Direct Answer
Software is not dead. The easy story is.
AI has made code dramatically easier to produce, so merely being able to build a small application is no longer the advantage it once was. At the same time, search behavior is changing, generic features are easier to copy, and AI usage introduces real variable costs. That weakens the old indie-hacker formula: build a narrow tool, publish SEO pages, and wait for high-margin subscriptions.
Greg Isenberg's answer is not to stop building. It is to move the advantage somewhere harder to copy: a deep niche, proprietary data, maintenance, customer trust, distribution, community, network effects, or an agent that owns a valuable outcome.
Watch Greg Isenberg's Live Q&A
Video credit: Greg Isenberg: "Everyone is saying SOFTWARE IS DEAD (LIVE Q&A)". Follow Greg on YouTube and on X.
Credits, Scope, and Evidence
Greg's stream combines founder commentary, examples shared on X, audience questions, and his own operating experience. This article reorganizes that discussion into a practical builder framework. It is independent and is not sponsored by Greg, Idea Browser, Google, PostHog, or any company mentioned.
The stream's examples are best treated as creator observations, not proof that every SaaS company or search-dependent business is declining. Where the article discusses search traffic, it distinguishes an academic field experiment from Google's aggregate platform claim. Business recommendations are JQ AI SYSTEMS analysis.
Useful Links
| Resource | Evidence type | Why it is useful |
|---|---|---|
| Full live Q&A | Creator commentary | Greg's complete argument, audience questions, and examples. |
| Greg Isenberg on X | Creator profile | Current startup, community, and distribution commentary. |
| Idea Browser | Product mentioned by Greg | A data-backed startup-idea and trend research product. |
| Pieter Levels: AI factories without traffic | Founder commentary | The distribution warning Greg reacts to during the stream. |
| AI Overviews field experiment | Academic working paper | Causal evidence on outbound clicks and zero-click searches. |
| Google's Search traffic position | Official platform statement | The counterpoint on aggregate click volume and click quality. |
| PostHog's product-market-fit guide | Operator guide | High-value events, ideal customer profiles, and retention measurement. |
What Actually Changed
The strongest part of Greg's argument is the distinction between software and a temporary software arbitrage. Ten years ago, a founder who could combine a web framework, payments, hosting, and SEO possessed a relatively scarce bundle of skills. That scarcity made a modest product defensible enough to earn attention and revenue.
AI coding agents reduce that scarcity. A competent competitor can reproduce a visible feature set quickly. The same agents can generate landing pages, comparison articles, support content, and thousands of experiments. More software can be produced than the market has attention to inspect.
| Old advantage | Why it weakened | Stronger replacement | What to build |
|---|---|---|---|
| Ability to code | Agents reduce build time and implementation scarcity. | Customer access and domain judgment | A vertical workflow for a narrowly defined buyer. |
| Generic SEO content | Answer engines satisfy more informational intent without a click. | Owned and diversified distribution | Newsletter, community, partnerships, creators, outbound, and paid tests. |
| Feature breadth | Visible features are easier to copy. | Outcome ownership | A system that completes, checks, and maintains the work. |
| Dashboard as product | Agents can operate tools without navigating every screen. | System of record and integration depth | Reliable data, permissions, audit logs, APIs, and agent interfaces. |
| Near-zero marginal cost | Inference, generation, support, and retries add variable cost. | Measured unit economics | Routing, caching, bounded context, quotas, and paid depth. |
This is not the disappearance of software. It is the end of treating software production itself as the business model.
The Search Traffic Question Needs Two Answers
Greg reacts to founders reporting weaker Google referrals, signups, and revenue. Their dashboards matter because a business lives on its own numbers. They do not, however, establish an internet-wide cause.
A 2026 field experiment by Saharsh Agarwal and Ananya Sen randomly varied whether users saw Google AI Overviews. Conditional on an AI Overview appearing, outbound organic clicks fell 39.8% and zero-click searches rose 34.5%. That is credible evidence that an answer shown on the results page can reduce downstream visits for affected searches.
Google presents a different aggregate view. In its August 2025 statement, the company said total organic click volume from Search had remained relatively stable year over year and that average click quality had increased. Google also acknowledged that traffic is shifting between sites and that users increasingly seek forums, video, podcasts, original analysis, and firsthand perspectives.
SEO remains useful for explicit commercial questions, original research, tools, calculators, comparisons, and proof. Generic informational content is easier for answer engines to absorb. The response is not to abandon search; it is to publish material worth visiting and build additional routes to the customer.
Seven Software Moats That Still Work
- A painfully specific customer. "Software for small business" is weak. "Permit and inspection coordination for independent commercial electricians in one regulatory market" is actionable.
- Proprietary or permissioned data. A model can imitate an interface; it cannot automatically reproduce a trusted dataset, feedback history, or operational record it does not possess.
