What are lookalike audiences, and how do site lookalikes differ?
Part of Lookalike websites
The short answer
A lookalike audience is a group of people an ad platform predicts resemble a seed list of the advertiser's own customers or visitors. A site lookalike is different: it is a list of websites whose articles resemble a few seed sites. The first models people from personal data; the second compares published content and needs no data about readers.
Key takeaways
- A lookalike audience starts from a seed list of real people, such as customers, site visitors or app users, and the platform models others who resemble them.
- A site lookalike starts from one to three publishers the client trusts and finds other publishers whose articles cover the same subjects.
- Uploading a customer list, even hashed, means handing personal information to a platform, which brings consent and privacy duties in the US and Canada.
- Removing sensitive fields from a model's inputs does not guarantee unbiased outputs, according to peer reviewed research on lookalike tools.
- A site lookalike suits briefs with thin first-party data, strict privacy limits or a client who can name the publications it wants to appear beside.
What is a lookalike audience?
Definition
- Lookalike audience
- A group of people that an ad platform selects because its model predicts they share characteristics with a seed list the advertiser supplied. The advertiser never sees who is in the group; it sees an estimated size and buys against it.
In Google Ads the current form is called Lookalike segments (opens in a new tab), which Google describes as "groups of people that share characteristics with others on an existing seed list." The seed is built from the advertiser's first-party data. The platform then looks across its own signed in users for people whose behavior resembles the seed and makes them reachable as a targeting option.
The word lookalike is doing two jobs in media planning, which can confuse a brief. When a client says "find more people like our customers," they mean an audience model. When they say "find more sites like the three we already love," they mean a list of publishers. The tools, the data and the legal questions behind those two requests are different, so the planner should ask which one the client means before anything is built. The lookalike websites hub covers the site side in full.
How is a lookalike audience built from a seed list?
On Google's Lookalike segments the mechanics run in stages. The advertiser assembles a seed, the platform matches the seed to its own users, a model learns what those matched users have in common, and the platform scores everyone else for resemblance. The advertiser chooses how close the resemblance must be, trading similarity for reach.
Five stages from a customer file to a modeled audience
- Assemble the seed. Google lists website visitors, app users, customer lists, Google Analytics audiences and YouTube audiences as seed types for Lookalike segments.
- Match the seed to platform users. For a customer list, Google's Customer Match documentation (opens in a new tab) says its system turns contact details into codes "using the secure hashing algorithm SHA256" and compares them with the hashed details of Google accounts.
- Check the seed size. Google requires the submitted seed lists to add up to more than 100 active matched people before a segment can be built.
- Choose the reach level. Google offers narrow, balanced and broad settings, each aiming at a larger share of users in the target location and a looser resemblance.
- Let the model refresh. Google says a segment is refreshed every one to two days from the available customer data, so the people inside it change while the campaign runs.
Two consequences follow for planners. First, the quality of the audience depends on the quality of the seed: a seed built from bargain hunters during a clearance sale can produce more bargain hunters. Second, the advertiser cannot audit the model. It can see delivery reports by placement and by demographic bucket, but it cannot see why any given person was included. The similar audiences replacement guide covers how these options fit inside Google Ads today.
How do lookalike websites differ from lookalike audiences?
A site lookalike answers a different question. Instead of asking which people resemble the customers, it asks which publications resemble the places where the brand already likes to appear. The planner gives one to three seed sites, the search compares the meaning of their articles with the articles on candidate sites, and the output is a list of domains a person can open, read and approve.
| Question | Lookalike audience | Site lookalike |
|---|---|---|
| What is the seed? | People: a customer list, site visitors, app users or video viewers | Publications: one to three websites or sections the client trusts |
| What is compared? | Behavior and attributes the platform holds about its users | The subjects and depth of published articles |
| What does the buyer receive? | A targeting option with an estimated size and no member list | A named list of domains, each with the articles that justify it |
| Where can it run? | Only inside the platform that built the model | Any buying system that accepts a domain list, or as a deal on an exchange |
| Can the client audit it? | Through delivery reports only; the model is not visible | Yes, by opening each site and its matching pages before launch |
| Main weakness | Opaque, tied to one platform and dependent on seed quality | Reaches readers of a context, not proven buyers, and needs quality checks on every site |
Neither approach is a substitute for the other in every case. A modeled audience can find likely buyers across many contexts, including ones a planner had not considered. A site lookalike gives up that reach in exchange for control: the brand knows where its ads will appear and why. A plan can use both, with the site list carried into an open web or deal buy and the audience model kept inside the platform that built it.
