What Happened to Google's Internal Prediction Market?
Google ran two distinct internal prediction markets: Prophit (2005–2011), the first large-scale corporate market in tech, and Gleangen (2020–present), a successor launched during the pandemic. Both produced accurate forecasts, but Prophit was shut down because organisational politics — not forecast quality — limited its operational value. The case is the most-studied example of why corporate prediction markets succeed technically and fail culturally.
Google has run two internal prediction markets. The first, Prophit, operated from 2005 to 2011, produced demonstrably accurate forecasts on product launches and operational metrics, and was shut down primarily because surfacing accurate cross-divisional information conflicted with managerial control of internal narratives. A successor market, launched in April 2020 during the pandemic, revived the format with a stronger focus on operational forecasts and explicit safety metrics. The Google case is the most-cited example of why corporate prediction markets typically work as forecasting engines but struggle as governance tools.
For B2B operators evaluating event-based intelligence — and for anyone integrating prediction-market data into a betting platform or research workflow — the Google story is the canonical lesson on the difference between forecast accuracy and organisational adoption.
Prophit (2005–2011): The First Large-Scale Corporate Market
Prophit launched in April 2005 as a play-money internal market open to Google employees. It was studied by Bo Cowgill (then at Google, later Columbia Business School) and academic collaborators including Justin Wolfers and Eric Zitzewitz, producing the most widely cited dataset on corporate prediction markets ever published.
Three findings from Prophit defined the field:
- Forecasts were accurate. Across launches, demand questions, and external events, Prophit prices were close to calibrated probabilities once optimism and event-novelty biases were controlled for.
- Information flowed where employees already talked. Traders sitting near each other, or sharing managers, traded on similar information — meaning markets revealed existing communication networks rather than aggregating wholly independent signals.
- Optimism bias was structural. Markets systematically priced positive Google outcomes higher than they should have, especially right after positive press cycles.
Prophit attempted a pivot late in its life and ultimately shut down in 2011. Internal accounts and the academic literature agree that the failure was not technical: it was organisational. Surfacing accurate cross-team forecasts cut against managers' interest in controlling what other parts of Google knew about their roadmap.
Gleangen (2020–present): The Second Attempt
In April 2020, almost exactly fifteen years after Prophit launched, Google launched its second internal prediction market. The relaunch was framed around operational and safety forecasting rather than launch-date games — including, notably, Waymo safety-metric questions that crossed division boundaries.
Two design changes mattered:
- Operational focus. Questions tilted toward measurable internal KPIs rather than novelty product bets, reducing the "fun but not actionable" critique levelled at Prophit.
- Explicit cross-divisional value. Some Waymo executives reportedly saw the safety-metric forecasts as a way to communicate consistent numbers across divisions — directly attacking the political problem that killed Prophit.
The same political tension still appears, however: at least one Waymo VP is documented as wanting to restrict the very metrics the market made visible, on the grounds that controlling who knew what was a core management lever.
Why Prediction Markets Die at Big Companies
The Google case generalises. Internal prediction markets at Microsoft, HP, Ford, and "Firm X" (anonymised in the Berkeley study) followed similar arcs:
| Failure pattern | Mechanism |
|---|---|
| Political resistance | Accurate forecasts undermine managerial narrative control |
| Manipulation risk | Insiders trade on information they should be reporting upward |
| Question design fatigue | Markets need a steady stream of well-specified, resolvable questions |
| Resolution friction | Settlement requires a trusted oracle; ambiguity erodes participation |
| Regulatory drag | Real-money markets cross legal lines; play-money loses incentive sharpness |
Prophit ran into the first, third, and fourth of these. Gleangen has been designed to mitigate all five.
What This Means for B2B Operators
For betting operators and intelligence-layer vendors, the Google case carries three operational lessons:
- Forecast accuracy is the easy part. Modern prediction-market infrastructure (LMSR pricing, AMM order books, automated resolution) is well understood. What's hard is institutional adoption.
- Cross-product intelligence has the same political dynamics inside operators. Surfacing accurate behavioural forecasts across sportsbook, casino, and prediction-market verticals creates the same "who controls the narrative?" tension Google saw between Search, YouTube, and Waymo.
- External prediction markets are now doing what internal ones could not. Polymarket, Kalshi, and ADI Predictstreet provide liquid, public, real-money probabilities that operators can ingest as data — bypassing the political problem of running an internal market entirely.
The Strategic Read
Google's prediction-market history is not a story of failed technology. It is a story of accurate forecasts running into the political reality of a large organisation. The shift from Prophit to Gleangen — and the parallel rise of external markets like Polymarket and Kalshi — suggests the next decade of prediction-market value sits in external event-data feeds that operators can consume, rather than in internal markets they have to socially manage.
Adkuu integrates external prediction-market signals into the same intelligence layer operators use for player segmentation, cross-product personalization, and bet-slip recommendations — turning event probabilities into operational inputs rather than a political project.
Last verified: April 2026