Key takeaways
- An audience signal is not targeting. It describes who is likely to convert so the platform can expand beyond your list, rather than limiting who sees the ad.
- Google Performance Max, Meta Advantage+, TikTok, LinkedIn, and Pinterest all run the same seeding architecture under different names.
- A bad seed audience does more damage than a bad target list, because the platform goes and finds more people like the ones you described.
- Most advertisers seed with a converter list, which is backward-looking, usually too small to model, and rarely refreshed.
- A good seed audience is in-market, representative rather than exceptional, comfortably above the platform minimum after matching, and updated on a schedule.
- The two ways to build one are resolving anonymous site traffic into identifiable records and licensing intent data, third-party or second-party data.
- Seed quality is now the main lever in paid media, because targeting control is not coming back.
Every major ad platform has spent the last four years taking targeting controls away and replacing them with automation.
Performance Max in Google Ads removed keyword and placement control. Advantage+ in Meta Ads removed most manual audience selection. TikTok, Pinterest, and LinkedIn have all shipped their own versions of the same idea.
Most advertisers read this as a loss of control. It is more accurate to say the control moved. You no longer choose who sees the ad. You tell the model who is worth finding, and the model expands from there.
That input has a name on each platform. Google calls it an audience signal. Meta calls it a suggestion. The mechanics are close enough that they can be treated as one discipline, and almost nobody manages it as one.
This post covers how audience seeding actually works, why most seed audiences are bad, and how to build better ones across search, social, programmatic, and offline channels.
What is an audience signal?
An audience signal is a list of people you give an ad platform to describe who is likely to convert. Unlike targeting, it does not limit who sees the ad. The platform treats it as a starting point and expands outward to find people with similar or stronger intent.
Think of it almost like a lookalike audience that is generated in real time. In the case of Performance Max and Advantage+, they refer to this as an audience suggestion.
What actually changed
Google is explicit in its own documentation that Performance Max does not support restrictive targeting. What it supports is signals, and the purpose of a signal is to tell the campaign who is most likely to convert so the system can look for people with similar or stronger intent across Search, Display, YouTube, and the rest of the network.
Meta works the same way. When you attach a custom audience to an Advantage+ sales campaign, the default behavior is to use it as a suggestion. A suggestion does not restrict delivery. It describes your core audience while leaving the system free to expand past it. You have to explicitly switch the setup and turn the suggestion toggle off to force delivery to stay inside the list.
The practical consequence: an audience you upload is no longer a boundary. It is training data.
If the seed describes the wrong people, the platform does not show your ad to a smaller wrong audience. It goes and finds more of them.
This is why the old instinct to upload a customer list and move on produces such inconsistent results. Under a targeting model, a mediocre list wastes some impressions. Under a seeding model, a mediocre list defines the direction of expansion for the entire campaign.
Audience signals vs. targeting, side by side
| Attribute | Traditional targeting | Audience signals |
|---|---|---|
| What the list does | Defines who is eligible to see the ad | Describes who the model should look for |
| Effect on reach | Caps reach at the list size | Reach expands beyond the list |
| If the list is small | Low delivery, high frequency | Weak or unstable modeling |
| If the list is wrong | Wasted spend on that list | Wasted spend on that list and everyone like it |
| Best input | A precise, narrow segment | A representative, forward-looking segment at scale |
| Main lever | Segment selection | Seed quality and refresh rate |
The last row is the one that matters. Under targeting, the skill was choosing the right slice. Under audience signal seeding, the skill is sourcing a good input and keeping it current.
Why the default seed audience is usually a bad one
When a platform asks for a seed audience, most accounts hand over the same thing: a list of past purchasers or converted leads, exported once and uploaded.
That list can have three structural problems:
It is backward-looking
A purchaser list describes people who finished buying. The campaign’s job is to find people who are about to start. Those two groups overlap on demographics and interests, which is why the approach works at all, but they differ on the thing that actually predicts a conversion, which is whether someone is in market right now.
Seeding an audience on completed purchases teaches the model to find people who resemble your customers rather than people who are currently shopping. This can work, but doesn’t always work at scale.
