Content negotiation

In HTTP, content negotiation is the mechanism that is used for serving different representations of a resource at the same URI, so that the user agent can specify which is best suited for the user (for example, which language of a document, which image format, or which content encoding).

Principles of content negotiation

A specific document is called a resource. When a client wants to obtain it, it requests it using its URL. The server uses this URL to choose one of the variants it provides – each variant being called a representation – and returns this specific representation to the client. The overall resource, as well as each of the representations, have a specific URL. How a specific representation is chosen when the resource is called is determined by content negotiation and there are several ways of negotiating between the client and the server.

The determination of the best suited representation is made through one of two mechanisms:

• Specific HTTP headers by the client (server-driven negotiation or proactive negotiation), which is the standard way of negotiating a specific kind of resource.
• The 300 (Multiple Choices) or 406 (Not Acceptable) HTTP response codes by the server (agent-driven negotiation or reactive negotiation), that are used as fallback mechanisms.

Over the years, other content negotiation proposals, like transparent content negotiation and the Alternates header, have been proposed. They failed to get traction and got abandoned.

Server-driven content negotiation

In server-driven content negotiation, or proactive content negotiation, the browser (or any other kind of user-agent) sends several HTTP headers along with the URL. These headers describe the preferred choice of the user. The server uses them as hints and an internal algorithm chooses the best content to serve to the client. The algorithm is server-specific and not defined in the standard. See, for example, the Apache 2.2 negotiation algorithm.

The HTTP/1.1 standard defines list of the standard headers that start server-driven negotiation (Accept, Accept-Charset, Accept-Encoding, Accept-Language). Though strictly speaking User-Agent is not in this list, it is sometimes also used to send a specific representation of the requested resource, though this is not considered as a good practice. The server uses the Vary header to indicate which headers it actually used for content negotiation (or more precisely the associated response headers), so that caches can work optimally.

In addition to these, there is an experimental proposal to add more headers to the list of available headers, called client hints. Client hints advertise what kind of device the user agent runs on (for example, if it is a desktop computer or a mobile device).

Even if server-driven content negotiation is the most common way to agree on a specific representation of a resource, it has several drawbacks:

• The server doesn't have total knowledge of the browser. Even with the Client Hints extension, it has not a complete knowledge of the capabilities of the browser. Unlike reactive content negotiation where the client makes the choice, the server choice is always somewhat arbitrary.
• The information by the client is quite verbose (HTTP/2 header compression mitigates this problem) and a privacy risk (HTTP fingerprinting)
• As several representations of a given resource are sent, shared caches are less efficient and server implementations are more complex.

The Accept header

The Accept header lists the MIME types of media resources that the agent is willing to process. It is comma-separated lists of MIME types, each combined with a quality factor, a parameter indicating the relative degree of preference between the different MIME types.

The Accept header is defined by the browser, or any other user-agent, and can vary according to the context, like fetching an HTML page or an image, a video, or a script: It is different when fetching a document entered in the address bar or an element linked via an <img>, <video> or <audio> element. Browsers are free to use the value of the header that they think is the most adequate; an exhaustive list of default values for common browsers is available.

The Accept-CH header

The experimental Accept-CH lists configuration data that can be used by the server to select an appropriate response. Valid values are:

The Accept-Charset header

The Accept-Charset header indicates to the server what kinds of character encodings are understood by the user-agent. Traditionally, it was set to a different value for each locale for the browser, like ISO-8859-1,utf-8;q=0.7,*;q=0.7 for a Western European locale.

With UTF-8 now being well-supported, being the preferred way of encoding characters, and to guarantee better privacy through less configuration-based entropy, most browsers omit the Accept-Charset header: Internet Explorer 8, Safari 5, Opera 11 and Firefox 10 have abandoned this header.

