In the weeks after 23andMe filed for bankruptcy in March 2025, about 1.9 million customers, roughly 15% of its base, asked the company to delete their genetic data while that same data was being offered to buyers. Each deletion shrank the asset on the auction block, which turns the usually abstract argument about how trust changes cash flow, risk, and valuation into something you can literally count in database rows. Most of the time the connection is harder to see, because it hides inside churn curves, discount approvals and an investor's sense of how risky a company feels. This post drags it into the open with a small Python model, three companies that each turned a different trust dial the wrong way, and six signals worth measuring before the next incident measures them for you.
Three dials, one function
Strip a valuation down to its skeleton and trust touches it in three places. The first is retention: how many customers are still paying next year, and the year after that. The second is price realization: how much of your list price survives the credits, discounts and "just this once" concessions that sales hands out when buyers get nervous. The third is the discount rate: the haircut investors apply to future cash because they are not sure it will arrive.
None of these is a soft metric. They are the same inputs a finance team already models, and the only new idea is admitting that trust is one of the forces moving them. A company people trust keeps customers longer, collects closer to list and gets its future cash valued more generously. When that trust breaks, all three dials turn against it at once, which is exactly what spreadsheets built one assumption at a time tend to miss.
The model
Here is the smallest version that still tells the truth: the present value of the gross profit a single customer produces over ten years. The inputs are illustrative (a $1,200-a-year subscription, 75% gross margin, 90% annual retention, a 10% discount rate), so swap in your own numbers before quoting the output to anyone.
def customer_value(retention=0.90, price=1.00, rate=0.10, arpu=1_200, margin=0.75, years=10): """Present value of the gross profit one customer is expected to produce.""" still_here, value = 1.0, 0.0 for year in range(years): value += arpu * price * margin * still_here / (1 + rate) ** (year + 1) keep = retention(year) if callable(retention) else retention still_here *= keep return value def spike(year): """Churn doubles for two years after an incident, then fully recovers.""" return 0.80 if year < 2 else 0.90 baseline = customer_value() scenarios = { "churn +5 pts, for good": dict(retention=0.85), "churn spike, then recovery": dict(retention=spike), "5% price concessions": dict(price=0.95), "risk premium +2 pts": dict(rate=0.12), "spike + concessions + risk": dict(retention=spike, price=0.95, rate=0.12), } print(f"{'baseline':<28}${baseline:,.0f}") for name, inputs in scenarios.items(): value = customer_value(**inputs) print(f"{name:<28}${value:,.0f} ({value / baseline - 1:+.1%})") Running it prints:
baseline $3,895 churn +5 pts, for good $3,327 (-14.6%) churn spike, then recovery $3,315 (-14.9%) 5% price concessions $3,700 (-5.0%) risk premium +2 pts $3,632 (-6.8%) spike + concessions + risk $2,947 (-24.3%) A few things stand out. The dial that gets the most airtime in meetings, price, is the most predictable one: a 5% concession costs exactly 5%, because nothing about it compounds. Two points of extra risk premium, which nobody on a product team will ever see on a dashboard, costs more than that. Churn is where the real damage lives, at nearly three times the cost of the discount.
The line that deserves a second look is the spike. Over a ten-year horizon, a two-year churn spike that fully recovers costs about as much as a permanent five-point increase in churn. Stretch the horizon to twenty years and the permanent case pulls ahead, but the spike never earns its money back, because the customers who left during the bad years are not around when trust returns. Fixing the problem stops the bleeding. It does not refund the blood.
The bottom row is the realistic one, since an incident rarely touches a single dial. None of its assumptions looks dramatic on its own, yet together they take almost a quarter of the value, which is precisely why this kind of loss gets waved through review.
23andMe: retention with no way back
The breach that started 23andMe's slide was not sophisticated. Attackers used credential stuffing, logging in with passwords leaked from other sites, to get into about 14,000 accounts. A product feature did the rest: through DNA Relatives, which let users share information with genetic matches, those logins exposed data on 6.9 million people. That is roughly 500 exposed people for every account actually compromised, and the ratio was set by a design decision long before anyone attacked it. The response made things worse. In a letter to lawyers for a group of victims, the company argued that users had "negligently recycled" their passwords, and mandatory two-factor authentication arrived only after the data was gone.
From there the dials moved the way the model predicts, only faster. 23andMe filed for Chapter 11 in March 2025, the deletion wave followed, and the company itself warned that its business could suffer if the trend continued. In the end a bankruptcy judge approved the sale of 23andMe to a nonprofit led by its co-founder for $305 million, a sliver of what public markets once paid for it. In model terms, a deletion is churn with no recovery path: the row is gone, and so is every future cash flow it could have supported.
