Every concept is a perspective. Which one are you looking for?

In practice

A growth rate without a period is meaningless; 10% monthly and 10% yearly are different businesses. High rates are easy on a small base, so always read them next to the absolute numbers.

In practice

Unlike a funnel, a loop feeds on its own output. Start by finding the weakest link, because a loop turns only as fast as its slowest step.

In practice

Hypergrowth is an outcome, not a goal, and it has a cost: if hiring, infrastructure and culture do not scale at the same pace, growth turns brittle. Check that unit economics hold before you chase speed.

In practice

A flywheel turns through many consistent pushes in the same direction, not one big move. Write down the wheel and its links first; spotting the link that feeds nothing else often clarifies the whole strategy.

In practice

A good model reveals which variable moves the outcome most. Do not model everything; start with three or four levers and compare it to reality every month.

In practice

The number of tools says nothing about maturity. The stack works when data lands in one place and everyone reads the same number; before adding a tool, check whether the ones you have are actually used.

In practice

A key result should be an outcome, not an activity: not 'run three campaigns' but 'raise trial-to-paid conversion'. If the target is easy to hit, it is not ambitious enough.

In practice

Every experiment starts with a hypothesis, a success metric and an end date. A failed test is still learning; the real mistake is not recording the result and running the same test again.

In practice

Judge the decision, not just the outcome: a good decision can still produce a bad result. If the lesson does not make it into next period's plan, the retrospective is just nostalgia.

In practice

Velocity can rise while the rate falls; that is normal on a larger base. The same holds for the team: how many experiments you run per week sets your learning speed.

In practice

The ceiling is usually the channel's, not the market's: once a channel saturates, costs rise and returns fall. When growth slows, first separate channel saturation from a product problem.

In practice

People consistently underestimate compounding: 10% a month is roughly 3x over a year. Small, steady improvements usually beat large one-off moves.

In practice

Levers are not equal: measure first, then rank them to find the big ones. Price often moves results more than cutting acquisition cost, yet it is tested far less.

In practice

A multiplier is only as good as what it multiplies: anything times zero is zero. Build something that works first, then add the multiplier.

In practice

Growth ops looks boring but sets your experiment speed. A team without a tracking plan, naming standards and a single source of truth ends up arguing about numbers instead of running tests.

In practice

The total can grow while losses quietly accelerate underneath; this method exposes that. Do not set strategy before you know whether growth comes from new customers or from existing ones staying.

In practice

Scale advantages do not arrive automatically; complexity grows alongside them. Do not assume what worked when small will work at scale; retest at each growth threshold.

In practice

Too many indicators work like none at all. One north star and three to five driver metrics are enough; every metric needs an owner and a target.

In practice

Revenue is a poor north star because it measures the outcome, not the value customers receive. A good one counts the moment value is delivered, such as 'users who do X at least once a week'.

In practice

There is no single number, but the signs are readable: strong retention, unprompted word of mouth, and customers who would genuinely mind if the product disappeared. Scaling before fit is pouring water into a leaky bucket.

In practice

A low-powered experiment does not only miss effects; it also makes significant results less trustworthy. Kohavi's table: when 10 percent of experiments truly win and power drops from 80 to 20 percent, 52.9 percent of significant results are false positives. For a low-traffic product the lesson is simple: fewer, bigger experiments rather than more (our reading). The Experiment Capacity tool runs this calculation for your own traffic.

In practice

The real question is how you define 'active': logging in, or doing something that delivers value? The DAU ÷ MAU ratio shows whether the product creates a habit.

In practice

A common mistake is counting one-off items, or annual prepayments not spread across months, in MRR. Look at the components of the change (new, expansion, churn), not only the total.

In practice

One blended conversion rate misleads; split it by traffic source, device and new versus returning. Raising the rate is easy; converting the right people is hard.

In practice

Counting only ad spend makes CAC look lower than it is; include salaries and tools. Calculate it per channel, because a blended CAC hides the bad ones.