What personal data does each approach use?
Privacy teams can ask this early, and the answer differs sharply between the two. A lookalike audience is personal data processing from start to finish. The seed is a list of identifiable people, the matching links them to platform accounts, and the model works on what the platform knows about its users. A site lookalike reads public articles and compares them with other public articles. It needs no list of customers and nothing about any reader.
In Canada, the Office of the Privacy Commissioner (opens in a new tab) states that information used to target behavioral ads "will generally constitute personal information", that collection should be limited to non-sensitive information where practicable, and that express consent is the usual approach for sensitive information such as health or financial data. In California, the California Privacy Protection Agency (opens in a new tab) describes a consumer right to ask businesses to stop selling or sharing personal information for cross-context behavioral advertising.
Bias is the second concern. A 2022 AAAI/ACM paper by Sapiezynski and colleagues (opens in a new tab) compared a lookalike tool with a version that had sensitive features removed from its inputs and found that removing demographic features did not reliably prevent biased audiences by gender, age and race.
When does a site based lookalike make more sense?
Signs a brief is better served by a site lookalike
- The advertiser's customer list is too small or too old to make a reliable seed.
- The client cannot, or will not, upload customer data to an ad platform under its privacy commitments.
- The product touches a sensitive category, such as health or personal finance, where profiling raises consent questions.
- The client can name publications it trusts and wants more places like them.
- The buy runs through an exchange deal or a buying system other than the platform that would build the audience.
- The client or its regulator expects a list of where ads appeared that a person approved before launch.
A modeled audience still makes sense when the advertiser holds a large, consented customer file, buys mostly inside one platform and cares more about finding converters than about where the ad appears. The two can also be layered: a site lookalike defines the places, and a broad audience setting runs inside them. The contextual targeting guide explains how content based selection works without any data about the reader.
How does PlacementGraph build site lookalikes without personal data?
What does choosing between the two look like on a real brief?
Fictional example
A fictional boutique fitness studio chain in Alberta
An unnamed chain of indoor cycling and strength studios in Calgary and Edmonton asks its agency for "more people like our members" ahead of a January membership push. The planner asks what the seed would be. The member file is small, mixes current and lapsed members, and the client's privacy notice does not mention sharing member contact details with ad platforms.
The planner proposes two tracks. Inside the platform, the agency uses only the studios' own website visitor lists, which the client's notice already covers, as a seed for a modeled audience. For the open web, the client names three publications its instructors read: a strength training magazine, a running and recovery blog and a Calgary food and wellness site.
The site lookalike search runs one theme at a time, training and local wellness, and returns specialist publications with a named reason for each. The client's privacy lead approves the site list in one review because it contains no personal data, while the audience track waits for an update to the privacy notice.
Frequently asked questions
Is a lookalike audience the same as retargeting?
No. Retargeting shows ads to the same people who already visited a site or used an app. A lookalike audience uses those people only as a seed and reaches others the model predicts resemble them. Retargeting talks again to known visitors; a lookalike audience looks for new prospects, so the two can be planned and reported as separate lines.
Can a lookalike audience built on one platform be used on another?
Not directly. The model and its members live inside the platform that built them, and the advertiser never receives a list of people. Each platform needs its own seed and builds its own audience. A site lookalike, by contrast, is a plain domain list that can be loaded into any buying system that accepts one.
Does hashing a customer list before upload make it anonymous?
No. The FTC's Office of Technology wrote in July 2024 that hashes can still be used to identify people and that an opaque identifier is no excuse for improper use or disclosure.
How large must a seed list be for a lookalike audience?
It depends on the platform. For Lookalike segments, Google requires the submitted seed lists to total more than 100 active matched people, and the matched count can be lower than the number of rows uploaded. A site lookalike needs no list of people, only one to three seed publications.
Sources
- Use Lookalike segments in Demand Gen campaigns as a Targeting Constraint (opens in a new tab)Google Ads Help
- How Google uses Customer Match data (opens in a new tab)Google Ads Help
- No, hashing still doesn't make your data anonymous (opens in a new tab)FTC Office of Technology, July 24, 2024
- Guidelines on privacy and online behavioural advertising (opens in a new tab)Office of the Privacy Commissioner of Canada
- Frequently Asked Questions (opens in a new tab)California Privacy Protection Agency
- Algorithms that "Don't See Color": Comparing Biases in Lookalike and Special Ad Audiences (opens in a new tab)Sapiezynski, Ghosh, Kaplan, Rieke and Mislove, AAAI/ACM AIES 2022
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.