Lookalike Audiences vs Interest Targeting (Real Test)

You have likely heard the common advice that mirroring your best customers is the only way to scale a paid social account. It is a frequent misconception in our industry that algorithmic similarity models always outperform broad interest buckets. In my ten years of managing multi-million dollar budgets, I have found that the “more precise” option often carries hidden costs that can erode your total return on investment.

Early in my career, I managed a large e-commerce account for a premium luggage brand. We were convinced that a 1% similarity model based on our highest-value purchasers would be our “silver bullet.” However, after three weeks of side-by-side testing against a simple group of people interested in “luxury travel,” the interest-based group had a 15% lower cost per acquisition. This experience taught me that the platform’s algorithm sometimes needs the freedom of a broader net to find the best buyers at the best price.

Understanding the Mechanics of Audience Selection

Choosing between manual interest selection and algorithmic modeling based on existing customer data is the most critical decision in any paid social campaign. This choice determines how the platform’s engine searches for your next customer among billions of active users. It sets the boundaries for the machine learning process.

To understand these concepts, we must first look at what they actually represent. Interest-based selection involves choosing specific categories, such as “Outdoor Photography” or “Personal Finance,” which platforms identify through user behavior and profile data. It is a proactive way to tell the platform, “Find people who care about these specific things.”

On the other hand, similarity modeling (often called lookalikes) uses a “seed” list of your own data—like a CSV of past buyers or website visitors—and asks the platform to find new people who “look” like them. The platform analyzes thousands of data points to find patterns that a human might miss. While one feels manual and the other feels automated, both require a strategic touch to function effectively.

• Interest Selection: Based on declared actions, page likes, and content engagement.
• Similarity Modeling: Based on deep-layer algorithmic patterns from a provided data source.
• Hybrid Approaches: Using interest layers to refine a large similarity model.

Why Platform Algorithms Influence Your Targeting Choice

The effectiveness of your targeting depends heavily on which platform you are using and how its specific engine processes user data. Each social network has a different “intelligence level” and data set, meaning a strategy that wins on Meta might fail on LinkedIn or TikTok.

Through my longitudinal platform tracking, I have observed that Meta’s algorithm is currently the most mature. It can often take a very broad interest group and optimize it into a high-performing audience faster than X or LinkedIn. TikTok, by contrast, relies heavily on immediate content resonance; if your creative doesn’t hook the user, even the most perfect similarity model won’t save the campaign.

Interestingly, LinkedIn’s professional data is so high-quality that interest-based targeting (like Job Seniority or Member Groups) often outperforms its similarity models for B2B lead generation. This is because professional identity is often more stable than consumer browsing habits. As a brand manager, you must match your targeting methodology to the platform’s inherent strengths.

Comparing Performance Metrics Across Audience Types

When evaluating which targeting method delivers the strongest return, we must look beyond just the click-through rate. A high click rate is meaningless if those users do not convert or if the cost to reach them is prohibitively expensive. We need to analyze the full funnel.

In a recent cross-platform marketing test for a SaaS client, I compared a 1% similarity model against a set of “Competitor Interests.” The results were surprising. While the similarity model had a higher average video watch time, the interest-based group had a much lower Cost Per Click (CPC). Because the interest group was cheaper to reach, it ultimately generated more leads for the same budget, even with a slightly lower conversion rate.

This highlights the “Crowding Effect.” Because many advertisers rush to use similarity models, the auction for those specific users becomes highly competitive. This drives up the Cost Per Mille (CPM). Sometimes, the “less perfect” interest group is so much cheaper to reach that it becomes the more profitable choice for your business.

• CPM (Cost Per Mille): Often higher for similarity models due to increased competition.
• CTR (Click-Through Rate): Usually higher for similarity models as the audience is “pre-qualified” by the algorithm.
• CPA (Cost Per Acquisition): The ultimate tie-breaker that determines which method stays in the budget.

Formulating a Real Placement Blueprint

A successful budget allocation requires a structured approach to testing these two methodologies side-by-side. You cannot simply guess which will work; you must build a framework that allows the data to speak for itself while protecting your overall ROI.