In practice

An average hides the extremes: a few large customers can inflate ARPU. Look by segment, and state clearly whether free users are included.

In practice

LTV is an estimate, not a measurement; a small change in churn moves it a lot. Use gross margin, not revenue, and trust it only for cohorts with real retention data.

In practice

An NRR above 100% means revenue grows even without new customers, one of the strongest signals in subscription businesses. Track churn separately by customer count and by revenue, because they can tell very different stories.

In practice

ROAS is measured on revenue, not profit; a high ROAS can still lose money on a low-margin product. Work out your break-even ROAS from gross margin and set the target above it.

In practice

The open-ended 'why' that accompanies the score is worth more than the score itself. NPS gives direction but not a diagnosis; if it drops, look for the cause in the comments.

In practice

Read retention from the cohort curve, not the average: if the curve flattens somewhere, the product has a loyal core. A curve that keeps falling means fixing the product, not spending more on acquisition.

In practice

In a cash-constrained company it is a more concrete signal than LTV, because it tells you when the money comes back. As a rough rule of thumb, around 12 months or less is often considered healthy for subscription businesses, but it varies by market and payment model.

In practice

A commonly cited rule of thumb is 3:1, but treat it as a starting point rather than a hard threshold. A very high ratio is a signal too: you may be underinvesting in growth.

In practice

Define activation first: not signup, but the action where value is first felt. The cheapest way to improve retention is often to improve activation.

In practice

The definition depends on context: likes and comments on social, frequency of the core action in a product. Separate shallow engagement (clicks) from engagement that creates value (completed tasks).

In practice

It varies widely by industry; compare only against similar business models and your own history. A rising RPE can mean efficiency or underinvestment.

In practice

Because it drives gross margin, classifying COGS correctly directly affects pricing decisions. In software companies hidden COGS usually sits in support and infrastructure costs.

In practice

Raising AOV through bundles, free-shipping thresholds or cross-selling is often cheaper than finding new customers. The average can mislead; look at the distribution and the median too.

In practice

Shipping a feature does not mean it was adopted. Track first use and repeat use separately; when adoption is low, the issue is often discoverability and messaging rather than the product.

In practice

Revenue-based churn shows the impact of losing a large customer, which a customer count hides. Separate voluntary cancellations from involuntary ones caused by failed payments, because the fixes differ.

In practice

Expansion is the most profitable way to offset losses, because it carries no acquisition cost. Check whether your pricing lets revenue grow as the customer grows.

In practice

When K goes above 1, growth feeds itself, but that is rare in practice. For most products the realistic goal is not K=1 but partial virality that lowers paid acquisition cost. The length of the invite cycle matters as much as K.

In practice

A blended retention number mixes old and new customers and hides improvement or decline. Cohorts are the most honest way to see whether a change you made actually made a difference.

In practice

How you define the market denominator determines this ratio; define it too broadly and penetration always looks low. As penetration rises, growth shifts from new customers toward deepening existing ones.

In practice

If you sell in a new category, your real competitor is indecision, and education is how you beat it. Measure the effect of educational content by completion rate and the activation that follows.

In practice

There is no single number; check whether the rates across the wheel's links (say satisfaction, referral, new customers) are improving in the same direction. Momentum comes from the links strengthening together, not from one campaign.

In practice

Value is realized in use, not at the point of sale, and the renewal decision reflects the gap. Agree with the customer up front on what success looks like, then report against that measure.

In practice

A share of voice above your market share often precedes market share gains, while the reverse signals falling behind. State which channel you measure on, because the result changes with it.

In practice

Leverage cuts both ways: when revenue falls, profit shrinks just as fast. A high fixed-cost structure is an advantage when you are confident in the growth forecast and a risk when you are not.

In practice

Speed means something only together with learning; shipping a lot and measuring none of it is not velocity. Shortening the time from idea to market lets you learn before competitors do.

In practice

Advocates are often the most satisfied customers, not the biggest spenders. Identify them, make sharing easy, and tag referred customers separately so you can measure the real effect.

