Fifty people work without paychecks for two years while the founder maxes out eleven credit cards — and then the 348th investor finally says yes.
Every entry is researched against 30+ structured fields
For about twenty years, Tim Westergren made his living as a musician. Bands, session work, and eventually film scoring, with stretches working as a nanny between gigs to cover rent. It is not the biography of someone positioned to build a technology company.
But the film work contained a problem worth noticing. Directors could not tell him what they wanted. They could not say "give me something in a minor key with a syncopated rhythm section" — they said the music felt wrong, or almost right, and he had to work out why. So he built taste profiles by hand. He would listen through stacks of a director's reference music, isolate the attributes that kept recurring, and construct a working theory of what that person responded to.
Separately, touring as a journeyman musician, he kept meeting a second problem: genuinely good artists with no mechanism for being discovered. The industry's distribution filter was narrow, and talent that did not fit through it simply never surfaced.
Two observations sat in one career. If musical taste could be decomposed into specific, nameable attributes, then a listener's affinity for a song they already loved could point them at a song nobody had marketed to them.
The company that became Pandora started in January 2000 as Savage Beast Technologies, co-founded with Will Glaser and Jon Kraft, and it was not a consumer product. It was business-to-business: license the Music Genome Project to retailers as a recommendation engine. There were pilots with Best Buy in-store kiosks and with AOL Music. Roughly $2 million in seed capital came in.
By 2001 the money was gone. The timing was close to maximally hostile — the dot-com collapse had closed the funding window, and the Napster shutdown had made any company with the word "music" in its description radioactive to investors.
The ordinary decision at that point is to shut down. Westergren's tipping point is that he made the other one: he asked roughly fifty employees to keep working without being paid.
They said yes. That arrangement — not the 2000 incorporation — is the moment the company actually committed to existing, and it lasted about two years.
It is also the part of this story that should not be told admiringly without qualification. Deferring the salaries of fifty people is a decision that transfers a founder's risk onto employees who did not raise the money and would not own the upside in proportion. It produced exactly the consequence you would expect: former employees sued over unpaid wages. Westergren has said he did not realize the arrangement violated wage law. Reporting on the episode has been skeptical of that explanation, given that the company was already paying for legal counsel on music licensing.
What the unpaid years bought was a dataset nobody else had, built the least scalable way imaginable.
The Music Genome Project is a rubric. Trained musicians listen to a single song and score it across a fixed set of attributes — roughly 150 for rock and pop, closer to 350 for rap, 400 or more for jazz and classical — each rated on a 0-to-5 scale in half-point increments. One song takes twenty to thirty minutes. About one in ten gets analyzed a second time, independently, purely to keep the scoring consistent between analysts.
The early work was done on pen and paper and aggregated into a spreadsheet. Inside a year they had classified more than ten thousand songs by hand.
Meanwhile Westergren fundraised as a volume problem rather than a persuasion problem. The number he has given is 348 pitches, though it appears in print as 300-plus and 347 as well, and no independent tally exists. In March 2004, the last of them landed: Larry Marcus at Walden Venture Capital led a rescue round into a company Walden's own materials describe as being on the verge of bankruptcy. Roughly $2 million in deferred wages was paid out.
Then the second unglamorous decision. Westergren did not keep the CEO job. Joe Kennedy, a former Saturn executive, was brought in to run the company, and Westergren moved to Chief Strategy Officer.
The consumer relaunch came in August 2005, and it launched with the wrong business model: subscription only. Users promptly discovered they could exhaust the ten-hour free trial and simply sign up again with another email address.
The fix was to stop fighting that behaviour. In November 2005 Pandora added a free, ad-supported tier; the first ad sold that December, nearly six years after the company was founded. The paid tier settled at $36 a year.
What followed was a decade of fighting over royalties. In March 2007 the Copyright Royalty Board tripled per-play rates, a ruling Kennedy publicly described as leaving no path forward. A 2009 settlement pulled the rates back. Pandora sued ASCAP in 2012 and, in 2014, a federal judge held its rate at 1.85% of revenue.
Pandora won those fights and lost the war anyway. It went public in June 2011 at $16 a share, valuing it around $2.6 billion, with roughly 80 million users. Listeners peaked above 81 million in 2014. But the company was built as radio — you get a station, not a song — and Spotify was selling on-demand. Pandora did not ship its own on-demand tier until March 2017, years late, having bought Rdio's assets for $75 million in 2015 to make it possible.
Westergren returned as CEO in March 2016 and stepped down again in June 2017. In February 2019, SiriusXM acquired Pandora for approximately $3.5 billion.
Read the ending honestly and it is genuinely ambiguous, which is why it is worth reading.
Three and a half billion dollars is a real outcome, and for a company that was functionally insolvent in 2001 it is an extraordinary survival. But it is below where the company traded at its peak, its audience had been shrinking for roughly five years by the time of the sale, it never established sustained standalone profitability, and it was bought rather than winning. Westergren has said publicly that they squandered an enormous opportunity by not moving to on-demand quickly enough.
The durable lesson is not the persistence, which is unrepeatable and was partly funded by other people's forgone wages. It is the method. Before Pandora was software, it was Westergren doing by hand — for film directors, one at a time — exactly what the product eventually did at scale. He had a manual process that worked, he wrote down the rubric, and only then did he encode it.
The order matters. Most people build the software first and go looking for the process afterwards.
The framework: run the valuable process by hand for real customers until you can write down the rubric, and only then automate it.
Move 1 — Identify the judgment call you already make well and produce it manually for someone who pays you. Westergren was profiling musical taste for directors before it was a product. The modern equivalent is a service business you can start this month: audits, curation, matching, diagnosis. You are not looking for revenue yet, you are looking for the repeatable steps.
Move 2 — Write the rubric down as explicit, scored attributes rather than instinct. Pandora's version was hundreds of named dimensions on a 0-to-5 scale, with one in ten items double-scored by a second person to catch drift. Do the same on a spreadsheet, with a friend re-scoring a sample. If two people applying your rubric disagree constantly, you do not have a product yet — you have a preference.
Move 3 — When you raise, treat it as attrition and set a hard stop in advance. Westergren's several hundred pitches worked, but the cost was borne by employees who went unpaid for two years and then sued. Decide up front how many months of your own money and how many rejections you will spend, and honour that number rather than financing persistence with other people's wages.
Budget line: a manual version of almost any matching or curation service costs under $500 to start — a spreadsheet, a landing page, and a payment link — plus about ten hours a week for three months to build the rubric on real customers. If you cannot get one person to pay for the hand-done version, no amount of software will fix that.
The company's arc — the 2004 rescue round, the 2011 IPO, the 2019 sale — is documented in filings and contemporaneous press. The near-death details are not. The pitch count appears as 300+, 347 and 348 depending on the telling; the personal debt figure is reported as both $150k and $500k; the number of employee wage lawsuits is given as two, four or five. All of it traces back to Westergren's own recollection. The rescue round is reported as $8M and $9M by different outlets. Wikipedia's Music Genome Project entry credits Will Glaser with conceiving the project in late 1999, while most press credits Westergren — noted here rather than resolved.
Last verified 2026-09-01
Same structure every time — the tipping point, the first ninety days, and a playbook with real numbers attached.