Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456]
What this covers
Gokul Rajaram, who helped build the ad and product businesses at Google, Facebook, Square, and DoorDash and has invested in more than 700 companies, speaks alone here. The conversation splits between how AI changes product work and what he learned scaling those companies. On the AI side, he argues that non-deterministic software makes evals a core job for PMs and researchers, often by writing AI to judge AI. He predicts that light, simple products will be absorbed by foundation model companies, so startups should aim at complex workflows that need custom data. He also contrasts per-seat vendors like Zendesk with sturdier incumbents, using data half-life to explain why Slack's messages are more precarious than accumulated records.
The ad material is the most concrete. He says there are only three ways to win in advertising: own a first-party surface, deliver a specific outcome at a set cost as AppLovin does, or serve one giant advertiser's demand as The Trade Desk does. He treats ChatGPT as a rare first-party surface and names the biggest threat to platforms as users who delegate repeat purchases to an agentic interface and stop opening apps. Along the way come stories of Larry Page shrugging at AdSense's growth, Sergey Brin scrapping publisher pre-approval, and Facebook's custom audiences emerging from Zynga's hunt for whales.
The rest is operating craft, and some of it is contested. He covers North Star metrics paired with check metrics, Square's shift of risk from merchant to transaction, picking a vertical to win a lighthouse customer, and Figma displacing Sketch from the bottom up. He also argues that authenticity is the key trait in a founder and that 12-to-18-month job stints are a hiring red flag.
Gokul Rajaram argues that AI agents are collapsing the traditional product-building pipeline and cheapening software, so durable value now comes from human judgment, data that decays slowly, switching costs (migration, money flow, hardware), and outcome-based pricing rather than seats or features.
- Long-running agents make product roles bottoms-up and non-deterministic software makes evaluation the core PM job
- Seat-priced software and short-half-life data are most exposed, while timeless data and migration costs protect incumbents
- Ad businesses have only three viable models, and agentic interfaces threaten ad platforms by removing app opens
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Unlike deterministic software where input X always yields Y, AI products can produce different outputs for slight variations of the same input, so someone must evaluate outputs across use cases, and PMs and researchers must own evaluation, often by writing AI to evaluate AI.
“Today, you could do X, Y happens. But if you do slight variation of X, something completely different happens, non-deterministic software.”
Square was founded after a glassblower lost a sale to lack of card acceptance, and it accepted over 90 percent of merchants rather than rejecting them up front, then applied machine learning risk models to each transaction, shifting risk from the merchant level to the transaction level.
“they put risk at the transaction level. So they accepted you as a person, as a business.”
A good North Star metric tracks customer value and should not be revenue, such as Square's gross payment volume and Facebook's daily active users, and it must be paired with check metrics for customer health and company health to prevent gaming, as DoorDash's GMV could be inflated to zero by removing delivery fees.
“North Star metrics, in my opinion, should not be revenue.”
Once an ad-free surface starts showing ads, user engagement tends to decline over time as ads siphon engagement, so companies should keep holdout groups that never see ads and set an engagement budget capping the revenue-driven engagement loss, as Facebook did with News Feed.
“once you start showing ads in a previously unmonetized zero ads surface, engagement of users goes down over time.”
Facebook's custom audiences, now the foundation of most ad systems, came from Zynga's struggle to acquire high-spending whales, which led Zuckerberg to suggest uploading known whale users so Facebook could find lookalike users, a method later applied to other customers.
“Why can't they just upload their whales into our system?”
Good design means a product is so intuitive that customers need no manual, and Square's point of sale can be downloaded and used immediately, in contrast with traditional point-of-sale systems that require days of barista training.
“Good design doesn't mean visually pleasing. It means a product that is designed so well that you don't have to give your customers a manual on how to use it.”
Because customers will not switch without data migration, AI-native challengers must spend a year or two building migration tools; one of Gokul's companies used engineers in Eastern Europe for two years to build one, and Square needed months or years of scripts to move a point-of-sale business away from a cheaper rival.
“So literally, it's a two-year effort to build migration.”
