Why Did Google Shut Down Its Internal Prediction Market?

Google operated two internal prediction markets and shut both of them down — Prophit in 2011 after roughly six years, and its 2020 successor after a similar arc. The cause was not a single failure but a recurring pattern: thin liquidity, weak participation incentives, and management ambivalence about whether crowd forecasts should override executive decisions. The Google story is one of the cleanest case studies for operators building prediction market products today, because every failure mode it exposed is still unsolved.

The Two Google Markets

Prophit (2005-2011). Launched in April 2005 as one of the earliest large-scale corporate prediction markets, Prophit allowed Google employees to bet play-money on internal questions: would a product ship on time, would a competitor launch first, would a metric hit a target. It attempted a pivot midway through its life and was ultimately shut down in 2011.

Gleangen (2020). In April 2020, almost exactly fifteen years after Prophit launched and one month after Google sent 150,000 employees home for COVID, the company quietly launched a second internal prediction market. Like Prophit, it survived for a few years before being deprioritized.

Why Both Markets Eventually Died

The post-mortems converge on three structural problems that any operator building prediction markets — internal corporate or external consumer — still has to solve.

1. Liquidity Was Always Thin

Internal prediction markets at Google never developed deep order books. Most questions had a handful of traders, prices moved on individual bets, and the resulting forecasts were noisy. This is the same cold-start problem that B2B operators hit when launching event-contract products: without market makers and a liquidity strategy, the prices generated are too unreliable to act on, which kills usage, which kills liquidity. The loop is brutal.

2. Incentives Were Misaligned

Prophit used virtual currency redeemable for small prizes. Gleangen experimented with similar structures. Neither generated stakes meaningful enough to draw participation from the people whose information would have made the markets accurate. As the Asterisk write-up noted, an engineer's salary and stock vesting will almost always be more valuable than any marginal benefit from trading on an internal market — and the engineer with the most valuable inside information may have a direct disincentive to express it publicly.

3. Management Did Not Want the Answer

The most quietly damaging issue was cultural. Prediction markets work by aggregating decentralized information, which is sometimes antithetical to the management instinct to control which forecasts circulate. When a market predicted a launch would slip and leadership had publicly committed to the original date, the market became inconvenient rather than useful. Over time, inconvenient tools get deprioritized.

What Operators Should Take From the Google Case

For B2B operators evaluating prediction market products today, the Google experience is instructive on every front:

• Liquidity is a product feature, not a side effect. Any operator launching event contracts needs an explicit market-making strategy on day one — typically an LMSR-style automated market maker or contracted liquidity providers. Hoping organic liquidity emerges is what killed Prophit.

• Real money changes everything. Kalshi and Polymarket succeeded where Google failed in part because the stakes are real. Play-money markets are research instruments; real-money markets are products. Operators building prediction surfaces inside existing iGaming platforms have a structural advantage here — the players already arrive funded.

• Pick categories where the operator wants the answer. A prediction market is most valuable when the host genuinely wants accurate forecasts. For iGaming operators, this maps to categories where existing odds-making is expensive and slow: niche sports, novelty events, entertainment outcomes. It does not map to categories where the operator has a strategic preference for a particular outcome.

• Cold-start with real questions and seeded liquidity. Every successful prediction market — corporate or consumer — has launched with a curated list of questions and seeded order books. Google launched with questions selected by employees, which produced uneven engagement. Operators should treat question curation as editorial work, not crowd-sourced.

The Quiet Comeback

The Google case also offers a hopeful note for operators. The fact that Google launched a second prediction market fifteen years after the first one died — and that Kalshi, Polymarket, and a wave of new venues are scaling now — suggests the underlying idea was never wrong. The execution has gotten better: real money, regulated venues, automated market makers, and crucially, audiences who already understand probabilistic thinking. The operators building the next wave can learn from Prophit's failures without repeating them.

---

Last verified: May 2026