not name Taiwan in the table. Read literally, the heading gives two options, and Taiwan is not the US, so All Other Countries is the column that would apply. The table itself does not say whether US refers to the advertiser or to the audience. For an advertiser and an audience that are both in Taiwan, either of the two readings the pages allow lands on All Other Countries, and they only diverge across borders, such as a company based in Taiwan advertising to US users or a company based in the US advertising to Taiwan users. The target-country reading comes from a troubleshooting message, which tells advertisers whose ads cannot serve in a selected country to compare the campaign’s target countries with the allowed categories.

This is an inference, not something OpenAI states.

The reading would change if OpenAI published a separate table for Taiwan, if a Taiwan-facing ad in a Not allowed category were approved, or if a rejection message cited a different rule.

If the reading holds, the rows land on businesses unevenly. A dietary supplement seller or a clinic offering minimally invasive cosmetic procedures would find Not allowed beside its category. A bank taking deposits or an insurer selling non-health insurance would find Restricted to Select Approved Advertisers, a law firm would find Not allowed, and a company whose category is Health software and infrastructure, or Financial software, would find Allowed.

Categories outside the table are not covered by these pages, so the table cannot be read as a list of who else may advertise. OpenAI’s full Ads Policies text is a separate document and is not covered here.

An eligible Taiwanese business can start an Ads Manager account, but access waits on review, and that answers a different question

Three questions get blurred here, and OpenAI’s pages answer them separately. The first is where the advertiser is based: the Ads Manager availability page lists Taiwan as Available and says the legal entity that will advertise and be billed must be based in a listed country, and the country tied to a ChatGPT login does not override that.

The second is where the ads will run: a troubleshooting message says the check compares the campaign’s target countries with the allowed categories, and the table itself does not define whether US means the advertiser or the audience. The third is who actually sees ads, which is the Free and Go rollout described above, and that rollout is still under way.

Answering one does not answer the others, and OpenAI’s troubleshooting page treats account verification, account approval, ad review and ad delivery as separate checks and says that clearing one does not clear the rest.

Context hints can describe keywords, but they are not exact-match keywords

OpenAI’s own advice on writing ads begins by saying that ChatGPT “requires a different approach than keyword-based platforms.” The system picks ads from several signals: the context and intent of the current conversation, the ad’s landing page, title and copy, the advertiser’s context hints and targeting selections and, when personalization is on, signals from the user’s wider ChatGPT use.

Context hints are short descriptions added at the ad-group level. OpenAI’s wording lets hints describe conversations, topics or keywords where a product may be relevant, says they help guide matching, and adds that they are “not exact-match keywords” and do not guarantee delivery in specific conversations. The advice for writing one is a clear, natural phrase about a single idea, not a list of disconnected keywords. One of OpenAI’s illustrative examples for running shoes is a sentence about beginners training for a first 5K, not a search term.

Hints also do not control where an ad runs. OpenAI’s example contrasts two sentences: a hint that describes residential plumbing services available in Chicago is a business description, while an instruction to show the ad only to people in Chicago is a delivery instruction, and hints cannot enforce geographic limits, schedules or exclusions.

Delivery controls live at campaign level: target countries are required, while platforms (Android app, Android web, Desktop web, iOS app and iOS web) and custom audiences are optional. Campaigns other than product-feed campaigns can also target regions, cities and postal codes, though OpenAI says location availability may vary by country and the pages do not say what Taiwan offers, while new product-feed campaigns are limited to country level.

The landing page does double duty, because OpenAI says its crawlers check pages submitted as ads for safety, and that it may also use content from the landing page to decide when the ad is most relevant to show. One Help page says OAI-AdsBot must be allowed and recommends allowing OAI-SearchBot as well, and another page words the rule more strongly for both bots. App-store links and pages that need a login may not give crawlers enough content to work with.

Among eligible ads, OpenAI describes a “relevance-weighted, second-price auction,” and says relevance and advertiser bids both count when it decides which ad to show first.

Put together, the landing page is both a safety check and an input to matching. For a site owner, preparing a page for an ad and preparing it to be crawled and understood may be harder to treat as separate jobs.