In practice

Assess every growth opportunity on a risk-adjusted basis: high-return moves you cannot undo should be treated differently from ones you can test in small steps. Decide up front what you can afford to lose.

In practice

The index you deflate with does not reflect your own price and cost structure. A business that raises prices much faster than CPI and one that can barely move prices get deflated by the same number. Treat real growth as a more honest first read than the nominal figure, not as a verdict.

In practice

When inflation is low the gap is small and the nominal number will do. When it is high, tying targets, bonuses or an investor deck to nominal growth means price increases can explain much of the growth, and rewards can be paid out before anyone notices. Separate how much of the figure came from units.

In practice

TMS 29 restates the statements; it does not calculate growth, and restated accounts do not replace a real growth analysis in management reporting. Scope is not everyone either: entities outside those frameworks are not covered. Tax rules (VUK) are a separate regime, so do not mix the two.

In practice

The denominator is net new ARR, with churn and contraction already taken out; use gross bookings and the multiple looks better than it is. A single quarter can swing hard, and one large contract can distort the number for a period (rule of thumb). Read the trend. In a profitable company there is no burn and the number means nothing. For bands, use the calculator.

In practice

The score moves with the profit metric you pick: Feld uses EBITDA, while Bessemer, under the name "efficiency score", adds ARR growth to free cash flow margin. Choose one and do not compare yours with figures built on another. Bessemer says that score generally only becomes meaningful after $25 million of ARR; Feld describes his rule for companies with at least $50 million in revenue. In the early stages the score says little. The same total can hide two very different businesses.

In practice

Burn does not stay fixed: hiring pushes it up and revenue growth pulls it down. The real question is not how many months are left but whether you reach profit before the money runs out. Paul Graham's default alive question assumes expenses stay flat and revenue growth continues at its recent pace. Dividing cash by gross burn is the more conservative calculation: it ignores revenue, so it shows a shorter runway. Say plainly which one you use.

In practice

Because expansion is excluded, GRR cannot exceed 100%. Set it next to Churn / NRR and the difference is clear: NRR can rise above 100% through upsell, cross-sell and upgrades, while GRR tells you how much of the existing revenue remains after churn and contraction. If NRR is high and GRR is low, part of the growth may be going into replacing losses (our reading). Calculate both on the same customer group and the same period, and read them side by side.

In practice

Runway counts the money in months; default alive asks whether that time is enough to reach profit. The answer rests on two assumptions: expenses stay flat and growth keeps its pace. Hiring raises expenses and growth can slow; either changes the answer (our reading). Graham argues against leaning on investors: he writes that they are so fickle that you can never do more than start to count on them. He also says you should know, as in write down, precisely what you would need to do to survive if you cannot raise more money. When you run the calculation, write down the assumptions too.

In practice

What affects the result most is which costs you treat as variable (our reading). Wall Street Prep gives purchases of inventory or materials, direct labor and shipping as examples. In an AI product, inference cost grows with sales, so it can be treated as variable (our reading); see the Inference Cost concept. Calculated per product or customer group, contribution margin shows which sales actually leave something toward fixed costs (our reading). State which items you count as variable and keep the same definition across periods.

In practice

There is no single "cost per query" figure: providers price input and output separately, and cached input and batch processing as separate line items. Caching has separate prices for writes and reads; if the hit rate is low, the saving is smaller than expected (our reading). If you set price for the average user, cost can exceed revenue on the heavy user; track cost as a distribution, not an average (our reading). Take the current table from the provider's page and do not bake an old figure into your model.

In practice

It looks logical for a product whose cost scales with usage, but revenue is exposed to swings. In 2024 Kyle Poyar reported that many people got spooked by that perceived uncertainty, and that during the past two years of SaaS slowdown, usage-based businesses saw revenue growth decelerate faster than traditional subscription businesses. The real tension sits at the application layer: in Poyar's May 2024 review of 40 application-layer AI startups, seven in ten have a subscription model and very few offer pure usage-based pricing; that is a small, selected sample, not a current measure of prevalence. It means cost can climb with usage while subscription revenue stays flat (our reading). Before tying revenue to usage, work out what the income statement looks like in a slow quarter.

