Look-alike, segments, exceptions

Look-alike (LAL) is a technology in which the advertising platform algorithm finds users similar to your “original” audience. On big data VK Ads, Ya.Auditory and Telegram this gives cold traffic with a CR 1.5–3 times higher than that of classic targets based on interests. In 2026, LAL is the main tool scaling performance campaigns after hot demand is exhausted.

how look-alike works under the hood

The LAL algorithm is a binary classifier. The site takes your original audience (e.g. 5,000 customers), extracts behavioral attributes of each person (subscriptions, activity, demographics, devices), and trains the model to “distinguish the buyer from a random user." After training, the model is applied to the entire VK / Yandex database and gives a score from 0 to 1 for each user.

“Accuracy 1” - we take the top 1% of the scoring (the most similar). "Accuracy 2" - top 2-3%. “Accuracy 5” is a wide audience, closer to niche interests. What is the accuracy? the more expensive the click, but the higher the CR. The wider, the larger the volume for scaling.

initial base: what to load and in what volume

The quality of LAL is 70% determined by the quality of the source base. Top sources descending values:

1. Buyers in the last 90 days. The most valuable base. Minimum 300, better than 1,000+.
2. Repeat customers. Who bought 2+ times. Less volume, but signal density is maximum.
3. Paid plan subscribers. For SaaS and subscription models.
4. Applications that have become a deal. Through CRM upload.
5. Visitors to the shopping cart and payment page. Weaker than buyers, but the volume is larger.
6. All visitors for 30 days. The widest and weakest signal. We use as a “seed” when there is nothing else.

The minimum for building a LAL in VK is 300 people, in Y.Audiences - 1,000. Less - the algorithm does not have enough data, the model will not learn. Recommended volume for sustainable LAL - 3,000–10,000 people.

lal by customers vs by visitors

The fundamental difference is in signal density. LAL for 1,000 customers will give a narrow, but a hot audience (about 500,000–1,500,000 people with an accuracy of 2 - 1,000 “super similar” forms an extension of 500-1500 times larger). CR of this audience is 2-3 times higher than the basic target by interests.

LAL of 50,000 visitors will give a wide audience (5,000,000+ people), but “the soup will be less rich” - the model learns from the mixture “came in by accident” and “really interested.” CR is lower, but the volume for scaling is larger.

A working strategy is to combine both:

• LAL by customers (accuracy 1-2) - launch campaign, narrow
• LAL by added to cart (precision 2) - extension
• LAL by category visitors (accuracy 3) - for scaling

custom audiences: alternative and addition to lal

Custom audiences are segments that you upload directly (email, phone, device id, Yandex id). This is not LAL - it is targeting specific people. Use:

• Reactivation of old clients. We load the customer base in 6–12 months that have not purchased in the last 90 days.
• Audience of competitors. If you have databases (purchased or collected through subscriptions), we load them into VK Ads and target them directly.
• Segmentation by LTV segment. “VIP clients with LTV 50,000+” — separate campaign with premium creative.
• Hypertarget for B2B. Email databases of executives in a specific industry, upload to VK Ads.

exceptions: a key part of segmentation

In every campaign with a “fresh” audience, three types of exceptions are required:

1. Current clients. We do not show advertising to those who are already buying. This is not only ethics - it is also economics: you are overpaying for people who are already yours.
2. Employees. Geo-office, corporate email, friend-graph employees.
3. Already interacted in 30 days. To cold campaign and retargeting did not intersect - otherwise the frequency cap will explode.

typical lal construction mistakes

• The base is too small. 80 buyers is not LAL, it is random noise. Minimum 300, better 1,000+.
• Database of “all clients” without segmentation. If you have VIP (LTV 50k+) and one-time (LTV 1k), mixing them you will get an “average” LAL, not optimized for any group.
• Do not update the database. LAL “by customers 2023” in 2026 — useless. We update at least once a quarter.
• Too narrow accuracy right away. Accuracy 1 gives CR, but does not volume. To scale, you need an accuracy of 2-3.

testing segments: correct structure

Don't launch 12 segments at the same time with one creative. Correct structure:

1. We prepare 2-3 key segments (LAL for customers, LAL for baskets, interests).
2. For each segment - 2-3 creatives.
3. Budget for the segment - minimum CPA × 30 (enough for 30 conversions for evaluation).
4. After 7 days, we cut off the worst segment and transfer the budget to the leaders.
5. After 14 days, we repeat the iteration.

This is the basic optimization cycle. Read more about A/B tests in Chapter 9 and in blog article about A/B testing.

how to read lal quality

Not “cheap click = good LAL”. LAL quality is read by three metrics in combination:

• CR in primary conversion (lead/cart) - higher than the base target by 30%+
• CR from primary to purchase - close to the average for CRM (which means the leads are high-quality)
• LTV of the LAL cohort is equal to or higher than the LTV of the hot search cohort

If the first metric is good, but the second or third is not, then LAL is catching the easy ones. leads” who don’t buy. We narrow the accuracy or change the initial base.

in the next chapter

Chapter 9 - Creatives and A/B Tests. Why the best creative gets 8x more CTR than the worst, how to properly structure the test and what volume is needed for statistical significance.