Every per platform LUFS table you have seen is a rumour

Spotify wrote its number down, and no YouTube page has

Spotify's loudness page does not only state the target. It states the direction: positive gain to softer masters, negative gain to louder ones, an album normalised as one unit so gain compensation does not change between tracks. It states the scope it does not cover, in its own words: the web player and third party devices such as TVs and speakers do not use loudness normalisation at all. A number, a direction and a stated exception. That is what a published target reads like, and a second Spotify page carries the same figure.

YouTube's upload encoding page, where a creator is sent for encoding requirements, is specific to the point of pedantry: AAC-LC, Opus or Eclipsa Audio, 48 kHz, 128 kbps for mono, 384 for stereo, 512 for 5.1. Not one loudness figure of any kind. No LUFS, no LKFS, no dB target.

The stronger evidence is the stable volume page, because it is about loudness. It says the feature balances the range between quiet and loud by continuously adjusting volume levels, that it is on by default, and that it is turned off for YouTube Music and official music videos. Three facts about a loudness feature and no level. The silence is not an oversight on one encoding page. It holds across the product.

The YouTube figure came out of a debug overlay, one video at a time

Ian Shepherd published the method on Production Advice on 29 September 2017. Open Stats for Nerds and read the Content Loudness value, which "indicates the difference between YouTube's estimate of the loudness and their reference playback level". A reading of 6 dB means the video is 6 dB louder than YouTube's distribution loudness and the page says a normalisation adjustment of -6 dB is applied to compensate. A negative reading means it is quieter, and nothing happens.

That is a real measurement and the best available. It is also a reading off a diagnostic panel, one video at a time, with no sample size on the page, and Shepherd's own text says the exercise is not about aiming for -14 LUFS. The tables built on it kept the number and dropped that sentence.

Podnews, in its loudness FAQ for podcasters, writes that Spotify and YouTube ask for -14 LUFS while most podcasters still produce a -16 LUFS file. Spotify asks. No YouTube or Google page asks for anything of the kind. That is how the figure travels: not through content farms but through careful publications repeating a number whose origin nobody re-checks.

Same number, two provenances: one is a sentence a platform wrote about itself, the other is a reading off a debug panel.

Spotify raises a quiet master, and the 2017 reading says YouTube leaves it quiet

The AES has the vocabulary for this. TD1008 calls it downward normalization when the audio sits above the target and is attenuated, and upward normalization when it sits below and positive gain is applied, which it notes may require peak limiting. Two directions, named separately, because a platform can implement one and not the other.

Spotify implements both and says so: "Positive gain is applied to softer masters so the loudness level is -14 dB LUFS." It says what it gives up to do it, leaving 1 dB of headroom for lossy encoding. Shepherd's reading of the overlay says the opposite about YouTube, with the verb italicised on the page: "YouTube doesn't turn up quieter videos."

Deliver a mix at -20 LUFS to Spotify and it comes back up to -14. Deliver the same mix as a video and, on the 2017 behaviour, it stays at -20 beside a feed of material sitting near the ceiling, and the viewer reaches for the volume or leaves. A table of targets cannot show that, which is why every table in circulation omits the property that decides the outcome.

That measurement is from 2017. Stable volume shipped later, and continuously adjusting volume levels is not obviously the same behaviour as leaving quiet audio alone, so date the claim to Shepherd's reading when you use it. Stable volume carries no figure either, so nothing published since would settle it.

A published target carries a tolerance, a meter and a revision history

EBU R 128 is what this genre looks like done properly. Recommendation (h) says programme loudness shall be normalised to a target level of -23.0 LUFS, permits plus or minus 1.0 LU where the target is not practically attainable, names live programmes as the example, then warns that drifting to the limits of that tolerance must not become standard practice. Recommendation (i) allows plus or minus 0.2 LU in quality control for measurement error. Recommendation (m) caps true peak at -1 dBTP. The current version is the fifth, Geneva, November 2023, and the document's history table records what moved: the relative gate from -8 to -10 LU in 2011, the tolerance in 2014.

The AES streaming recommendation everybody cites is a range, not a point. TD1004, dated 19 October 2015, recommends that target loudness not exceed -16 LUFS and not be lower than -20 LUFS. Its successor TD1008, 24 September 2021, publishes a table instead: -18 LUFS where speech is measurable, upper tolerance 1 LU, and -16 LUFS for track normalized music, upper tolerance 0.2 LU, with the instruction that distribution loudness is not to be targeted to the upper tolerance. Anyone quoting the AES at -16 is quoting one end of a band and dropping the other.

Two more things no table tells you. TD1008 recommends, in its own text, that the industry reconvene and lower its distribution loudness recommendations by 6 LU to harmonise with EBU R 128 and ATSC A/85 at -23 to -24 LUFS, so the body that wrote the streaming numbers thinks they are too loud. And its scope excludes sound with picture content, so the document most often cited for streaming loudness is not a video recommendation at all. Its official AES download link returns 404, which is a small comment on how carefully it gets read.

Congress legislated loudness and stopped at the television set

47 CFR 73.682(e)(1) is short: effective 13 December 2012, television broadcast stations must comply with the ATSC A/85 recommended practice, incorporated by reference, insofar as it concerns the transmission of commercial advertisements. Read the whole of subsection (e) and you will not find a number. It arrives by reference: the edition of A/85 dated 12 March 2013 says the target loudness value should be -24 LKFS, anticipates variations of about plus or minus 2 dB, and asks for true peak below -2 dB TP. The one rule with legal force borrows its number.

The statute behind it, 47 U.S.C. 621, names the ATSC recommended practice by title and applies to television broadcast stations, cable operators and multichannel video programming distributors. Online video is not in it. The IAB Tech Lab says as much in its August 2022 ad format guidelines: the CALM Act does not yet apply to digital video online, and publishers are likely to normalise anyway to protect their brand. Then the same document borrows the broadcast numbers, -24 LKFS plus or minus 2 in the US and -23 LUFS plus or minus 1 in the EU, because there is nothing else to borrow.

Google borrows them too, where it sells advertising: Ad Manager normalises audio to a target loudness value of -24 LKFS, and Display and Video 360 states the same figure and credits the IAB spec by name. Google also publishes -16 LUFS for stereo and -19 LUFS for mono on its Assistant developer page, with the arithmetic, since duplicating a mono track across two channels doubles the energy and adds 3.01 LU. Three loudness figures across three Google pages, and not one of them about YouTube playback.

So what do you actually mix a video to

No platform in this set publishes a loudness target for video, so the choice is between borrowing a number and choosing a behaviour. Borrow -14 LUFS because a table says YouTube wants it and you are borrowing a 2017 reading off a debug panel. Borrow -16 LUFS from the AES and you are borrowing a figure its own authors want 6 LU lower, from a document whose scope excludes anything with a picture.

What survives is the asymmetry. A mix that is too loud gets turned down, and the only cost is that somebody else's limiter decided. A mix that is too quiet has nothing documented to rescue it, on the one behaviour anyone has measured. So do not aim at the middle of the band: -16 LUFS is the loud end of the AES range, and -20 LUFS is a bet on a platform doing something no page says it does.

What to do about it

1. Before mixing to a per platform number, find the sentence on the platform's own page. If there is no sentence, you are borrowing somebody's measurement, and the brief should say whose and from when.
2. Aim at -16 LUFS rather than -20 LUFS, the two ends of the AES streaming band, because the only normalisation ever measured on YouTube turns loud audio down and leaves quiet audio where it is.
3. Hold true peak at -1 dBTP and record the integrated loudness of every render, so a number exists to argue with when somebody says the video sounds quiet.