Companies like Zendesk that price per seat, where each seat maps to a human doing a task, are the most endangered because an AI agent can sit beside the product and let customers cut seats gradually, such as going from 50 seats to 20 plus 30 agents, as a two-way door decision.
“Instead of paying for 50 Zendesk seats, you can pay for 20, and I can have 30 AI agents sitting next to Zendesk.”
Enterprise companies should target one or two specific verticals and win the best lighthouse customer there, because peers in an industry evaluate a product based on who else in their vertical uses it, while a Procter & Gamble win does not persuade JPMorgan.
“if you get JP Morgan to use your product, I promise you every single bank will then evaluate your product.”
ChatGPT combines intent data, which Google had, with identity data, which Facebook had, making it a rare first-party ad surface, and its multi-step queries provide richer targeting than single Google searches.
“ChatGPT, their combination of intent and identity data is unparalleled.”
Staying in jobs only 12 to 18 months is a major red flag for hiring managers because meaningful impact takes three to four years, and candidates who repeatedly hop may be rejected without knowing why.
“I think it takes minimum three to four years to have impact at a company.”
AI platforms must keep ads from influencing the content served or recommendations given, while holding a high bar for engagement, because ads reduce engagement over time as some engagement is siphoned off.
“ads should not influence the content that is served to me, or the recommendations that AI gives to me.”
Light, simple products are doomed because foundation model companies will absorb them, so startups should target complex, high-value workflows that require custom data.
“if you try to build a company that is light, that is not a hard problem, the foundation model companies are going to eat you.”
The biggest threat to major ad platforms is consumer behavior change where users automate repeat transactions through AI agents and stop opening apps, eroding ad opportunities and customer relationships, so platforms should watch early adopters who link accounts to ChatGPT and compare their app engagement.
“Consumer behavior change, where they don't open up the apps anymore, but they use agentic interfaces.”
There are only three ways to succeed in advertising: own a coveted group of users on a first-party surface like Google Search or Facebook, drive a specific outcome at a set cost like AppLovin's app installs, or become the exclusive provider for a large advertiser's demand like The Trade Desk with Procter & Gamble.
“There are three fundamental ways to succeed in the ads business. Three, and only three.”
Software companies are insulated when their utility is built on data that has been captured over time and stays valuable, while those whose data has a short half-life, like Slack messages, are more precarious.
“The more timeless the data is, the more protected they are.”
Sergey Brin argued AdSense should drop its upfront publisher approval system and instead review pages in real time after a URL reaches around 100 impressions, because pre-approval is slow and can be gamed, leading Google to scrap a system that half the engineering team had built.
“Let it load for a hundred times. And after a hundred impressions, if any URL hits a hundred impressions, then start reviewing it.”
To detect agentic disintermediation, companies should watch users who connect their accounts to ChatGPT apps and measure whether their app opening frequency falls, then decide whether to make the app less compelling or incentivize opening it.
“I'm going to look to see these people who have connected. How's their behavior on my app?”
Larry Page was disappointed by AdSense's strong growth because it was under 1 percent of all internet ads; he cared about serving every ad on the planet rather than a revenue figure of a billion or ten billion dollars.
“What percentage of all ads on the internet are you? Less than 1%.”
Products that spread through individual users can infiltrate and displace incumbents in ways a top-down direct sales motion cannot, as Figma displaced Sketch after a mid-level design manager adopted it, and Cursor is used across large companies through engineers rather than top-down deals.
“you can infiltrate and be an insurgent in a unique and powerful way, which direct sales motion could never produce”
The most important thing to assess in a founder is authenticity, meaning a founding story rooted in a lived experience and genuine curiosity, because founders who start companies just to start with a friend lack the compelling reason to persist.
“The most important thing I look for is founder authenticity.”
When Larry Page saw internal customer systems used for large accounts, he ordered every feature made available to small customers, and self-serve users adopted faster and proved the most sophisticated, revealing product capability that sales-led models miss.
“End it right now. I want to make sure that everything you're building for large customers is also available to small customers.”
Opening a product to self-serve users reveals its real capabilities, since sophisticated small agencies and hustlers exploit features in ways internal sales teams never anticipate, and it forces companies to make onboarding excellent.