It is usually too small
Ad platforms enforce minimum matched list sizes before a segment becomes usable, and the practical threshold for stable modeling is well above the published minimum. Plenty of advertisers seed with a few hundred matched records and then conclude that audience signals do not work.
Verify the current minimums for each platform before you build, since they change, but assume that a few thousand matched records is the floor for a seed that behaves predictably rather than the ceiling.
It goes stale
A list uploaded in January and left alone is describing a market that has moved. Intent decays quickly. A seed that is not on a refresh cadence quietly becomes a historical artifact while the campaign keeps optimizing against it.
It doesn’t match as well on B2B email data
Most ad platforms want personal email addresses, as the match rates are typically much higher. Think about your own behavior: when was the last time you signed up for a platform using your business email address?
For B2B paid media campaigns, uploading an audience signal full of business emails may not produce a high match rate, leading to lower counts and improper signals.
What a good seed audience looks like
The best seed audiences to produce the highest signal typically have the following in common:
- Forward-looking: It describes people showing interest now, not people who transacted in the past.
- Large enough to model: It clears the platform minimum by a comfortable margin after match rate is applied, not before.
- Representative, not exceptional: Your top 50 accounts are a bad seed. They are outliers, and the model will chase outliers.
- Refreshed on a schedule: Weekly or monthly, matched to how fast buying cycles move in your category.
- Portable: The same segment should be usable across every channel you buy, so you are seeding one definition of the customer rather than five.
- Consented and documented: Match quality is irrelevant if the segment fails a platform policy review.
Lookalike audiences are seeds too
A lookalike audience, or an actalike on Pinterest, is nothing more than a seed plus expansion. The platform has no information about your business other than the source list you handed it, so every criterion above applies with more force, not less.
If you are asking what to use as a seed audience for a lookalike, the answer is the same as for any audience signal: a representative in-market segment large enough to model, refreshed on a schedule, rather than your best customers or a stale converter export.
That last point on portability is where most programs leak value. A seed audience built inside Meta cannot be reused in Google. A Google Customer Match list cannot be pushed to a DSP.
If your segment definition lives in a platform, you are rebuilding it every time you add a channel, and the definitions drift apart. We’ll touch on this more later.
Channel by channel: what to seed and how
The naming differs, the concept does not. What varies is how much the platform lets the model expand and how much control you retain.
| Channel | What the seed is called | How it behaves |
|---|---|---|
Google Performance Max
|
Audience signal on the asset group | Guides the model, never restricts. Pair with search themes. |
|
Google Search & Demand Gen
|
Customer Match list, data segment | Can run as observation or targeting. Observation first, then bid adjust. |
Meta Advantage+
|
Custom audience used as a suggestion | Expands past the list by default. Toggle off to force retargeting. |
TikTok Ads
|
Custom audience, lookalike source | Seeds both lookalikes and automated campaign types. |
LinkedIn Ads
|
Matched audience, contact or company list | Company-level seeding is unusually strong for B2B. |
Pinterest Ads
|
Customer list audience, actalike source | Actalike expansion behaves like a lookalike. |
| Programmatic DSPs | Onboarded audience segment | Closest thing left to real targeting. Less model mediation. |
| CTV | Household-matched segment | Addressable at the household level rather than the individual. |
| Direct mail | Matched postal file | Not model-mediated at all. The list is the campaign. |
Two things fall out of that table.
First, the walled gardens are the most automated and therefore the most sensitive to audience seed quality. Programmatic and offline channels still let you buy the segment directly. If you have a segment worth trusting, it is worth running in both places, because the DSP and direct mail versions give you a cleaner read on whether the segment itself works, separate from the platform’s modeling.
Second, match rate compounds across the stack. Plan seed size against the matched number rather than the raw one, and ask any data provider for match rate per destination rather than a single blended figure. We cover how matching works in more detail in our guide to audience targeting.
Where seed data actually comes from
There are three sources, and most programs only use the first.
Your CRM and First-Party Data
Utilizing your first-party data is the perfect place to start, especially if you’re an established brand or company. It’s free, accurate, but could be too small and too backward-looking to carry a program on its own. Worth using as one signal among several, not as the only seed audience in a campaign mix.