- Maintenance ownership. Customers pay to keep integrations working, data clean, permissions current, and exceptions resolved. Maintenance is not residue after the product; it can be the product.
- Workflow depth. The defensible layer is often the complete path from trigger to approved outcome, including edge cases, audit trails, and handoffs.
- Distribution and affinity. Attention creates discovery. Repeated useful contact creates trust. Together they reduce dependence on a single algorithm.
- Network effects and community. Products become harder to replace when the participants, shared knowledge, reputation, or collaboration improve the experience.
- A clear point of view. Greg cites companies such as 37signals as examples of businesses whose mission and operating philosophy attract customers, not merely their feature lists.
A moat does not have to be grand. A founder who answers customer calls, understands a peculiar workflow, ships the missing integration, and fixes failures every week can be harder to displace than a better-funded generic app.
Do Not Bolt AI Onto the Old Product
Greg uses the mobile transition as the useful analogy. Instagram was not simply Facebook squeezed onto a smaller screen; it was designed around the camera, feed, and behavior of a mobile device. Agent-native software deserves the same rethink.
Ask these questions before adding a chatbot:
- What outcome can the agent complete instead of merely describing?
- Which facts must come from the product's trusted data?
- What can run automatically, and what requires approval?
- How does the agent verify that the result is correct?
- Can another agent become the user through an API or MCP interface?
- What durable record remains after the conversation ends?
An MCP server can be a valuable interface, but "MCP-first" is not a complete business. The installed audience is still smaller than the mainstream software market, protocols evolve, and customers pay for completed work rather than connector novelty. Build the valuable workflow first; expose it through the interfaces customers actually use.
Build Distribution Around Attention and Affinity
At 35:29, Greg reduces distribution to two jobs: earn attention and create affinity.
| Job | Question | Useful channels | Evidence to track |
|---|---|---|---|
| Attention | How will the right person discover us? | Search, short video, creator partnerships, communities, outbound, paid ads | Qualified reach, replies, visits, demo starts |
| Affinity | Why will they trust and remember us? | Newsletter, live sessions, case studies, support, events, customer stories | Return visits, direct traffic, referrals, branded search, sales velocity |
| Conversion | What proof makes action feel safe? | Demo, paid pilot, sample deliverable, guarantee, reference call | Activation, close rate, time to value |
| Retention | Does the result remain valuable? | Product usage, review calls, success reporting, community | High-value events, retained cohorts, expansion, churn |
A founder does not need to become an influencer. Distribution can be a partner channel, a niche directory, local events, outbound research, a practitioner community, or paid acquisition with sound economics. The requirement is an intentional route to customers that exists before launch day.
Greg's Audience-Community-Product framework is one route: earn a relevant audience, deepen the relationship through community, then offer a product shaped by what that group repeatedly needs. It is not the only route, but it forces distribution into the design instead of leaving it until after the build.
The Service-First Path Is Underrated
Asked how he would make an initial $25,000, Greg chooses a niche service with a one-time fee and recurring component. Later, when asked whether to sell agents or sell a service fulfilled by agents, he again starts with the service.
That sequence is practical because a service exposes the real work:
- Sell one measurable outcome. Avoid "AI transformation." Sell a shorter response time, cleaner pipeline, faster reporting cycle, or reduced rework.
- Deliver with human-reviewed agents. Keep the customer buying the result, not the internal tool stack.
- Record exceptions. Every manual correction reveals a requirement the product must eventually handle.
- Standardize the repeated path. Turn intake, analysis, execution, review, and reporting into a bounded operating system.
- Productize only proven repetition. Build software after several customers pay for substantially the same workflow.
The resulting business may remain a high-margin service, become software, or settle into a hybrid. The label matters less than whether the customer receives reliable value and the delivery economics improve with repetition.
Technical Literacy Still Matters
Greg does not argue that technical understanding is obsolete. He recommends learning the architecture-level concepts that help a founder direct and verify agents: APIs, MCP, web application stacks, data structures, basic algorithms, deployment, and how system components interact.
The goal is not to read every generated line. It is to recognize unsafe architecture, ask better questions, design acceptance tests, understand cost, and know when the agent is confidently wrong. AI reduces the price of implementation; it increases the value of judgment.
Technical founders now need the complementary craft too: customer interviews, positioning, offers, sales, content, partnerships, and retention. Greg is blunt about the transition: if code was your old unfair advantage, learn to market.
Design the Economics Now, Not After Token Prices Fall
AI products can look like traditional SaaS while carrying very different costs. Long contexts, repeated tool calls, image or video generation, model retries, human review, and customer support all grow with usage.
