What does website similarity mean, and what does each measure prove?
Part of Lookalike websites
The short answer
Website similarity is a claim that two sites share something: their topic, the meaning of their articles, their readers, the sites that link to them, or their owner. Each measure proves only its own kind of likeness. Topic and meaning say nothing about quality, and audience overlap is the one measure that needs data about people.
Key takeaways
- There are at least five distinct ways two websites can be similar, and a planner should say which one a claim of similarity rests on.
- Topical and meaning based similarity describe content; they do not show that two sites share readers or meet the same standards.
- Of the five measures, audience overlap is the one that depends on data about individual visitors, which brings privacy questions with it.
- Shared links and shared ownership can reveal networks of related domains, which matters more for list hygiene than for relevance.
- Two sites can match closely on every content measure and still differ completely in editorial quality, ad load and traffic sources.
What does it mean for two websites to be similar?
Definition
- Website similarity
- A measured relationship between two websites along one stated dimension, such as subject matter, the meaning of their articles, shared readers, shared inbound links or a shared owner. Without the dimension named, the word similar is not a checkable claim.
When a client says two publications are alike, they can mean several things at once: same subjects, same seriousness, similar readers. A measurement system cannot capture all of that in one number. Each method looks at one kind of evidence and is blind to the rest. A planner who reports that a site is similar to a seed should be able to say which evidence the claim rests on, the same way a researcher names the variable behind a correlation.
| Dimension | Evidence used | What it proves | What it misses |
|---|---|---|---|
| Topic | Content labels, such as categories from the IAB Tech Lab Content Taxonomy | Both sites are filed under the same subject | Depth, angle and whether most of the site is really about that subject |
| Meaning | Text of sampled articles turned into vectors and compared | The articles discuss the same things in similar ways | Who reads them, who owns them and whether the content is original |
| Audience overlap | Records of the same people visiting both sites | The sites share some readers | Why readers visit, and whether the overlap reflects interest or bought traffic |
| Link patterns | Which pages link to both sites, or which sites both link to | Other people on the web treat the sites as related | Recent changes in focus, and links bought or exchanged inside a network |
| Ownership | Shared analytics or ad account codes, ads.txt owner fields, registration data | One operator probably runs both sites | Anything about content fit; related sites can cover unrelated subjects |
How is topical similarity different from audience overlap?
Topical similarity and meaning based similarity look at what a site publishes. Two cooking sites that both carry the same content label are topically similar; two sites whose sampled articles both explain weeknight bread baking for beginners are similar in meaning, which is a tighter match. The website categorization guide covers how labels are assigned, and the semantic targeting guide covers how meaning is compared.
Audience overlap looks at who visits. It asks what share of one site's visitors also visit the other. That is useful for a different reason: two sites on unrelated subjects can share readers, such as a regional news site and a local sports blog, and two sites on the same subject can share almost none, such as a beginner gardening site and a professional horticulture journal. Neither measure substitutes for the other.
Why can two similar looking sites differ in quality?
Content measures reward resemblance, and resemblance can be cheap to produce. Google's spam policies name the patterns directly. Scaled content abuse is "when many pages are generated for the primary purpose of manipulating search rankings and not helping users", including scraping or rewriting other content. Expired domain abuse is where an expired domain is bought and repurposed "by hosting content that provides little to no value to users." A rewritten copy of a respected publication can sit very close to it on topic and meaning while sharing none of its standards.
Quality differences that similarity scores do not catch
- The candidate's articles carry named writers, dates and specific detail, like the seed's, rather than generic summaries.
- Distinctive sentences from the candidate do not appear word for word on other domains.
- The candidate's domain history fits its current subject, rather than a recent change from an unrelated former use.
- The share of each page given to advertising is comparable to the seed's, on a phone and on a desktop.
- The candidate's readers appear to arrive through search, direct visits or social sharing rather than paid link units.
- The candidate has published on the subject for long enough to show a track record.
These checks belong after the similarity search, never inside it. Folding quality into one similarity score hides the reason a site was kept or dropped. The AI generated content sites guide covers one source of lookalike copies.