Redis: when a competitor sets your price
On March 20, 2024, Redis moved its core from the permissive BSD license to source-available terms, mainly to stop large cloud providers from profiting from it without contributing back. Eight days later the Linux Foundation announced Valkey, a fork backed by AWS, Google Cloud, Oracle, Ericsson and Snap. Then AWS made a move that belongs in every pricing deck: in October 2024 it priced ElastiCache Serverless for Valkey 33% below the other engines it supports and node-based Valkey 20% below, and let customers carry their reserved-node discounts across. Trust that Redis spent with its community was converted, within seven months, into a competitor's discount.
Redis has since walked much of it back. Its creator Salvatore Sanfilippo rejoined the company, and Redis 8 shipped in May 2025 with AGPLv3 as an open source option. CEO Rowan Trollope has said the license change had no effect on the business and that growth hit records afterward, and on the income statement that may well be true. Look where the cost went instead. By May 2026, AWS said tens of thousands of ElastiCache customers had moved hundreds of thousands of clusters to Valkey. In October 2026, Percona announced Valkey-proxy, now an official Valkey project, so that apps built for a single instance can run on a Valkey cluster without rewrites, with migration from Redis as an explicit goal. A whole ecosystem is spending engineering hours on making it cheaper to leave. That is what a broken promise does to a developer platform: it may never show up as next quarter's churn, but it does show up as other people dismantling your switching costs for free.
Okta: when investors discount your own numbers
Okta sells identity, which makes its product trust with an API. In late September 2023, an attacker used a stolen credential to get into its customer support system. BeyondTrust, a customer, caught an attack on its own Okta admin account on October 2 and reported it, but Okta did not confirm the breach to BeyondTrust until October 19. Cloudflare said it had spotted the activity more than a day before Okta notified it. When Okta went public, its first estimate was that around 1% of customers, 134 organizations, were affected. Shares closed down 11.5% that Friday, and by Monday's close more than $2 billion in market value was gone.
Then the story changed. On November 29, Okta said the attacker had downloaded a report containing the names and email addresses of every user of its customer support system, a much larger breach than the company had first described. The stock slipped again that day, hours before Okta reported a quarter that beat expectations. The second drop was not about a weaker business. It was about learning that the company's first estimate had been wrong, and that is the discount-rate dial in its purest form: a business that still looked healthy, valued with less confidence.
Measure trust the way you measure latency
You can't put "trust" on a dashboard directly, but you can track the things that move the three dials, and most of the raw data already sits in your logs, your CRM or your support queue.
- Restatement rate. How often your first public statement about an incident, from the number of affected users to the root cause, later has to be revised. Okta's second drop was a restatement. Review it every quarter, next to change failure rate.
- Time to acknowledge. The gap between the first credible outside report, whether from a customer, a researcher or a bug bounty submission, and your confirmation. Okta's 17 days with BeyondTrust is the number to beat.
- Blast-radius ratio. For any feature that shares data between accounts, how many other people's records one compromised login can reach. At 23andMe it was roughly 500 to 1. Compute it in staging before an attacker computes it in production.
- Exit intent. Weekly counts of data exports, account deletions, API key revocations, downgrades and "how do we migrate off" tickets, with an alert on spikes. This is the earliest view of the retention dial you will get.
- Discount drift. The average concession on renewals and new deals, quarter over quarter. Nervous buyers ask for a better price before they ask for a cancellation link, so this number moves first.
- Alternative gravity. For developer tools and open source projects, the stars, downloads and contributor counts of forks and compatibility layers relative to your own. Fork activity is public data; the only question is whether anyone on your team is looking at it.
Restatements and slow acknowledgments feed the discount rate. Exit intent and blast radius feed retention. Discount drift and alternative gravity feed price. Put the six on one page and you have something most companies never build: an early-warning system for the inputs their valuation actually depends on.
Pass trust explicitly
Every team already passes trust into its decisions. Most pass it implicitly, like a default argument nobody remembers setting, which is how a surprise pricing change, a vague incident update or a data-sharing feature with a huge blast radius gets through review without anyone asking what it does to the three dials. The fix is not a new department or a values statement. It is a habit: before you ship, name the dial a decision touches and the direction it pushes, then run the numbers with your own inputs. The model above is 30 lines long, and every company in this post would have paid a fortune to see its output a year earlier.