Your own website traffic
Most sites can only build audiences from the small share of visitors who convert or log in. Website visitor identification closes some of that gap by resolving anonymous traffic to people you can push into an audience, which turns high-intent traffic that would otherwise be invisible into a usable seed.
This is the highest-intent first-party source available to most advertisers and it is almost always underused.
Third-party and second-party intent data
Buyer intent data providers in this category include Untitled, Bombora, G2, and 6sense, and they differ mainly in how intent is observed and at what granularity. This is what solves the scale problem. Intent segments built on content consumption cover people who are researching a category right now but have not touched your site yet, which is precisely the population a prospecting campaign is supposed to find.
Used as a seed rather than as a target list, good intent data gives the model a forward-looking definition of the buyer at a volume no CRM export can reach.
How Untitled fits in
Untitled supplies both of the sources above. The Untitled ID Tag resolves anonymous site visitors against a person-level identity graph, and the resulting audience can be pushed into ad platforms or integrated into your own CDP environment through our public API.
Untitled’s intent data is topic-classified, built on content consumption across more than 300,000 publisher websites, with counts refreshed weekly.
Segments push as Custom Audiences to Meta, Google, TikTok, and LinkedIn at roughly a 75% match rate, and the same segment can run through Untitled’s own CTV DSP or be mailed as a postal file.
Two honest limits worth stating: the intent data is topic-level rather than query-level, so it tells you someone is researching a category and not which exact search they ran, and the 75% figure is an average that varies by destination.
How to tell whether your audience signals are working
Knowing whether your audience signals are working can be hard to measure. The platforms will not tell you which conversions came from expansion versus the seed audience itself.
Three approaches get you close:
- Run the experiment: Performance Max supports campaign experiments. Run the same campaign with two different audience signals and let it settle. This is the only clean read available inside the walled gardens, and most accounts never use it.
- Watch new customer share, not blended ROAS: A good prospecting seed should raise the proportion of conversions coming from people who have never bought before. Blended ROAS will often look flat while the customer mix improves underneath it.
- Test the segment outside the model first: Run the segment through a DSP or automated direct mail where you are buying the audience rather than seeding a model. If it performs there, the segment is sound and any weak result inside Meta or Google is a modeling or creative problem, not a data problem. That separation saves a lot of arguing.
Common mistakes
Most of these are habits carried over from the targeting era, when they cost very little. Uploading a list once and leaving it alone was fine when that list only defined who was eligible to see the ad.
Under a seeding model the same habits compound, because the platform treats whatever you hand it as a description of your ideal buyer and spends the rest of the campaign looking for more people who match it. The six below account for most of the cases where an audience signal underperforms.
- Seeding with only your best customers: Outliers make bad training data. Use a representative segment.
- Uploading once: A static seed decays. Put it on a refresh schedule.
- Reading a suggestion as a target: If you needed the campaign to stay inside the list, you had to say so explicitly. Check the toggle.
- Ignoring consent settings: Uploaded data needs to be labeled correctly at the account level or it will not be usable.
- Sizing on raw records: Plan against the matched count after the platform applies its own identity resolution.
- Running one seed for every campaign: Prospecting and retention want different seeds. Feeding both the same list collapses the distinction the model is supposed to learn.
Frequently asked questions
Is an audience signal the same as targeting?
No. Targeting decides who is eligible to see your ad. An audience signal tells the platform who is likely to convert so it can find similar or higher-intent people beyond your list. Delivery is not limited to the people you upload.
Do Performance Max audience signals actually work?
They influence how the campaign expands, but they are an input rather than a control. Accounts that report no effect are usually seeding with a list that is too small to model or too backward-looking to be useful. Test it with a campaign experiment rather than judging on a single before-and-after.
What is the best seed audience for Performance Max?
A representative, in-market segment that clears the platform minimum after matching and is refreshed regularly. In practice that means combining recent high-intent site traffic with a third-party intent segment, rather than relying on a converter list alone.
My custom audience is too small to use. What should I do?
You have three options: widen the definition, resolve more of your anonymous site traffic into identifiable records, or supplement with third-party intent data. Widening the definition is free but dilutes the seed. The other two add scale without weakening what the seed describes.
