Send two references, not five. Level match them to your mix before anyone listens, usually a 4 to 6 dB pull, because a mastered record played next to an unmastered mix always wins and the comparison teaches nothing. Then attach a timestamp and one attribute to each: "1:12, the low end there is continuous." That is the whole method. Most of the wasted revision rounds I see are not disagreements about taste, they are two people listening at two different volumes and describing two different things.
What is a reference track actually for?
A reference answers one question at a time. It is a measuring instrument, not a target.
The useful questions are narrow: how wide the guitars sit, whether kick and bass share the bottom or split it, how much of the vocal is dry, how loud the snare is against everything else. The useless question is "make it sound like this," because a record is the sum of a song, an arrangement, a performance, a room and a mastering pass, and only two of those are on the mixing desk.
The practical consequence is that a reference is worth more early than late. During the arrangement it tells you there are too few parts or too many. During the mix it tells you where the low end lives. After the mix is 90 percent finished it mostly makes everyone nervous.
Why does the reference always sound better?
Two reasons, and one of them is arithmetic.
Commercial releases are mastered, and a typical pop master sits several dB above an unmastered mix. Loudness reads as quality at equal fidelity, which is why every mastering house level matches before it demonstrates anything.
The second reason is playback. Streaming players normalise, so on the listener's end the gap shrinks or disappears. Spotify normalises playback to roughly -14 LUFS by default, meaning the loud master you are comparing against gets turned down on the phone that plays it. A good part of the gap you hear in the session is an artifact of the session.
How do you level match a reference correctly?
Use an integrated loudness reading, not a peak meter and not your fader.
Any meter implementing ITU-R BS.1770, the loudness algorithm underneath both broadcast and streaming specs, will do. Play the loudest chorus of your mix, note the short-term value, then trim the reference bus until its loudest chorus reads within about 0.5 dB of that. Not the intro, and not an average across the whole song, because arrangements differ and an average flatters whichever track has longer quiet sections.
Then fix the monitor level and leave it. Balance judgements move with playback level, because the ear's sensitivity at the frequency extremes changes with level, which is what the ISO 226 equal-loudness contours describe. A chorus balanced at one volume and a reference judged at another are two different mixes being compared.
Two mechanical details. Import the reference at the session sample rate rather than letting the DAW convert on the fly, and route it to a bus that bypasses the mix bus entirely. A reference passing through your own bus compressor is not a reference any more.
How many references, and how do you choose them?
Two. One for tone, one for arrangement density. Three is already a committee.
Choose them from the same production era and roughly the same budget class as the record you are making. A 2024 pop single and a 1976 rock album are both instructive, but not at the same time, and the notes they generate contradict each other. The same goes for instrumentation: if your record is one acoustic guitar and a voice, a reference with twelve tracks of production tells you nothing you can act on.
The test I apply before accepting a reference is whether I can name what I am supposed to hear in it in one sentence. If not, it is a mood board, and mood boards do not survive contact with a fader.
What do you actually send with the reference?
Three fields per reference. Timestamp, element, attribute.
"0:48, the vocal, it sits further back than mine" is actionable in one listen. "It has more energy" is a guess, and every guess costs a full revision round. If the reference is only right in one section, say so, because half the confusion in reference notes comes from people sending a whole song when they mean eleven seconds of it. And name what you do not want from it. "The drums, not the vocal sound" saves an afternoon.
Where do references quietly lie to you?
The version matters. A streaming version, a vinyl master and a 2009 CD of the same song can differ by several dB in the low end and a full dB of high shelf. If you are chasing a tonal balance, the version you send is part of the note.
Spectral match plugins are the second trap. Matching your mix's curve to a mastered reference imports the mastering EQ and the limiter's effect on the spectrum into your mix, which is why match-EQ'd mixes so often sound close and lifeless at once. Use the curve as a diagnostic, then make the moves by hand on the elements causing them.
Third, references have arrangements. A record whose low end sounds continuous usually has a bass part that plays continuously. That is a writing decision made months before anyone opened a session, and no amount of EQ produces it after the fact.
What happens to the reference on the producer's side?
References go on their own stereo bus, no processing, trimmed once at the start of the project against the first rough. I check against them about three times: after the arrangement is locked, once when the low end is set, and once near the end. In between the session stays closed to them, because constant referencing is how mixes turn into averages of other people's records.
At the end I do one pass with the references muted entirely, quiet, on a single speaker. If the record still tells its story there, the references have done their job and can go away.
Working across countries adds one step, and it is why I ask for references in week one rather than week four. A reference chosen by an artist in Kyiv and a reference chosen by a label in Los Angeles are frequently different records with different assumptions about vocal loudness and low end, and finding that out early is much cheaper than finding it out late. The [credits page](/credits.html) lists the records where that conversation had to happen before anyone touched a fader, with the source for every role.
What to do with this
On the next mix you send out, put two references in the same email as the files, each with a timestamp and one attribute, and level match them yourself first so you know what you are actually hearing. If a note survives level matching, it is real. If it does not, you just saved a round. More on how I work is on the [about page](/about.html).