ChatGPT Ads custom audiences for B2B: what the rules actually allow, and how company data feeds them

A custom audience is a reusable audience object you create in OpenAI's Ads Manager from identifiers your business already holds. The current workflow accepts email addresses, phone numbers and Google Advertising IDs, either in plain text or hashed with SHA-256. Since a September 2026 update one file may mix identifier types, audiences can be edited after creation through add, remove and replace operations, and the old five-million-identifier ceiling is gone. Files are CSV or TXT up to roughly 500 MB, and processing typically takes twenty to thirty minutes before the audience becomes usable.

Once an audience is ready you can apply it in three distinct ways, and they are not variations on a theme. Campaign-level inclusion restricts delivery to people in one or more selected audiences. Campaign-level exclusion suppresses them — existing customers, churned accounts, your own staff. Ad-group bid multipliers, documented from 0.1x to 10x, change what you are willing to pay when a matched user shows up; when several audiences match, the highest multiplier wins.

What you do not get is a customer database inside Ads Manager. You cannot see which individuals matched, you cannot pick a person out of the result, and the upload is a dated snapshot of your own source system. Collection, consent, corrections, suppression and retention all stay your responsibility.

An audience must reach 25,000 matched users before it can be used for inclusion or for bid adjustment, and OpenAI recommends 100,000 or more. Below that floor the audience is still created, and since the September 2026 change it can be used for exclusion — which is genuinely useful, because suppression lists are usually small by design.

The word doing the work is matched. A file with 25,000 rows is not a 25,000-person audience. Duplicates, malformed values, role addresses and identifiers OpenAI cannot tie to an account all fall out of the usable count. Match rates on business email addresses are typically lower than on the personal addresses people use to register consumer accounts, so a B2B team planning an inclusion campaign should assume it needs a multiple of the floor in raw records, not a comfortable margin above it.

This is where the arithmetic bites niche targeting. There are not 25,000 matchable dental practice owners in a single mid-sized country. If your ideal customer profile is narrow, inclusion targeting on ChatGPT Ads is arithmetically out of reach on its own — you either widen the definition, aggregate across several markets and adjacent job families, or you use the channel for exclusion and bid shaping rather than for gating.

OpenAI's documentation is explicit: upload only first-party audience data you have the right to use for advertising, and do not upload broker-sourced data. You are also expected to confirm you hold the required rights, notices, consents, permissions and legal bases before uploading, and to get privacy or legal sign-off for the use case.

That single line rules out the workflow most people imagine when they hear that a platform accepts list uploads. You cannot buy or licence a prospect file, upload it, and treat the platform as a targeting layer over purchased data. The compliant path is slower and, in practice, more durable: use company data to define and reach your market through channels that permit business outreach, and let the relationships you build there become the first-party list you are allowed to upload.

In other words, company data is upstream of ChatGPT Ads, not a shortcut into it. It tells you which organisations match your technical and firmographic profile, who decides there, and how to reach them by phone, email, post or messaging. What comes back — enquiries, webinar registrations, trial signups, customers, newsletter subscribers who opted in to marketing — is the material a custom audience can legitimately be built from.

Custom audiences are not supported for campaigns targeting the European Economic Area or Switzerland, because personalised advertising is not yet available there. This is the rule that quietly invalidates most European B2B plans written about this feature. A campaign aimed at Germany, Belgium, France, Denmark, Spain, Italy, Ireland or Luxembourg cannot use an uploaded audience today, no matter how clean or how large your list is.

Where it does work right now is outside that perimeter: the United States, Canada, the United Kingdom and Australia are the markets worth planning for. If your growth plan includes Europe, the honest sequencing is to build the list and the demand-generation engine now through channels that do work there, and treat ChatGPT Ads as a channel you switch on per market as availability changes.

Check the current position before you budget. This is a feature that has been rewritten repeatedly through 2026 — the minimum, the editability, the identifier rules and the size caps have all moved — and regional availability is the most likely thing to change next.

For software and service vendors, the hard part is not the ad platform. It is defining a target list that is specific enough to be worth advertising to. A vendor selling a migration service for a particular CMS does not want 'companies in retail'. They want companies of a certain size, in certain sectors, in certain countries, that show the technical characteristics their product replaces or complements.

That is a company-data selection problem. Filter on country and region, legal form, sector code, registration date, headcount and revenue bands, group structure, number of locations and register events such as new registrations, capital changes, director appointments and dissolutions. Layer technical attributes on top — the stack a company runs, the marketing automation or commerce platform it uses, the presence of specific tooling — and the result is a segment you can write one message for instead of ten.

The same segment definition then drives every channel at once. Direct mail to the registered office, outbound calling to the named decision-maker, email to business addresses, messaging where that is accepted, and lookalike or matched audiences on the ad platforms that allow them. One definition, five delivery mechanisms, one reporting view — and consistency between them is what makes attribution legible later.

Drimble publishes company records from official registers across the United Kingdom, France, Spain, Ireland, Denmark, Belgium, Canada, Australia, Greenland and India, with the firmographic filters, register events and business contact fields that B2B teams need. Note that we do not provide company data for the Netherlands. Coverage expands market by market; the selection tool and the export format stay the same as it does.

Start with the segment, not the channel. Write your ideal customer profile as concrete filters — country, sector, size, structure, technical attributes — and count the result. That count tells you immediately whether inclusion targeting on ChatGPT Ads is even arithmetically available to you, or whether the channel belongs in the exclusion and bid-shaping column of your plan.

Then run the channels that are open to you on business data: postal, phone, email and, where accepted, messaging. Instrument everything so that every response, download, demo and purchase lands in your CRM with a lawful basis and a marketing permission attached. That record is what turns into an uploadable audience over the following quarters.

When your own first-party list crosses a size where a 25,000 match is plausible, start with exclusion — suppress existing customers from acquisition campaigns and you improve efficiency on day one, with no minimum to clear. Move to inclusion and bid multipliers only when the matched count supports it, and keep a control group so you can tell whether the channel actually added anything.

Finally, treat every upload as a decision with a paper trail: what the list is, where the permission came from, which audience object it created, when it was refreshed and when it was archived. The platform rules will keep moving. Documentation is what lets you adapt to the next rewrite without starting over.