Lookalike Audiences: How Segment Fragmentation Can Improve Programmatic Targeting
Lookalike audiences (LAL) sound almost magical: start with a group of valuable users, find people who resemble them, and expand your reach. 🪄 But the real challenge is not just finding more people. It is finding more people who are relevantly similar.
1. What is a lookalike audience in programmatic advertising?
A lookalike audience is a group of users selected because their characteristics or behaviors resemble those of a reference population, often called the seed audience. Your seed might consist of visitors to a product page, existing customers or users who have completed a valuable action.
From that seed, a technology provider looks for patterns associated with similar users and builds a broader audience that advertisers can activate. The ambition is straightforward: increase reach without losing too much targeting relevance.
Different platforms use different modeling techniques—some rely on machine learning, embeddings or proprietary features, while others make extensive use of segment-level affinities. There is no single universal lookalike algorithm.
2. The intuition behind affinity scoring
One accessible way to think about lookalike modeling is to compare how frequently a seed audience appears within a segment versus the general population:
For example, if 20% of a segment belongs to the seed audience, compared with 5% of the overall population, its affinity index is 4. That segment is four times as concentrated in seed users as the baseline population (assuming comparable, properly measured populations).
This is a useful illustration of relative concentration, not a guarantee that every member of the segment will convert. In practice, vendors may use more complex scores and additional signals.
3. Three limitations every Media Trader should understand
Scale versus precision. Expanding a lookalike too aggressively can dilute its similarity to the seed. Restricting it too much may produce an audience that cannot spend.
Seed quality. A model trained on a weak or misleading seed will reproduce those weaknesses. Product-page visitors, for example, are not automatically equivalent to purchasers.
Data granularity. A publisher or advertiser whose DMP contains only ten broad segments may struggle to distinguish valuable combinations of interests and behaviors. The original categories can be too coarse to identify the best pockets of affinity.
These constraints connect to a broader challenge we have discussed in audience fragmentation across the Open Web: useful data may exist, but its structure and activation paths matter enormously.
4. The fragmentation strategy: look at intersections, not just categories
Imagine that your DMP contains a broad segment labeled Sports Fans and another labeled Technology Enthusiasts. Both might show moderate affinity with your seed. But what about people who belong to both groups?
That intersection could show a much stronger relationship with your seed than either broad segment alone. The fragmentation approach therefore tests combinations or sub-segments instead of evaluating only the original categories.
Think of it as using a finer sieve: rather than treating each large audience bucket as homogeneous, you examine the smaller groups hidden inside it. 🪨⛏️
5. From ten segments to a richer set of candidates
Consider a publisher with ten audience segments. The original approach scores those ten segments individually. A fragmentation approach can also evaluate their intersections:
| Approach | Candidate groups | What it can reveal |
|---|---|---|
| Original segments | 10 broad segments | Which individual categories over-index on the seed |
| Pairwise intersections | Up to 45 distinct pairs | Whether combinations have stronger affinity than their components |
| Individual segments + pairs | Up to 55 candidates | A more granular shortlist for activation |
| Higher-order intersections and other splits | Potentially many more | Additional patterns, subject to data availability and statistical reliability |
In the original example, ten segments could produce around a hundred sub-segments when multiple combinations or splitting rules are considered. That is a useful illustration of the potential, not an automatic mathematical outcome: ten segments produce 45 unique unordered pairs, or 55 candidates if the ten originals are included.
Crucially, more candidates do not mean more unique users. Intersections overlap, and many will be too small to support reliable estimates.
6. How to put segment fragmentation into practice
The good news: a publisher does not necessarily have to create and store dozens of new audience objects manually. Depending on the DMP and its query capabilities, much of the exploration can happen through virtual calculations or SQL-based analysis.
- Choose a meaningful seed. Prioritize business outcomes and ensure the seed has enough volume.
- Inventory your available segments. Check their definitions, freshness, consent basis and coverage.
- Calculate segment intersections. Test combinations that are operationally meaningful rather than generating every possible permutation blindly.
- Score the candidates. Compare affinity against the baseline, while monitoring sample size and confidence.
- Select qualifying groups. Use a query or DMP rule to assemble a combined LAL audience from high-affinity candidates, deduplicating users where possible.
- Activate and test. Compare performance with the original broad-segment lookalike.
The precise SQL formula depends on the data model, identifier structure and privacy constraints of the platform. There is no single portable query that works across all DMPs.
7. Does a fragmented lookalike actually perform better?
A high affinity score is an interesting signal, but campaign performance is the real test. Compare the fragmented LAL against a conventional LAL using a consistent budget, inventory mix and measurement framework.
Depending on the objective, examine qualified reach, CPM including data fees, conversion rate, cost per incremental action or other funnel-aligned KPIs. Where possible, use holdouts or controlled experiments to test whether the new audience generates additional outcomes rather than merely claiming existing demand.
Watch for sparse groups, overfitting, overlapping segments and selection bias. An intersection that looks extraordinary on a tiny sample may collapse when activated at scale.
This is also why our articles on incremental targeting and attribution models emphasize separating a clever targeting mechanism from its measurable business impact.
8. What this means for programmatic curation
Here in France, publishers, retailers and advertisers have valuable audience signals that do not always translate into scalable programmatic activation. Better audience engineering—including carefully validated segment fragmentation—can help unlock that value.
Through Curator Seats, relevant data and inventory strategies can be packaged into curated deals for activation in buyers' DSPs, subject to platform integrations, permissions and available identifiers.
At GreyFox Ads, we believe that effective Open Web targeting is not about piling on more segments and fees. It is about finding the right signals, combining them intelligently and making activation simple and transparent for media traders entering the French market.
Make French Open Web targeting work harder
Need to activate relevant audiences and quality French publisher inventory without multiplying your operational complexity?
Discover our Master DealTalk to GreyFox AdsOriginal post: Read Max Legrand's LinkedIn post. The strategy and illustrative DMP example are drawn from the author's original commentary; the statistical caveats and combinatorial clarification are added for context.