“when you open up your system to self-serve, you learn so much more about the capability of your product”
AppLovin reached over $100 billion by delivering mobile app installs at a set cost, and it now controls the buy side, sell side, and middleware, so it effectively controls the auction for most mobile apps.
“AppLovin is a $100 plus billion company. They drive one outcome really well. Mobile app installs.”
Eric Schmidt required Google's annual strategy to be presented using only images, arguing that people do not remember words but remember how things made them feel, so presenters should find a compelling image for each business such as a hockey-stick graph for YouTube uploads.
“People don't remember words. They remember how things made them feel.”
Seat-priced software companies must move to outcome-based pricing, for example from $20 to $30 per seat to $0.20 to $1 per ticket resolved, a transition so difficult that many may need to go private to make it.
“That's why I think many of them probably need to go private because they have to make this business model transformation in private.”
Because model capabilities change too fast to prescribe a product in advance, companies have moved to bottoms-up building, where PMs only articulate customer needs at a high level and act as guardians of the why, while engineers, researchers, PMs and designers build the product together in code.
“So almost everybody has gone to a bottoms-up approach, where it's not done by product management anymore.”
A Fortune 500 CIO told Gokul he would skip startups and use Gemini's agent builder and ChatGPT Enterprise's agent builder with a thousand retrained IT engineers, so any startup whose product a customer can build with horizontal agent tools will fail.
“Why do you need any startups?”
Building an ad business as a middleman on top of Google or Facebook fails because those platforms learn from every new capability built on them and absorb it, so ad companies must stay outside their ecosystems, and optimization tools for ChatGPT ads or answer engines will not create durable income.
“Every time you build a new capability on top of Google, turns out Google learns what you're building.”
Being first to build an AI ad network does not matter much, because companies with first-party inventory cannot lose it, and later entrants can learn from earlier iterations, and Gemini can wait to monetize while positioning itself as a zero-ads product.
“being first doesn't matter because especially if you're in category one, which we described, you control your first party inventory.”
In the AI era, companies should avoid hiring middle managers and instead hire builders who manage AI agents, and a span of control below 10 direct reports should not be allowed because managers with 15 reports meeting weekly cannot fill the rest of their time productively.
“a span of control of less than 10 should not be allowed at any company at this point.”
The most relevant skill within one to two years is becoming a functional expert who builds AI agents for that function and orchestrates an army of agents, as a non-technical Meta PM did by building agents that made engineers ask to learn from him.
“the number one skill that is going to be relevant two years from now, probably even one year from now”
Every vertical has a system of record where data lives, and agent companies can no longer just sit on top of it, so they must ambitiously build and offer a full system of record themselves.
“So every company I know is now trying to figure out how do I build the entire platform, and not just a system that does some workflows.”
Public software markets are not distinguishing between companies whose value is data that decays slowly and companies whose value can be siphoned off by a next-door AI agent, so their low multiples reflect a blunt category discount.
“the software public markets are not distinguishing between these two types of companies.”
NetSuite and similar ERP systems are protected because replacing a system that runs the whole business is career-limiting for the buyer, even if AI-enabled ERP competitors exist, and NetSuite can also bundle AI agents using its own data.
“there is no compelling reason for someone to put their career at stake by ripping out NetSuite.”
A weekly CEO email should have three sections, top of mind, performance update, and miscellaneous, with top of mind taking 60 to 70 percent of effort, and content organized across product, business, and team dimensions, with repetition used to make messages stick.
“But the most important section where you should spend 60% or 70% of your time on, is top of mind.”
Judgment is the one thing likely to remain durable in AI because when engineers can produce unlimited code, the scarce skill is deciding which code and which products are worth building, and humans must review critical code and assess design fit.
“The one thing I think that's going to be truly future-proof is judgment.”
No feature should ship without a clear hypothesis stated as a customer behavior change, such as moving customers from doing X to doing Y or from X minutes a month to Y minutes, grounded in data or a non-obvious customer insight, and CEOs hiring a first PM should ask why.
“you should not let any feature go out if there's not a clear hypothesis behind this feature.”
To be durable in the AI era a company needs ownership of a scarce asset such as a license or regulatory position, a control point over money or data, hard-to-replace hardware, an essential workflow, or network effects, and ideally builds these into the workflow from the start.