I recommend a “70/30” budget split when starting a new campaign. Allocate 70% of your spend to the targeting method that has historically performed for your brand, and 30% to the challenger. If you are starting from scratch, I suggest putting 70% into interest-based groups. This is because similarity models require a “learning phase” that can be expensive if you don’t have enough initial data to feed the engine.

Building on this, you should avoid “targeting overlap.” This happens when your interest groups and similarity models contain the same people. If you don’t exclude one from the other, you end up bidding against yourself in the auction. This is a common rookie mistake that inflates costs and makes it impossible to tell which audience is actually driving the results.

Step-by-Step Testing Sequence

1. Establish a Baseline: Run a broad interest-based campaign for 14 days to collect data.
2. Generate the Seed: Use the purchasers from that initial period to create a 1% similarity model.
3. The Head-to-Head Test: Run both audiences in separate ad sets with the same creative and budget.
4. Analyze the Divergence: Look for where the CPA starts to stabilize.
5. Reallocate: Shift budget to the winner, but keep the loser running at a low “maintenance” spend to monitor for shifts.

Asset Customization for Different Audience Behaviors

The content you show to an interest-based audience should often differ from what you show to a similarity-modeled group. These two segments are at different stages of the “awareness” journey, and your creative must reflect that to maximize performance.

Users in an interest-based group may not know your brand at all, but they care about the topic. Your creative here should be educational or problem-focused. For example, if you are targeting “Home Gardening,” your ad should solve a gardening problem. You are earning their attention through the shared interest.

Conversely, a similarity model is finding people who behave like your customers. They might respond better to social proof, testimonials, or direct “buy now” offers. They already “look” like buyers, so you can often be more aggressive with your sales messaging. I have seen countless campaigns fail because the manager used a generic “one-size-fits-all” ad for both groups.

• Interest Creative: High-value information, “How-to” content, and brand introductions.
• Similarity Creative: Direct offers, customer reviews, and “People Also Bought” style imagery.
• Retention Signals: Watch how long each group stays on your site to judge audience “quality” beyond the initial click.

Troubleshooting Metric Discrepancies and Algorithm Shifts

One of the biggest pain points for marketing managers is when a once-profitable audience suddenly stops performing. This “ad fatigue” or “audience saturation” happens at different rates for interest groups versus similarity models.

In my experience, similarity models tend to fatigue faster. Because the audience is often smaller (e.g., a 1% lookalike in a small country), the frequency—how often a person sees your ad—rises quickly. When people see the same ad too many times, they stop clicking, and your costs skyrocket. Interest groups are usually much larger, providing a “buffer” that allows the campaign to run longer before needing a creative refresh.

If you see your CPA rising, check your frequency first. If the frequency is above 3.0 for a similarity model, it is time to either expand to a 3% or 5% model or switch back to interest targeting. This constant “seesaw” between the two methods is what keeps a multi-channel portfolio healthy over the long term.

1. Check Frequency: Is the audience seeing the ad too often?
2. Verify Seed Quality: Is your similarity model based on recent buyers or “junk” traffic?
3. Review Platform Updates: Did the platform change how it defines a specific interest?
4. Analyze Placement-Level Data: Is one targeting method performing better on Stories than on the Feed?

Calculating Holistic ROI Across Networks

To justify your budget to a board or client, you must be able to explain the “Why” behind the “What.” You need a unified reporting structure that compares these targeting methods across all platforms in a way that makes sense for the business’s bottom line.

I use a “Channel Performance Report Card” that tracks the organic-to-paid engagement ratio and the platform-native retention signals. This allows me to see if a specific interest group on TikTok is bringing in “high-quality” traffic compared to a similarity model on Facebook. Sometimes, a higher CPA is acceptable if the “Lifetime Value” of that customer is significantly higher.

For instance, I once managed a campaign for a high-end furniture brand. The interest-based targeting on Pinterest had a 40% higher CPA than the similarity model on Meta. However, the Pinterest customers spent three times as much on their first order. By looking at the holistic ROI rather than just the initial acquisition cost, we were able to justify keeping the “more expensive” channel active.