This is an inference, not something OpenAI states.

One fact cuts against it: OpenAI says ads run on systems separate from the chat model. The inference would fall if live campaigns show that the landing page has no effect on matching.

This inference concerns the landing page as a matching signal for ads. It is a separate matter from OpenAI’s statement that ads do not influence ChatGPT’s answers, and it says nothing about how ads relate to those answers.

OpenAI names three official channels plus self-service, and the self-serve route runs through a review queue

The announcement names three channels: OpenAI’s Ads Solutions team, agency partners, and technology partners. The agency partners it names include dentsu, Havas Media, Omnicom Media, Publicis Groupe and WPP. Self-service access through Ads Manager comes on top of those for eligible businesses, and the availability page adds that larger advertisers with more complex needs can contact the OpenAI Ads team.

On the self-serve route, each advertiser needs its own account, and the details must be those of the entity that will advertise and be billed. Country, billing currency and time zone cannot be changed after the account is created. In the flow that starts inside ChatGPT, agencies creating advertiser accounts on a client’s behalf are not supported, although an agency can be invited once the client has created the account.

Then comes the wait. Account applications go through a rolling queue, and OpenAI says it cannot expedite account review or ad review at this time. Business verification, one step of the process, may take up to seven business days, after which OpenAI asks advertisers to contact Ads Support, and that figure covers verification, not the whole setup.

New self-serve accounts add a limit: OpenAI says some new accounts can initially advertise only in their home country, and other countries open up after identity verification is approved and the account reaches the required spend from campaigns in its home country, an amount OpenAI does not publish.

Read with the legal-entity rule above, this suggests an extra layer for a foreign advertiser who wants to reach Taiwan: the entity must be based in a listed country, and a new account may start out limited to its home country. This is an inference from two passages, not something OpenAI states, and it would change if a new account with a foreign legal entity can advertise in Taiwan directly in practice.

Two money figures appear in OpenAI’s pages. The Minimum Campaign Spend table lists a minimum daily budget of TWD 60 for accounts billed in Taiwan dollars, which is a floor and not a recommended budget. The table is organized by billing currency, not by country. Separately, for campaigns billed per click, OpenAI recommends starting with a maximum bid of $3 to $5 USD per click. That is a suggested starting bid, not a price actually paid, and OpenAI has not published one.

A floor and a suggested bid measure different things, and neither says what an effective campaign costs. One reading follows from the pieces above: for a small Taiwan business, the obstacles that show up first in OpenAI’s own pages are the review queue and the category table, not the daily minimum.

This is an inference, not something OpenAI states.

It would change if live campaigns turn out to be expensive to run effectively, or if reviews clear quickly. It also leans on the reading that the All Other Countries column applies to Taiwan, so part of it falls if that reading is overturned. And the category table covers only health, financial and legal advertising, so for a business such as a shoe shop or a coffee roaster, the first hurdle would be the review queue.

In OpenAI’s announcement and Help pages of September 24, there is no Taiwan user count, no results and no prices paid

The table below sets what the pages say next to what they leave open. The right-hand column is the reason this article is a snapshot and not a forecast.

Measurement exists but runs with delays. Ads Manager Beta currently reports impressions, clicks, spend, click-through rate, average CPC, average CPM and conversions, the last if conversion measurement is set up. Impressions, clicks and click-through rate refresh about every 15 minutes, spend can lag by 7 to 8 hours, and OpenAI asks advertisers to allow 24 to 48 hours for attributed conversions to appear and says totals may change. OpenAI says advertisers receive aggregated, non-identifying information, such as total views or clicks, and that new measurement solutions are coming.

Nothing in these pages says whether advertising budgets will move to ChatGPT from other platforms, and this article does not guess. The question needs data OpenAI has not published, and one announcement dated September 23 cannot supply it.

---

As of September 24, 2026, OpenAI still describes the program as a test, so the rules, tables and figures above may change, and OpenAI’s own pages are the place to confirm them.

This article summarizes OpenAI’s published announcement and help pages as of September 24, 2026. It is not campaign or purchasing advice.