1,200+ custom songs and 20+ genres: what running Magical Song taught us about AI songs as gifts

We run a product studio called Inithouse. One of our products, Magical Song, turns a personal story into a studio-quality custom song with real vocals. The user writes a few sentences about a person or a moment, picks a genre, and gets a finished song in minutes.

We crossed 1,200 generated songs a few weeks ago. This post is about what that volume taught us about AI-generated music when the use case is gifting, not production.

The gift angle changes everything

Most AI music tools (Suno, Udio) are built for creators who want to make tracks. The user thinks in terms of prompts, styles, stems, loops. Our users think in terms of "my mom turns 60 on Saturday."

That distinction shaped every product decision we made. The interface asks for a story, not a prompt. The output is a shareable link, not a downloadable WAV. The genres are labeled "acoustic love ballad" and "reggae birthday jam," not "lo-fi chill hop 120bpm."

When we looked at the data across 1,200+ songs, birthday songs accounted for roughly a third. Anniversaries, weddings, and "just because" gifts split the rest fairly evenly. Holiday spikes (Valentine's Day, Mother's Day, Christmas) were real but shorter than we expected, usually a 3-4 day window.

What makes a story "singable"

Not every story translates well into lyrics. We learned this the hard way by reading through output quality reports.

Stories that work well tend to have three things: a specific person, a concrete memory, and at least one sensory detail. "My dad taught me to fish at Lake Lipno when I was seven, and he always burned the sausages" produces a song people actually cry over. "My dad is the best dad ever and I love him" produces something generic that could be about anyone's dad.

We don't reject vague inputs. But we noticed that the songs from detailed stories get shared about 2x more often than the ones from generic messages. Sharing is the closest proxy we have for "this actually landed as a gift."

Early on, we considered adding a guided input flow that would prompt for specifics: "Name a place," "Name a food," "What did they always say?" We haven't shipped it yet, but the data keeps pointing that direction.

Where AI songs break

Genre matters more than we assumed. Some genres handle personal lyrics gracefully. Acoustic, pop ballad, country, and reggae consistently produce clean results. The vocal delivery matches the emotional tone of a personal story.

Hard rock, metal, and hip-hop are trickier. The vocal style can clash with sentimental content. "Happy 40th birthday, Mom" in a death metal arrangement is either hilarious or terrible depending on context, and we have no way to know which one the buyer intended.

We offer 20+ genres. About 80% of orders land in five or six genres. The long tail exists mostly for the users who specifically want a joke gift or something unexpected. That is a valid use case we try not to break, but it is harder to quality-control.

Another breakage point: names. Non-English names, especially Czech, Polish, or Hungarian names with diacritics, sometimes get mispronounced by the vocal model. We have partial fixes for this (phonetic hints, pronunciation guides fed into the generation pipeline), but it remains the single most common complaint at a 4.9/5 average rating.

Three builder-level decisions that shaped the product

One: no account required for preview. The recipient gets a link, taps play, hears their song. No sign-up wall, no app download. This matters because the song is a gift, and gifts should not require the recipient to create an account before they can open them. It cost us some tracking granularity but improved the share-to-listen conversion rate significantly.

Two: one-time purchase, not subscription. A birthday song is a one-time event. Forcing a subscription model on a gift product felt wrong, and the churn would have been brutal. We charge once per song. Repeat usage comes from the same person making songs for different occasions, not from a recurring billing cycle.

Three: shareable link as the default output. We could have defaulted to a downloadable MP3. Instead, the primary output is a hosted page with the song, the story behind it, and a play button. This turned every song into a distribution channel. When someone shares the link on social media or in a family group chat, the page shows what the product does without us having to explain it.

What we are measuring now

We track two things closely: the ratio of previews to paid unlocks (tells us if the AI output quality is good enough), and the share rate per song (tells us if the gift actually landed).

Both numbers have been stable over the last few months, which either means we found a local optimum or we are not iterating fast enough. We are currently testing whether adding a short video version of the song (lyrics on screen, album-art-style visuals) increases share rates. No results yet.

The broader question we keep coming back to: is "AI custom song as a gift" a category, or is it a feature that a larger music platform will absorb? Suno and Udio keep adding personalization features. Our bet is that the gift workflow (story in, shareable link out, no music knowledge required) is different enough to sustain a standalone product. But we will see.

If you have shipped a product where the buyer and the end user are different people (gifts, cards, surprise experiences), what patterns did you see in how people describe what they want?