Greg rejects building an unprofitable product on the assumption that inference prices will eventually rescue it. That is sensible. Lower model prices can be offset by heavier usage, larger contexts, more ambitious workflows, or new platform pricing.
| Metric | Calculation | Why it matters |
|---|---|---|
| Cost per accepted outcome | Models + tools + compute + review + support / accepted results | Measures the real delivered unit, not the cheap-looking API call. |
| Contribution margin | Revenue minus variable delivery cost | Shows whether more usage creates or destroys cash. |
| Correction burden | Human repair minutes per completed outcome | Reveals hidden service labor inside "autonomous" software. |
| Activation | Customers reaching the first high-value event | Separates signups from experienced value. |
| Retention by ICP | Repeat high-value use among the target cohort | Tests whether the chosen customer truly keeps the product. |
A free plan should expose enough value to create confidence, but it needs a boundary. Charge for depth, volume, collaboration, proprietary data, automation, integrations, guarantees, or ongoing maintenance. "Unlimited AI" is not positioning; it is an unpriced liability.
Software Durability Scorecard
Score each item from 0 to 2: 0 means absent, 1 means plausible, and 2 means proven. A score below 10 suggests that the idea is still easy to copy or difficult to distribute. A score above 15 is worth a focused paid pilot, not a guarantee of success.
| Question | 0 | 1 | 2 |
|---|---|---|---|
| Does one specific buyer have this painful problem repeatedly? | Generic user | Named segment | Paid evidence |
| Do we own or earn access to unique data or context? | No | Can collect | Already compounding |
| Does the product complete and verify an outcome? | Advice only | Partial workflow | End-to-end with evidence |
| Is there a reason to stay after the first result? | One-off | Convenience | Record, learning, or collaboration |
| Can we reach customers without one gatekeeper? | One channel | Two tests | Diversified and repeatable |
| Does maintenance create continuing value? | No | Occasionally | Core promise |
| Are variable costs bounded at expected usage? | Unknown | Estimated | Measured per outcome |
| Do users retain after reaching value? | Unknown | Early signal | Flattening cohort |
| Can agents use the system safely? | No interface | Experimental API | Scoped tools and audit trail |
A 30-Day Builder Plan
| Week | Work | Required evidence |
|---|---|---|
| 1: Problem | Interview ten people in one niche. Map one frequent, expensive workflow and its current workaround. | Repeated problem language, owner, frequency, current cost, and buying trigger. |
| 2: Service | Sell a narrow paid pilot and deliver it with a human-reviewed agent workflow. | Payment, baseline metric, accepted result, correction log, and testimonial permission. |
| 3: System | Standardize intake, execution, verification, reporting, permissions, and exception handling. | Second delivery is faster, cheaper, or more reliable than the first. |
| 4: Distribution | Run one attention channel and one affinity channel. Publish proof, contact partners, and ask for referrals. | Qualified conversations, source attribution, close rate, and next experiment. |
Do not build a portfolio in month one. Greg's advice is to take one product to real product-market fit before expanding. PostHog's practical measurement guidance is useful here: define high-value events, identify the ideal customer profile, and look for retention rather than celebrating registrations.
Key Moments in the Live Q&A
| Time | Topic |
|---|---|
| 09:34 | The "software is dead" discussion begins |
| 15:06 | Founder traffic and revenue examples |
| 20:08 | What remains worth building |
| 22:06 | Distribution, niches, data, and maintenance as new advantages |
| 30:27 | Why coding scarcity changed |
| 35:29 | Distribution as attention plus affinity |
| 37:56 | The technical concepts founders should understand |
| 54:14 | How Greg would pursue the first $25,000 |
| 1:00:44 | Audience, Community, Product |
| 1:04:46 | Why demand capture is harder |
| 1:06:45 | Go agent-first and make distribution part of the product |
| 1:10:21 | What agent-first software means |
| 1:23:03 | Sell the service and let agents help fulfill it |
| 1:30:47 | Profitability and token-cost assumptions |
| 1:38:42 | Why creative and customer-connected founders have an opening |
Bottom Line
AI did not kill software. It made software production less scarce and exposed weak products whose real advantage was code plus borrowed traffic.
The new indie-hacker opportunity is less passive and more interesting: understand a narrow customer better than a general model can, own the difficult workflow, accumulate useful data, maintain the result, build several paths to demand, and design the product for agents as well as people. Start with a paid outcome, measure what customers retain, and let the software emerge from repeated truth.
That is harder than generating an app in an afternoon. It is also much harder to copy.
Sources
- Greg Isenberg: Everyone is saying SOFTWARE IS DEAD (LIVE Q&A)
- Greg Isenberg on YouTube and on X
- Pieter Levels: Indie hackers build fancy AI factories but have no money or traffic
- Agarwal and Sen: The Impact of Google AI Overviews on Publisher Traffic and User Experience
- Google: AI in Search is driving more queries and higher quality clicks
- PostHog: Measuring product-market fit is more than vibes
- Idea Browser