Property proposals that send themselves.

A demo page with enough markup, scrolling and forms on it to see whether the replay pipeline actually carries the shape of a page.

Read more
4,281proposals sent this month
18smedian time to first open
62%read past the gallery

Three things worth testing

Hover the cards, scroll the page, type into the form. Every one of those should survive into the replay, and none of the typed values should.

Scroll depth

Long pages are where the interesting behaviour lives. If the replay truncates, this is where you find out.

Form masking

The fields below are masked by default. The account number is masked explicitly. The block below is hidden outright.

Chunk ordering

Chunks arrive every five seconds, plus a final uncompressed one on unload. Sequence is what makes the replay join up.

Try the form

Type whatever you like. It should be recorded as a field interaction with the value masked.

Masked automatically by rrweb.
Masked explicitly with .rr-mask.
Scratch pad

Type an email, a phone number or a card number here. Text masking should turn each of them into *** at record time, and leave this sentence alone.

A contenteditable paragraph, so text masking can be checked somewhere that is not an input.

This whole block has .rr-block

It should appear as an empty grey rectangle in the replay, not as its real content.

Something to scroll through

Lorem ipsum is a cop-out, so this is prose about the thing itself. A session replay is a stream of DOM mutations: an initial snapshot of the whole document, then every subsequent change encoded as a typed operation against a node id. Added nodes, removed nodes, attribute changes, text changes, scroll positions, mouse coordinates, input events.

The consequence is that a replay is not video. It is a reconstruction, which means it renders at whatever viewport the player has, and it weighs far less than a screen recording of the same length. It also means anything the recorder chose not to capture is gone, not blurred, which is why masking has to happen at record time rather than at playback.

The other consequence is that chunk boundaries matter. Events are only meaningful in order, so a missing chunk is not a dropped frame, it is a broken reconstruction. The player is forgiving about this in ways that make bugs easy to miss, so it is worth checking the chunk count against what the tracker thinks it sent.

Payload size is the thing to watch. The initial snapshot dominates: on a page with a lot of markup and inline styles it can be tens of kilobytes before any mutation happens. Subsequent chunks are small by comparison. If a property ends up with a large average chunk size, the usual cause is a page that repaints something expensive on scroll, not a page with a lot of text.

Sampling is the lever. Mouse movement at full rate is noise and bytes for no analytical gain, so it gets sampled down. Scroll is cheaper to keep. Media elements emit positions at a slow interval, and inputs are deduplicated to the last value in a tick, since intermediate keystrokes tell you nothing extra.

None of this is specific to this demo. It is the same shape whether the page is a property proposal, a checkout, or a marketing site with a modal on it. Which is really the point: the pipeline should not care what the page is, only that the events line up.

recording