Platform Evaluation Checklist

• [ ] Does the platform have enough data to support a similarity model? (Minimum 1,000 events recommended).
• [ ] Are the interest categories specific enough to avoid “junk” traffic?
• [ ] Have you excluded past purchasers from your “top-of-funnel” interest groups?
• [ ] Is the budget split at least 70/30 to allow for meaningful data collection?
• [ ] Are you using platform-native tracking to verify conversion data?

Practical Next Steps for Budget Allocation

Managing a diversified portfolio is about balance, not perfection. You will never find a “perfect” audience that lasts forever. The goal is to create a system where you are constantly testing the “known” (Interests) against the “predicted” (Similarity Models).

Start by auditing your current campaigns. Identify which ad sets are using which methodology and look for the “CPM gap.” If your similarity models have a CPM that is 50% higher than your interest groups, ask yourself if the conversion rate is high enough to justify that premium. If not, it is time to rebalance.

Finally, remember that the “seed” data is the heart of any similarity model. If your website tracking is broken or if you are feeding the algorithm “low-intent” data (like people who just visited your homepage), your similarity model will be weak. Focus on high-intent seeds—like “Added to Cart” or “Completed Purchase”—to give the algorithm the best chance of success.

Frequently Asked Questions

Which targeting method is better for a brand-new product launch?

Interest-based targeting is generally better for new launches. Since you likely do not have a large list of existing customers to use as a “seed,” similarity models will lack the data they need to be accurate. Starting with specific interests allows you to find your initial core audience and collect the data necessary to build similarity models later.

How large should my “seed” list be for a similarity model to work?

While most platforms say you only need 100 people, my experience shows that you need at least 1,000 high-quality events (like purchases) within the last 30 days for the algorithm to truly understand the patterns. Using a seed list that is too small often leads to erratic performance and high costs.

Why is my interest-based audience performing better than my 1% lookalike?

This often happens because of “Auction Competition.” Because similarity models are a popular “best practice,” many of your competitors are bidding for that same narrow group of people. This drives up the cost to reach them. Interest groups are often larger and less “crowded,” allowing you to get more impressions for the same dollar.

Can I use both targeting methods in the same campaign?

Yes, but you must be careful about overlap. It is best to run them in separate ad sets so you can track their individual performance. I recommend excluding your similarity model audience from your interest-based ad set to ensure you are reaching unique people in each group.

Does “Interest Targeting” still work after major platform updates?

Yes, but it has changed. Platforms now use “Interest Expansion” or “Advantage+” features that allow them to look beyond your chosen interests if they think they can find a conversion. This makes interest targeting more like a “hint” to the algorithm rather than a strict boundary.

How often should I refresh the “seed” data for my similarity models?

You should ideally use a dynamic seed that updates automatically via a platform pixel or API. If you are uploading manual CSV lists, I recommend refreshing them at least once a month to ensure the algorithm is looking for people who resemble your current customers, not your customers from a year ago.

What is the ideal budget split between these two methods?

For a stable account, a 60/40 split is often effective, with 60% going to the top-performing methodology (usually similarity models for established brands) and 40% going to interest-based groups to “prospect” for new segments and prevent audience fatigue.

Is TikTok’s similarity modeling as good as Meta’s?

Not yet. Meta has decades of historical data on user purchasing habits. TikTok’s algorithm is incredibly good at predicting what content you will watch, but it is still catching up on predicting what you will buy. On TikTok, I often find that broad interest targeting or “hashtag” targeting can match or beat similarity models.

Should I use “Broad” targeting with no interests or similarity models?

“Broad” targeting—where you only select age, gender, and location—is becoming more popular as algorithms get smarter. However, I only recommend this for brands with very high daily spends (over $500/day) and a lot of historical conversion data. The algorithm needs a massive amount of “signal” to work effectively with no guardrails.

How do I explain a higher CPA in interest groups to my board?

Focus on the “Reach” and “New Customer Acquisition” metrics. Interest groups often reach people who would never be found by a similarity model. Explain that these groups are essential for “filling the top of the funnel” and ensuring the brand doesn’t run out of new people to talk to as the similarity models become saturated.

(This article was written by one of our staff writers, Jonathan Mercer. Visit our Meet the Team page to learn more about the author and their expertise.)