Something’s shifted in the kind of requests I get behind the counter. Used to be customers walked in with three sources: a magazine article, a friend’s recommendation, or a record they’d heard at a party. Now there’s a fourth, and it’s becoming the dominant one. They’ve been chatting with an AI model — Claude, ChatGPT, whatever — and it’s recommended six obscure records and they want to know if I stock any of them.

Last Saturday a young bloke came in with a printed list of fourteen records. “AI said these are essential if I like the Velvet Underground.” Some of them were spot on — Television, Modern Lovers, the obvious moves. Some were genuinely interesting picks I wouldn’t have made — a Galaxie 500 record, a Felt album, some Slapp Happy. And one was completely hallucinated. The “1974 second album by Crystal Mainline” doesn’t exist. Crystal Mainline doesn’t exist.

This is the discovery economy now and it’s worth thinking about what it’s doing to the second-hand market.

What the AI tools actually do well

I’ve been playing with several of them to see what they recommend. They’re surprisingly good at the canonical stuff — if a record has been written about extensively, sat on rateyourmusic lists, or got a long Wikipedia entry, the AI will dig it up and explain why someone who likes X might like Y. That’s useful. It functions like a polite, patient music nerd who’s read every Pitchfork review back to 2003.

Where they fall down is anything regional or anything that wasn’t extensively documented online. Australian indie from the 80s and 90s, for example. Ask Claude about the No Frills label or about early Aberration Records, and you’ll get a lot of vague hedging and a few outright fabrications. The training data has thin coverage of anything that didn’t get an international release and a heap of music writing in English.

For shops like mine, this is actually good news. The records that the AI confidently recommends are selling. The records that the AI can’t speak to — which is most of what I actually specialise in — still need a human to recommend them. The shop’s role hasn’t gone away. It’s shifted.

Demand spikes are weird now

Used to be that demand spikes on specific records came from clear events. A reissue, a death, a film soundtrack, a high-profile cover. Now I’m seeing demand spikes that don’t seem to correlate with anything visible, and the only explanation I can come up with is that an AI model has, for whatever reason, started recommending that record more often.

A few examples from the last six months. Three customers in two weeks asking for the first Slits album, none of whom had heard it before, all of whom had been “recommended” it. A run of about ten requests for original UK pressings of the first Cleaners From Venus album — that one’s properly obscure, and clearly somebody’s AI tool has decided it’s the perfect entry point for indie-pop deep diving. A whole month of people asking about a specific Brigitte Fontaine record that I’d previously sold maybe one copy of in five years.

You can’t predict it and you can’t plan inventory around it. The discovery layer is now sitting outside of music journalism and outside of word-of-mouth, and shop owners are noticing the downstream effects.

What this means for new customers

The customers I’m seeing who come in via AI discovery are different to the ones who came in via Pitchfork or YouTube algorithms five years ago. They’ve often got a more eclectic starting list — the AI doesn’t push them down a single genre rabbit hole the way an algorithm does. They’ve usually got context about why each record matters, even if some of that context is slightly off.

They’re also more willing to be told they’re wrong. A customer who heard about a record from a friend will get defensive if you tell them the version they’re after isn’t the one they want. A customer who got it from AI is happy to be corrected because they already half-suspect the AI was making things up.

The flip side: they’re less invested in any individual record. The list of fourteen they walked in with came from a conversation, not from years of building anticipation. If you don’t have something they’re after, they’ll shrug and check the next shop. There’s no fierce attachment.

A few thoughts on getting it right

For anyone running a shop, a couple of observations:

First, don’t be precious about the AI factor. The customer didn’t choose to find records this way because they don’t respect record stores — they’re just using the discovery tool that’s in front of them. Treat them like any other curious buyer.

Second, the conversation matters more than ever. Because AI lists are sometimes wrong or partial, the moment a customer walks in is when you add real value. “You like those? Then you should also hear these.” That’s the part the AI tools can’t do because they don’t know what just came in, what’s actually playing in the shop, or what your gut says.

Third, it’s worth having a view on tools like Discogs being the main reference layer for AI. Most music recommendation tools are sourcing from Discogs, rateyourmusic, and a smattering of music journalism. That means whatever isn’t documented well in those places isn’t going to surface via AI. If you want your local scene to show up in future AI recommendations, somebody needs to be writing it down somewhere indexable.

A friend who works with an AI consultancy — they help businesses build their own AI tools — made an interesting point: the discovery problem here is essentially a data coverage problem. The recommendations are only as good as what’s been written publicly about a given artist. If Australian indie from 1985-1995 wants to be findable via AI, somebody has to write the long-form articles, populate the Wikipedia entries, and feed the source material that future models will be trained on. Otherwise it stays invisible.

What I’m doing about it

Practically, I’m writing more. Posting more. Trying to put solid descriptions of stock onto the website so that next time somebody asks an AI about a specific Australian record, there’s actually something out there to base an answer on. It’s a slow grind but the shops that take it seriously now will have an outsized influence on the next decade of taste.

That, and I keep recommending records I love to customers who walk in. Some things don’t change.