Which similarity measures need personal data?
| Measure | Personal data needed | Why |
|---|---|---|
| Topic labels | None | Labels describe published pages |
| Meaning of articles | None | Only public article text is read and compared |
| Link patterns | None | Links are public parts of published pages |
| Ownership traces | None about readers | Codes and ads.txt fields describe the site; registration records can name a person and should be handled with care |
| Audience overlap | Yes | Someone must record that the same browser or person visited both sites |
| Country popularity (context only) | None at the buyer's end | Public aggregates such as the Chrome UX Report give popularity by country, inferred from IP addresses, without any individual records |
Audience overlap data comes from tracking people across sites, through panels, browser extensions, cookies or device identifiers. In Canada, the Office of the Privacy Commissioner (opens in a new tab) takes the position that tracking and targeting data used for behavioral advertising will generally constitute personal information. A planner buying an overlap report does not collect that data directly, but should know where it came from before relying on it. The Chrome UX Report (opens in a new tab) is a different kind of source: it publishes aggregated records by country and device, and the CrUX popularity rank guide explains what it shows.
How does PlacementGraph measure similarity between sites?
What does reading similarity evidence look like on a real brief?
Fictional example
A fictional small press in the US Southwest
An unnamed independent publisher of Southwest fiction and regional history, based in Albuquerque, wants display ads in Arizona and New Mexico beside literary reviews and author interviews. Its agency receives a vendor report listing 30 sites "similar" to a respected regional literary magazine.
The planner asks which dimension each entry rests on. Twelve entries come from audience overlap: several are local weather and traffic sites that share readers with the magazine but publish no books coverage. Eight come from meaning: book review blogs and a university press newsletter. Ten come from link patterns, and six of those link mainly to each other and share one analytics code.
The planner keeps the meaning based matches that pass quality checks, treats the six linked domains as one network entry that then fails the page review, and moves two local news sites with strong overlap and an active books section to a separate test line. The final list names, for every site, the evidence behind the word similar.
Frequently asked questions
Is a similarity score a percentage of shared content?
Not for meaning based scores. They measure closeness between vectors, so a number like 0.82 means the texts sit close together in the model's space, not that 82 percent of the words match. Scores from different models or vendors are not comparable, so ask for the matching articles instead of the number alone.
Can a site be similar to a seed in only one section?
Yes, especially with general publishers. A city magazine may have a books section that matches a literary seed closely while the rest of the site covers dining and real estate. Measure similarity on the section's own articles, and if the buying system allows it, plan that section rather than the whole domain.
Does a high similarity score mean two sites perform alike?
No. Similarity describes content, readers, links or owners, not campaign results. Two very similar sites can deliver differently because of ad load, viewability, price, the exchanges that sell them and the season. Treat similarity as a reason to test a site, and judge performance from the campaign's own reporting.
Should sites with the same owner both stay on a list?
They can, but they should be reviewed and counted as one entry. A publisher group may run several distinct, high quality titles. The risk is a network of thin copies filling a list with what is really one site. Check each domain's content separately, then decide how much of the plan one owner should hold.
Sources
- Finding related pages in the World Wide Web (Computer Networks, 1999) (opens in a new tab)Jeffrey Dean and Monika Henzinger, ISTA Research Explorer record
- Hide and seek in Slovakia: utilizing tracking code data to uncover untrustworthy website networks (opens in a new tab)Mintal, Kalman and Fabián, arXiv, 2021
- Who Funds Misinformation? A Systematic Analysis of the Ad-related Profit Routines of Fake News sites (opens in a new tab)Papadogiannakis, Papadopoulos, Markatos and Kourtellis, The Web Conference 2023
- Spam policies for Google web search (opens in a new tab)Google Search Central
- ads.txt specification version 1.1 (opens in a new tab)IAB Tech Lab
- Guidelines on privacy and online behavioural advertising (opens in a new tab)Office of the Privacy Commissioner of Canada
- CrUX dimensions (opens in a new tab)Google Chrome for Developers
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.