In practice

Bessemer presents hybrid as the model that wins when you are uncertain and calls it an effective middle ground for early-stage startups; do not stretch that scope to everyone. The hard part is how much usage the base fee includes: if the allowance is generous, the fixed fee will not cover the heavy user's cost (our reading). Set the allowance from your own usage distribution, not by instinct. One more caution: in Poyar's 2026 report, companies with hybrid pricing are generally the happiest, yet struggle to explain their pricing to customers; a two-part bill can give up some of the "simple and clear" pricing principle.

In practice

Bessemer spells out the conditions: be confident in your AI's performance, be able to absorb cost variance, and have an outcome that is unambiguous and measurable. A hard case can consume far more compute than expected; the fee stays fixed while the cost moves. Pricing before you define the outcome means writing a contract that invites disputes (our reading). Poyar describes a shift toward charging for the work delivered by software and AI agents; do not read that as "everyone is moving to outcome-based".

In practice

In the 40 AI startups Poyar and Broe reviewed in 2024, most companies charge based on the number of users, which the authors link to AI being sold as a copilot that assists the employee. According to the authors, customers can end up buying more seats than they need; monthly active user rates between 20% and 40% are not uncommon (the authors' observation, not a measured sample). As products deliver the work rather than assist the employee, the seat fee can detach from value (our reading). In 2025 Poyar makes the point about agents too: the value of AI agents does not scale naturally with buying more seats. That does not mean the seat is dead; where value grows with the number of users, a seat can still make sense (our reading). As the product gets better at automation, the customer needs fewer seats, so revenue can fall as the product improves (our reading). Before choosing a seat, ask: do my cost and the value the customer gets grow with the number of users?

In practice

For growth, an agent is less a product feature than a decision about authority: it sets which step of the funnel (acquisition, activation, retention) you hand to whom, and with how much autonomy (our reading). Anthropic and OpenAI start from the same place: the simplest solution first. Anthropic says outright that this might mean not building an agentic system at all. OpenAI recommends agents in three situations: complex decisions, rule sets that have become hard to maintain, and unstructured data. Before handing a growth step to an agent, ask: would a fixed workflow do it just as well, for less?

In practice

Your competitor can use the same model; the difference comes from the customer context the model works with (our reading). In personalization, activation and retention, context often matters more than the model: an agent that does not see the customer's plan, recent usage and open support ticket at the right moment gives everyone the same answer (our reading). Context is not free either; every piece you add is tokens, which means inference cost. The test: is what the agent sees about the customer the same as what a good account manager checks before a call?

In practice

Evals are the agent-world counterpart of an experimentation culture: a growth team without A/B tests decides on instinct, an AI team without evals decides on complaints (our reading). Anthropic describes the loop of a team without evals: wait for complaints, reproduce manually, fix the bug, hope nothing else regressed. Its advice is to start small: 20-50 simple tasks drawn from real failures is a great start, and evals get harder to build the longer you wait. The growth test: does your set measure the work the customer pays for, or the model's general ability?

In practice

For retention, a human in the loop is a condition of trust before it is a cost line: one agent that cancels a customer's subscription by mistake can wipe out the trust built by many good answers (our reading). OpenAI describes this intervention as the way to find failures and edge cases early and build a solid evaluation cycle; Anthropic also notes that agents can pause for human feedback at checkpoints. It belongs in unit economics too: every task handed to a person adds cost. Report the agent's success rate together with its handoff rate (our reading).

In practice

According to METR's March 2025 measurement, over the past six years this length followed an exponential trend, doubling roughly every seven months. Know the limits: it is measured on multi-step software and reasoning tasks, and the authors themselves note the possibility of substantial model error. The growth reading: give agents short, reversible work first (one email variant, one segment list) and keep long, multi-step work (running a campaign end to end) under human approval (our reading). Order the agent work in your growth roadmap by task length, but do not tie the schedule to a trend line (our reading).