Can I use third-party intent data as a seed audience?
Yes, provided it is delivered in a format the platform accepts and the data is consented and labeled correctly at the account level. Intent segments are often a better seed than a customer list because they describe current research behavior rather than completed purchases.
Does a bad audience signal hurt performance?
It can. Because the signal directs expansion, a seed that describes the wrong people pushes the model toward more of the wrong people. This is the main way seeding differs from targeting, where a bad list simply wastes its own impressions.
How often should a seed audience be refreshed?
Match the cadence to your buying cycle. Weekly is appropriate for fast-moving consumer categories and for intent segments, where the value of the data is that it reflects current behavior. Monthly is usually sufficient for longer B2B cycles.
How do I build a custom audience without a customer list?
Resolve your own anonymous site traffic into identifiable records, or license a third-party intent segment. Visitor identification turns high-intent traffic you already have into an audience, and intent data supplies scale for people who have not reached your site yet. Both produce a list you can upload the same way you would a customer file.
What are the alternatives to Google Customer Match?
Customer Match is a feature rather than a vendor, so the real question is usually where to get a list worth uploading. If your own data is too small or too stale, the alternatives are visitor identification on your site traffic and licensed intent segments. Both feed the same Customer Match upload rather than replacing it.
Where can I buy intent data for Google Ads?
Several providers license intent segments that can be uploaded to Google Ads as customer lists or data segments, including Untitled, Bombora, G2, and 6sense. The things to compare are how intent is observed, the granularity of the classification, refresh frequency, and the match rate into Google specifically rather than a blended average.
How do I get better first-party data for Meta ads?
Most sites can only build audiences from the small share of visitors who convert or log in. Resolving anonymous traffic into identifiable records closes that gap and gives you a larger, higher-intent first-party audience to push to Meta. Server-side conversion tracking improves the signal Meta receives, but it does not increase the size of the audience you can seed with.
What should I use as a seed audience for a lookalike?
A representative in-market segment rather than your highest-value customers. Lookalikes amplify whatever the source describes, so an outlier list produces an outlier audience. Size the source comfortably above the platform minimum after matching, and refresh it on the same cadence as any other seed.
TL;DR: the short version
You have less control over targeting than you did five years ago and you are not getting it back. What you do control is the quality of the input every one of these systems is now built to consume. Treat the seed audience as a managed asset rather than a one-time export, build one segment definition that travels across Google, Meta, TikTok, LinkedIn, programmatic, CTV, and mail, and refresh it on a schedule that matches how fast your buyers move.
That is the whole discipline. It is unglamorous, and it is currently the highest-leverage thing most paid media programs are not doing.
About Untitled
A seed audience worth feeding the model
Untitled is a target audience platform built on identity resolution. It exists to solve the problem this post describes, which is that most advertisers are seeding automated campaigns with a list that is too small, too old, or too backward-looking to model well.
There are two ways to fix that, and Untitled supplies both:
- The Untitled ID Tag resolves anonymous visitors on your own site against a person-level identity graph, turning high-intent traffic that would otherwise stay invisible into an audience you can actually use.
- Untitled Intent Data covers the people researching your category who have not reached your site yet. Topics are classified from content consumption across more than 300,000 publisher websites, with counts refreshed weekly.
Segments built either way push as Custom Audiences to the major platforms via our integrations at roughly a 75% match rate, and the same segment can run through Untitled’s CTV DSP or be mailed as a postal file. Audiences can also be pushed into your own warehouse or CRM through the Untitled API.
If you’re interested in learning more, explore our case studies, and when you’re ready, book a demo today and start seeding your audiences for better paid media results.
Check Out Other Articles
- Bot Traffic Is Wasting Your Ad Spend: How to Target Real Humans Instead
- What Is Audience Intelligence? Definition, Use Cases, and Tools
- Person-Level Website Visitor Identification: How to See the Real People Behind Anonymous Traffic
- The Future of Website Visitor Identification: 7 Trends Reshaping How Brands Identify Anonymous Traffic in 2026
- Customer Data Platform Alternatives: Why Marketing Teams Are Moving to a Composable Stack
- How Identity Resolution Unlocks Customer Insights