“you need to have ownership of a scarce asset.”
The best hiring method is a realistic work project performed without AI, such as Square Corp Dev asking for one company to acquire with synergy analysis, because other functions allow candidates to bluff without doing real work.
“Every function needs to have a work project that you need to put them in a room without AI and get them to do the project”
Companies should never invite someone to the board before spending at least a year with them, first using an advisory board and multiple board meetings, because a board role is like a marriage.
“a board role is like a marriage.”
Each board member should be a buddy to a management team member, meeting monthly or texting between meetings, because the in-between interactions are where things actually get done rather than in formal board meetings.
“It's not the board meeting that truly matters. It's all the things between the board meetings that are the real thing when things get done.”
In enterprise sales, founders should stop leading with what the product does and instead present an outcome they can deliver, as Palantir does by promising to solve a customer's key problem in six months with payment only if successful.
“you cannot lead with what your product does anymore. You've got to lead with what is the outcome you can deliver”
Gokul failed to push Figma top-down into a design team that preferred Sketch, but two years later a mid-level design manager brought Figma from a prior company, and it spread across the team and displaced Sketch, showing that insurgent bottoms-up adoption can beat incumbents.
“they brought in Figma from their prior company and they got it to be used across, and it kicked out Sketch.”
The best designers and product people minimize steps and edit down features, asking what two things drive the customer outcome, because a PM's role is editing rather than adding features, as Jack Dorsey called the PM role product editor.
“the role of the product manager is not to add more features.”
Public software markets are pricing both seat-based and data-durable software companies the same, although the two types face very different levels of disruption risk.
“the software public markets are not distinguishing between these two types of companies.”
Self-serve products force companies to design onboarding and reach moments of delight quickly, and with word of mouth can reach millions of customers, whereas 100 salespeople might reach around 10,000, as shown by Cursor's engineer-led spread.
“with a self-serve product, with the right word of mouth, you can reach millions of customers.”
Teams given a choice between an extra designer and an extra engineer now pick the engineer because design systems already exist and AI can generate designs from them, so the ratio of designers and PMs to engineers has moved from roughly 1:3 or 1:10 to about 1:20.
“The number of designers and product managers relative to the number of engineers, when I was growing up in product, it used to be 1:3 or 1:10. It's going to 1:20 now.”
The best product people define outcomes as customer behavior changes, because the simplest product job is moving users from non-customer to customer, to loyal customer, to paying customer, while poor products move loyal customers to churn, and customer behavior is a leading indicator of business outcomes.
“Because customer behaviors are leading indicators for every business outcome.”
A good board needs product or technology experience, which 10 to 15 years ago was rare but is now present on most tech boards, and it needs a voice of the customer, as Square gained through Shake Shack's CEO Randy Garutti.
“Now, every good board has one or two product or tech people.”
Starting in 2024, systems of record like Slack began restricting API access to AI agent companies that treat them as dumb databases, with Slack cutting off Glean, and other verticals either blocking APIs, bundling free agents, or charging per call such as $2 per API call.
“Slack has done it most publicly. Slack is owned by Salesforce. They cut off access to Glean”
Founders building AI applications should start with deep, compelling problems by finding industries where highly paid people do somewhat repetitive work that agentic software can now perform, because software has shifted from a tool humans use to something that can do the job of people.
“what industry are there roles of people that are highly paid, that are doing somewhat of a repetitive job, and that can be done by software?”
Over December 2025 and January 2026 it became clear that product development has fundamentally changed because of the notion of a long-horizon, long-running agent that is resilient to failure and usable by non-technical people.
“it's become very clear that something has fundamentally changed. And what that thing is, is the notion of a long horizon, a long-running agent.”
A product manager's job is to balance customer needs and business needs, acting as the keeper of why a product is built, since building something customers love without business value or raising prices at customers' expense both fail.
“balancing customer needs and business needs at the highest level is what I think of the product.”
Companies should communicate along three dimensions (product, business, team) and repeat messages many times, because repetition once, twice, thrice or four times is what makes communication seep into people.
“don't be afraid of repetition.”