In practice

For growth, MCP is less an integration detail than a distribution channel: if you expose your product's data and actions through an MCP server, the AI assistant your customer already uses can call your product (our reading). It reopens the old question of integration strategy: which work, once opened to agents, grows usage, and which turns your product into an invisible data pipe in the background? (our reading). Pick the one action you will expose first, then measure the usage that comes through it separately.

In practice

In A2A an agent advertises its capabilities in a JSON "Agent Card"; a client agent reads these cards to pick the best agent for the task. If there is an agent on the buyer's side, your product is chosen by a capability description a machine reads, not by a landing page (our reading). It is an early standard; do not tie your acquisition plan to it (our reading). But ask: can a machine understand what your product does from a single description? (our reading)

In practice

The acquisition channel may be shifting. Menlo Ventures' December 2025 report finds that 27% of AI application spend comes through product-led growth (PLG) motions, against 7% in traditional software. If users already find and try products on their own, the next step is an agent doing it for them (our reading; Menlo makes no such claim). The test: could an agent find your product, try it and get the right job done without writing to your sales team?

In practice

According to Rodriguez and Immerman (a16z, 2025), this wave makes early retention curves a less decisive indicator of long-term retention; their fix is to rebase retention and CAC from M0 to M3. In their words, M3 is the point where the hobbyists have mostly churned out and the users who found high-impact use cases remain. M3 is not a rule: look at where the curve flattens in your own cohort matrix (our reading). Instead of trying to hold on to tourists, find the job the stayers are doing and build activation around it (our reading).

In practice

For a founder the real question is market size: which budget does your customer pay from, software or people and services? Bek writes that for every dollar spent on software, six are spent on services; an autopilot goes after the second budget. His other argument is about defensibility: sell the tool and you are racing the model; sell the work and every model improvement makes your service faster and cheaper. The price shows up in unit economics: whoever delivers the work also carries the operations and the liability for mistakes, and gross margin can look more like a services firm than a software company (our reading).

In practice

The line from a16z's Erten and Wang (August 2026) gets to the heart of it: a credit is a currency, not a value metric; what matters is what the credit buys. In their survey of 50 technical AI buyers, 27 preferred credits tied to recognizable work and 14 preferred tokens; a small survey, not a measure of prevalence. Their advice is to price at the highest layer of value you can reliably measure, attribute and defend. For growth, a work unit ties the customer's bill to the growth of their own work; but set the credit's exchange rate wrong and you carry your heavy-usage segment at a loss (our reading).

In practice

Bessemer's 2025 report says the fastest-growing AI startups ("Supernovas") run low margins and are often trading distribution for profit in the short term; the same report adds that this fast adoption may reflect low switching costs or a novelty that may not align with long-term value. A fast-growing product with thin margins may be buying its growth with its margin (our reading). Track margin by cohort and usage band, not by the average customer; the heavy user can be margin-negative (our reading). The question to ask: is your growth rate buying your gross margin?

In practice

Bessemer's 2025 report finds this measure at 4-5x a typical SaaS benchmark among the fastest-growing AI startups, but also low gross margins in the same group. Read the two together: a small team may have moved its cost into the inference bill, so ARR per employee can be high while contribution margin stays weak (our reading). An agentic growth team has the same trap: using agents instead of hiring raises the ratio, but if agent cost and the human hours spent fixing errors are left out, the ratio misleads (our reading).

In practice

AI made building free, not knowing: the sample-size formula has no term for who wrote the variant, only traffic, base rate and effect (our reading). Ten experiments a week on the same page means traffic that is enough for none of them; the result is more false positives and exaggerated winners. The right measure is not the number of experiments but the number of decisions changed per unit of traffic (our reading). Kohavi's low-power table and Georgiev's review of 1,001 tests (median effect 0.08 percent) show the two ends of this mechanism (our reading).