Concepts
Co-citation in SEO: what shared linkers show
Understand co-citation, shared linking domains and retained overlap. See a worked example and the limits of similarity scores.
Co-citation describes two items being cited by a common source. In domain-level website research, a shared linker is a source domain with links to both target domains. Those links can appear on different pages. Shared linkers show observed graph overlap; they do not prove affiliation, matching audiences or a Google ranking benefit.
Where co-citation comes from
Henry Small’s 1973 co-citation paper studied relationships between documents cited together. Adapting that idea to website domains gives us a way to compare their incoming link neighborhoods. The historical definition does not establish that a search engine uses an equivalent ranking signal.
Keep the unit explicit. Two pages may link to the same sources; two targets may receive links from the same domain; or two retained source lists may overlap. These are different observations. This article uses domain-level incoming links unless it says otherwise.
Three observations that need separate labels
| Observation | Supported statement | Still unproved |
|---|---|---|
| One page links to A and B | Both targets appear as links on that captured page | Endorsement or business affiliation |
| One domain links to A and B | The domain is a shared linker | Same-page or same-sentence proximity |
| Two retained linker lists overlap | They share the retained sources shown | Complete overlap across the web |
A direct link from A to B is another relationship entirely. Co-occurrence can also describe names appearing together in text without a hyperlink. Use the narrower term that your evidence supports; a domain graph alone cannot reconstruct the surrounding words.
Illustrative graph direction: source X → target A; source X → target B. X is the shared incoming linker. A does not need to link to B.
A worked co-citation calculation
Consider these fictional retained lists. Target A has publisher.example, directory.example and forum.example. Target B has publisher.example, forum.example and news.example. These are teaching examples, not live LinkCensus results.
The intersection contains two domains: publisher and forum. The union contains four distinct domains. Retained-set Jaccard similarity is intersection divided by union: 2 / 4 = 0.5. In this example, 0.5 is a ratio between these two lists, not a probability that the targets are competitors.
The LinkCensus method separates embedding-based ranking from retained shared-linker explanations. In co-citation mode, the overlap score uses retained intersection divided by retained union. An embedding score represents a different calculation and must not be read as a shared-domain count.
Why the overlap count and ratio behave differently
LinkCensus presents shared linkers (among top 200 retained). Retaining only a subset can hide observed common sources. The displayed shared count is therefore a lower bound on overlap in the underlying full lists from that graph. It is not a complete count of current web relationships.
The normalized ratio has no equivalent lower-bound guarantee. Suppose both lists also contain one additional, different nonshared source. The full intersection stays at two, while the union becomes six: 2 / 6 is about 0.33. Removing those nonshared sources raises the retained ratio to 0.5.
Now consider a different full-list extension: both targets also share club.example. The full intersection is three and the union is five, giving 0.6. Omitting that shared source lowers the retained ratio to 0.5. Truncation can move the ratio in either direction.
This distinction matters when comparing results. A retained count can understate observed overlap, while a retained ratio may be higher or lower than the full-set ratio. Neither calculation establishes the percentage of the internet shared by two websites.
Read the sources before drawing a conclusion
Common Crawl’s graph documentation, reviewed on 26 September 2026, explains that technical resource links are included. A font host and a specialist journal can both contribute graph relationships, but they supply different evidence about subject matter. Inspect the source context when it matters to your decision.
Use the five-stage provenance checklist to preserve the source date, aggregation unit and product edition. When you have two retained datasets, the edition comparison workflow helps test whether their units and filters are comparable before interpreting a changed overlap. The difference between referring domains and backlinks also explains why a domain count cannot be treated as a page-level link count.
For practical discovery, look up a domain in LinkCensus, inspect its retained shared linkers and validate candidate pages against a stated purpose. Our similar-website workflow shows how to record those checks. Treat the resulting sites as research leads until current evidence supports a stronger label.
What is a co-citation?
A co-citation occurs when a common source cites two items. In a domain-level link graph, the analogous observation is a source domain linking to both target domains. The two links may come from separate pages. Always specify the unit and capture context when describing this relationship.
Shared-linker research also gives no reason to manufacture links for rankings. Google’s spam policies identify links created mainly to manipulate rankings as link spam. Use observed relationships to guide research; do not present co-citation as a proven shortcut to search visibility.
