Some advertisers who are participating in the ChatGPT ads beta now have the ability to create, target, and exclude custom audiences based on lists of email addresses and phone numbers. I am one of those advertisers.
I admit this has become a repetitive theme, but I have to say from the start how incredibly basic this is. I’m not exaggerating by saying I’d compare this to when Facebook (before there was “Meta”) introduced custom audiences more than a decade ago.
But it’s something. And honestly, it does say something about OpenAI’s plans related to targeting control that remarketing is even in the initial plans.
I’m not saying this is good or bad. But in the spirit of looking at this from a completely different lens than I look at Meta ads, I’m open-minded to any feature OpenAI gives us until proven otherwise.
With all that said, there are signs that at least the initial launch of custom audiences will be inaccessible to the vast majority of advertisers. But hopefully that will change.
Let’s take a closer look…
Create an Audience
Under the Tools section on the left, you may see a new link for Audiences.
Before you create your first audience, you’ll see this message:
Upload raw or hashed emails and phone numbers to use as audience filters for campaigns.
Click “Create Audience,” and you’ll see this…
You’ll need to name the audience and select an identifier type.
The four data options you can upload:
- Phone
- Hashed Email (SHA-256)
- Hashed Phone (SHA-256)
You’ll need to upload a CSV or TXT file (up to 500 MB) of email addresses or phone numbers.
OpenAI will then match your list to OpenAI users. In my example above, I’m getting messages that there are no matched users and the audience is “too small.”
More on this in a moment.
Using Audiences: 100,000 Minimum?
In order to apply these audiences, OpenAI added a new section to campaign setup.
As you can see in the screenshot above, you can either target or exclude these audiences.
Unfortunately, audiences aren’t eligible to be used until they exit the “too small” designation. And how big does that audience need to be?
Well, according to a rare tooltip in the ChatGPT ads beta UI, the “audience must include at least 100,000 identifiers.”
Is that 100,000 identifiers before matching or 100,000 matched records? Either way, that’s quite the number, and it applies to both targeting and exclusions.
Help Article Documentation: NONE
After that, the details are difficult to find. In fact, I don’t think OpenAI has published anything yet.
I assume it’s rather straightforward that a raw upload of email addresses or phone numbers should only consist of a single column of that information. No first name, last name, or any other columns. There’s never a step to map columns to data types, so that would make sense.
Does it matter if you include a header row? I don’t know, but I’ve tried both. I’ve also tried both hashed and raw versions.
I’ll be curious to find out what, if anything, this “100,000 identifier” minimum means. Is it incorrect information added to a tooltip for a new product? Is it a temporary limitation that will be loosened? If this is a hard number, it could make the custom audience feature inaccessible to most advertisers.
OpenAI is matching a single data point to users, and a ChatGPT account wouldn’t be as ubiquitous as a Facebook account. And considering people may have used a different email address or phone number to create their OpenAI account in the first place, the match rate is going to be low. So if you need 100,000 matched addresses (and not just 100,000 total), that’s going to be a rare feat.
Is This a Big Deal?
The one thing I’ll say is that this is interesting since all initial indications were that OpenAI was going to lean heavily into a more algorithmic targeting approach. Until now, the only targeting input an advertiser could provide outside of country was Context Hints — a prompt of sorts to describe your audience.
Otherwise, no age or gender even. Just location.
So, does this mean that OpenAI will offer more targeting control? Or is this just a feature that only the biggest companies will get to experiment with?
Ultimately, whether remarketing matters in the first place will depend a whole lot on how good OpenAI’s algorithmic signals are. If they’re primitive, the way Facebook ads were a decade ago, options like remarketing will prove useful.
Time will tell.
Your Turn
Do you have this feature? What do you think?
Let me know in the comments below!