The Accept-CH-Lifetime header

The Accept-CH-Lifetime header is used with the Device-Memory value of the Accept-CH header and indicates the amount of time the device should opt-in to sharing the amount of device memory with the server. The value is given in miliseconds and it's use is optional.

The Accept-Encoding header

The Accept-Encoding header defines the acceptable content-encoding (supported compressions). The value is a q-factor list (e.g.: br, gzip;q=0.8) that indicates the priority of the encoding values. The default value identity is at the lowest priority (unless otherwise declared).

Compressing HTTP messages is one of the most important ways to improve the performance of a Web site, it shrinks the size of the data transmitted and makes better use of the available bandwidth; browsers always send this header and the server should be configured to abide to it and to use compression.

The Accept-Language header

The Accept-Language header is used to indicate the language preference of the user. It is a list of values with quality factors (like: "de, en;q=0.7"). A default value is often set according the language of the graphical interface of the user agent, but most browsers allow to set different language preferences.

Due to the configuration-based entropy increase, a modified value can be used to fingerprint the user, it is not recommended to change it and a Web site cannot trust this value to reflect the actual wish of the user. Site designers must not be over-zealous by using language detection via this header as it can lead to a poor user experience:

• They should always provide a way to overcome the server-chosen language, e.g., by providing a language menu on the site. Most user-agents provide a default value for the Accept-Language header, adapted to the user interface language and end users often do not modify it, either by not knowing how, or by not being able to do it, as in an Internet café for instance.
• Once a user has overridden the server-chosen language, a site should no longer use language detection and should stick with the explicitly-chosen language. In other words, only entry pages of a site should select the proper language using this header.

The User-Agent header

The User-Agent header identifies the browser sending the request. This string may contain a space-separated list of product tokens and comments.

A product token is a name followed by a '/' and a version number, like Firefox/4.0.1. There may be as many of them as the user-agent wants. A comment is a free string delimited by parentheses. Obviously parentheses cannot be used in that string. The inner format of a comment is not defined by the standard, though several browser put several tokens in it, separated by ';'.

The Vary response header

In opposition to the previous Accept-* headers which are sent by the client, the Vary HTTP header is sent by the web server in its response. It indicates the list of headers used by the server during the server-driven content negotiation phase. The header is needed in order to inform the cache of the decision criteria so that it can reproduce it, allowing the cache to be functional while preventing serving erroneous content to the user.

The special value of '*' means that the server-driven content negotiation also uses information not conveyed in a header to choose the appropriate content.

The Vary header was added in the version 1.1 of HTTP and is necessary in order to allow caches to work appropriately. A cache, in order to work with server-driven content negotiation, needs to know which criteria was used by the server to select the transmitted content. That way, the cache can replay the algorithm and will be able to serve acceptable content directly, without more request to the server. Obviously, the wildcard '*' prevents caching from occurring, as the cache cannot know what element is behind it.

Agent-driven negotiation

Server-driven negotiation suffers from a few downsides: it doesn't scale well. There is one header per feature used in the negotiation. If you want to use screen size, resolution or other dimensions, a new HTTP header must be created. Sending of the headers must be done on every request. This is not too problematic with few headers, but with the eventual multiplications of them, the message size would lead to a decrease in performance. The more precise headers are sent, the more entropy is sent, allowing for more HTTP fingerprinting and corresponding privacy concern.

From the beginnings of HTTP, the protocol allowed another negotiation type: agent-driven negotiation or reactive negotiation. In this negotiation, when facing an ambiguous request, the server sends back a page containing links to the available alternative resources. The user is presented the resources and choose the one to use.

Unfortunately, the HTTP standard does not specify the format of the page allowing to choose between the available resource, which prevents to easily automatize the process. Besides falling back to the server-driven negotiation, this method is almost always used in conjunction with scripting, especially with JavaScript redirection: after having checked for the negotiation criteria, the script performs the redirection. A second problem is that one more request is needed in order to fetch the real resource, slowing the availability of the resource to the user.