How Tor Chooses Relays: Geography, Guards, Bandwidth, and Network Diversity

A common assumption is that Tor behaves like a content-delivery network and automatically selects nearby relays.

It does not.

A Tor user in Japan is not automatically routed through Japanese relays, just as a user in Germany is not guaranteed to begin a circuit in Germany.

Tor deliberately prioritizes anonymity properties and load balancing over geographic proximity.

A typical Tor circuit

A conventional Tor circuit normally contains three relay positions:

Client
  ↓
Guard
  ↓
Middle
  ↓
Exit
  ↓
Destination

The three positions have different responsibilities.

The Guard sees the client’s IP address but normally does not know the final destination.

The Middle Relay connects the Guard to the Exit and sees neither the original client nor the final destination in a directly useful form.

The Exit Relay connects to the destination server and therefore sees the destination traffic leaving the Tor network, although it does not directly know the original client.

Tor does not perform simple geographic nearest-node selection

A user in Tokyo might receive a circuit resembling:

Japan
  ↓
Guard in Europe
  ↓
Middle in North America
  ↓
Exit in another European country

Another circuit might involve a nearby relay, but this is not because Tor intentionally chose the geographically closest machine.

Relay selection considers factors such as:

• relay flags;
• measured bandwidth;
• consensus weights;
• relay family constraints;
• network diversity;
• exit policies;
• current Guard selection state.

Physical distance is not the main routing criterion.

This means that increasing the number of Japanese relays does not create a private Japanese Tor region or automatically reduce latency for Japanese users.

Why Guards behave differently

Guard selection is intentionally conservative.

If a client selected a completely new first-hop relay for every circuit, an attacker operating enough relays would eventually have a greater chance of becoming the client’s entry point.

Tor therefore maintains a relatively stable Guard set.

Conceptually:

choose suitable Guards
        ↓
keep using them for a long period
        ↓
replace them only when necessary

This reduces exposure to malicious entry relays.

Consequently, even if many excellent new relays appear in a user’s home country, existing Tor clients may not immediately begin using them as Guards.

Middle Relays are globally useful

A new non-exit relay usually begins its life primarily as a Middle Relay.

Its traffic may originate from clients anywhere in the world.

For example:

Guard in Canada
      ↓
Middle in Japan
      ↓
Exit in France

is entirely normal.

The Japanese Middle Relay is therefore not “serving Japanese users.” It is contributing routing capacity to the global Tor network.

This is an important distinction.

The value of a relay located in Japan is not primarily:

providing a geographically local service to Japanese Tor users.

Its more important contribution is:

adding another independent routing location inside the global anonymity network.

Bandwidth influences selection probability

Tor does not distribute traffic equally among relays.

A relay capable of sustaining significantly more throughput will generally receive more traffic than a much slower relay, assuming both are otherwise eligible for similar roles.

Very roughly:

Relay A
measured bandwidth: high

Relay B
measured bandwidth: low

Relay A is normally assigned a greater probability of being selected.

However, raw bandwidth alone does not determine everything. Consensus weight calculations also balance traffic across Guard, Middle, and Exit positions.

For this reason:

advertised bandwidth
≠
measured bandwidth
≠
consensus weight

These are related but different concepts.

Why network diversity matters

Tor also attempts to avoid constructing circuits that concentrate too many hops within the same network.

This matters because two relays operated in closely related infrastructure may be observable by the same organization.

For example:

Guard → major hosting provider A
Middle → major hosting provider A
Exit → another server in provider A

would provide poor infrastructure diversity.

Tor therefore uses information such as relay families and network prefixes when building paths.

From a wider architectural perspective, this is why independently operated relays in less common autonomous systems can be valuable even when they do not provide enormous amounts of bandwidth.

Geography still affects performance

Although Tor does not select relays primarily by distance, geography still affects latency.

A circuit involving three nearby regions will often have lower round-trip time than one repeatedly crossing oceans.

For example:

East Asia → East Asia → Southeast Asia

will generally have lower physical propagation delay than:

East Asia → Western Europe → North America

But deliberately optimizing circuits for geographic closeness could weaken anonymity properties or increase predictability.

Tor therefore accepts some performance cost in exchange for stronger routing diversity.

A Tor relay is part of a global pool

The most useful mental model is not:

Japanese relay
→ serves Japanese users

but instead:

relay located in Japan
→ becomes one candidate in a global anonymity network

A stable relay may eventually participate in circuits created by users from dozens of countries.

That is why the significance of a relay cannot be measured only by how much traffic it serves locally.

Tor is not a national network connected to other national Tor networks.

It is one distributed global network, and every relay increases the number of possible paths through it.