For gaming companies and casinos, roughly 80 percent of all revenue comes from whales, which is why game companies focus on acquiring them.
“80% of all revenue for any gaming company comes from whales.”
Gmail, internally codenamed Caribou and launched April 1, 2004, offered one gigabyte of storage, roughly 100 times the 10 megabytes of Yahoo Mail, which exemplified Larry and Sergey's technology-first philosophy.
“Yahoo Mail was the dominant product, and it gave 10 megabytes of storage. So this thing had 100x more storage.”
Square was founded after Jim McKelvey lost a sale because he could not accept credit cards, and Jack Dorsey and McKelvey realized the iPhone audio jack could carry a card reader, letting small businesses accept payments.
“the iPhone, which had just been released a couple of years ago, had this thing called the audio jack”
Tony Xu gave candidates $10 or $20 and asked them to acquire a thousand DoorDash customers, with few succeeding in reaching even a hundred, and the test measured how many approaches people tried within hours and who opted out.
“He would give people either $10 or $20 and ask them to acquire a thousand customers.”
Gratitude for health, family, and the random lottery of circumstances, which places one in the top sliver of global circumstances, creates a responsibility to lead the best possible life and to feel pain when others suffer.
“in almost every run of the simulation, I could be in a million different worse circumstances than I am today.”
Gokul's sense of purpose comes from recognizing how fortunate he is to be healthy and loved, which places him in the top 1 percent of the 1 percent of the 1 percent of situations, and he feels a responsibility to lead the best life he can.
“we are sitting in literally the top 1% of the 1% of the 1% of situations right now.”
Google's investments in Street View, TPUs, and Waymo reflect over ten years of spending on uncertain futures, and Gokul argues that betting on technology capability eventually pays off, though it took a decade.
“knowing that if you invest in technology, good things are going to happen, and good things happened, but it took a decade.”
Mark Zuckerberg is one of the best at designing consumer products for engagement and growth, critiquing product flows directly, and learning new domains by shadowing teams until he generated ideas, as he did with ads within about a year.
“within a year, he got to the point where he was generating ideas for the ads team.”
Salesforce customer records, Zendesk complaints, and Jira product work should be linked, creating value for unifying platforms, but no customer will move until a seamless migration exists, making it a multi-year effort that otherwise requires an outside integrator like Accenture.
“So all these three systems should be linked together, but they're all three different companies.”
Bob McDonald's decision to hire Gokul despite his lacking product management experience and visa constraints was the kindest act, and it shaped Gokul's practice of paying forward favors with no expectation of return.
“he bet on me.”
The strongest product candidates question the premise of a project, as one candidate did by interviewing 10 Square users and showing that none wanted the Premium Insights product, which signals the agency and voice-of-customer judgment Gokul values.
“You want people to reject the premise or question the premise in the first place.”
Companies that combine financial services with software, such as business banking products where customer money flows through the platform, are sticky because switching requires untangling regulations and embedded processes.
“So I like things that are a combination of financial services and software because of that.”
The best founders have studied alternative and historical approaches to their problem, and Gokul tests this by offering five or six other solutions to see whether founders can justify their chosen path, citing the Collisons' study of payments company failures and successes.
“I think the best founders are students of history in that industry.”
For consumer companies, influencers now drive discovery, and TikTok works as a local search engine for younger people who find restaurants that Google Maps and Yelp miss, which is pushing a set of companies to scale influencer marketing by capturing viral long-tail waves.
“Somebody said that TikTok is the best local search engine.”
Leaders should be more candid with employees about what is top of mind, because asking people for their input on hard decisions surfaces better ideas, especially in small companies.
“I personally think more candid is better than less.”
PMs are beginning to check code into production repositories using Codex or Claude Code, and companies now include an explicit prototyping interview in hiring loops to force PMs to be hands-on, with AI tools expected to review code before engineers commit.
“Now, there's an explicit interview, in the interview loop called prototyping. It literally forces product managers to be hands-on.”
The AI capability frontier is being pushed out roughly every two months, so a product person's model of what is possible goes stale quickly and they fall behind if they keep reasoning from older assumptions.
“the capability frontier is being pushed out every two months”