Browsers can set the Save-Data: on header to indicate that a user is on a connection that's slow or has high data costs. Websites can then adjust their responses to take that into account.
How many websites have added support for this header? We tested the homepages of 1092 sites and found 4 where the Save Data header had a noticeable impact (~0.4%).
Who supports Save-Data, and what impact does it have?
These websites supported the header to some extent:
- Answers.com
- Nike.com
- Okzone.com (an Indonesian news website)
- Shopify.com
Answers, Nike and Okezone all just decrease the quality of their images. Answers and Nike both use Cloudinary, which automatically applies this optimization if the image quality level is set to q_auto.
Shopify also saves some data by serving smaller images – in this case by only serving 1x versions of each image instead of high-resolution ones at double the size.
But most of Shopify’s savings come from not including a tag manager script if Save-Data is enabled. This brings down the number of requests from 82 to 24 and saves 5s of CPU time. Overall page weight goes down 55% and the Lighthouse Performance score jumps form 50 to 88.
Do browsers send the Save-Data header?
There's no point in sending less data in response to requests containing the Save-Data header if browsers don't actually set the header. So do they?
Tim Kadlec collected some data in August 2019 and found that on his site around 4% of sessions send the header. Shopify reported 20% of requests from India and Brazil including the header.
For this blog it's about 1.7%, based on a sample of 12,000 page views.
Methodology
We ran Lighthouse 10 times for each site, half the time with the header and half the time without. The page weight chart above compares the median of each of those 5 tests.
After that we isolated each request URL and checked if the response size is affected by the header.
There are probably other sites where the header has some impact, but it may be very small (<2% saved) or hard to detect because of underlying variance in the site's page weight.
These sites were tested:
- Tranco top 1000
- 100 services from the SaaS 1000
- 111 Indian news sites
Some sites were removed because they were duplicates or untestable.