Which online businesses are most AI-proof now?

Has AI already shown us which online businesses are easiest to replace?

AI has already exposed the weakest online businesses: generic information, routine digital production and simple tasks that customers can describe clearly enough for a model to do them.

Chegg gives us one of the cleanest examples. Its latest quarterly revenue fell 51% year over year to $51.8 million. Chegg built much of its historical consumer value around helping students get explanations and answers quickly. General-purpose AI now does a large part of that job instantly, interactively and at little extra cost.

Fiverr gives us a broader view because its marketplace covers many types of digital work. Its latest marketplace revenue fell 15.5%, while annual active buyers dropped 21.9% to 2.7 million. Fiverr's own management said rapid AI adoption was hurting the categories most exposed to automation.

The interesting part appears one level deeper. Fiverr clients completing projects worth more than $1,000 actually increased 13% over the previous twelve months. Upwork shows something similar: AI-related project volume grew more than 22%, AI strategy and consulting grew over 50%, and spending through its higher-touch Business Plus offer jumped 174%.

AI is already sorting digital work by how easy the outcome is to specify. “Write these descriptions,” “summarize this document” and “make this basic graphic” are becoming cheap. “Understand our business, fix this messy problem and be accountable for the result” is holding up far better.

Online business	What the latest evidence says	AI exposure
Generic information subscriptions	Chegg revenue down 51%	Very high
Low-value freelance tasks	Fiverr says AI is absorbing transactional work	High
Higher-value freelance projects	Fiverr $1,000+ projects up 13%	Lower
AI consulting and implementation	Upwork AI consulting volume up 50%+	Currently growing

Are SEO websites still a good AI-proof business?

Pure SEO publishing is one of the weakest online businesses to build today because AI is squeezing both the cost of producing content and the number of clicks that content receives.

The latest Ahrefs study makes the traffic problem harder to dismiss. Ahrefs re-ran its analysis across 300,000 keywords and found that a Google AI Overview was associated with a 58% lower click-through rate for the number-one organic result versus its modeled counterfactual. Its earlier study had measured 34.5%.

That jump from roughly one-third to well over half is more important than another anecdote about one publisher losing traffic. It suggests that the effect intensified as Google's AI answers spread.

The production side is moving in the same direction. Researching a common question, writing a decent article, generating supporting images and publishing variations of the page have all become dramatically easier. A site built mainly around doing those things no longer owns much scarcity.

There are still strong information businesses. They usually possess something harder to reconstruct: proprietary research, original reporting, a trusted expert, a valuable audience, private data, paid access to people or a transaction that follows the content.

A page answering a public question has become much easier to replace. A business that happens to publish useful pages can still be excellent.

Is SaaS actually safe from AI?

SaaS is far too broad to call AI-proof because AI agents can weaken some software products while making others more valuable.

Gartner recently estimated that as much as $234 billion of enterprise application spending could be exposed to what it calls “agentic arbitrage” by 2030, roughly 20% of projected SaaS spending. The idea is simple: once an AI agent operates several applications for an employee, companies have less reason to pay for a separate human seat in every interface.

That puts pressure on software whose main value lives in the interface. A dashboard that mostly rearranges information, a thin reporting product, a basic writing app or a narrow workflow with little proprietary context can increasingly be bypassed.

The latest numbers from harder-to-bypass software look very different. Veeva grew revenue 18% and subscription revenue 16%. CrowdStrike's annual recurring revenue grew 25% to $5.84 billion. HubSpot's revenue grew 20%. These companies are adding AI aggressively rather than watching AI make the underlying platforms disappear.

The dividing line is becoming clearer. We would worry about software that helps a user manipulate information an agent can obtain elsewhere. We would feel much better about software that holds the authoritative customer record, security policy, regulated document, transaction history or operational workflow the agent needs before it can act.

Why does vertical SaaS look much safer than generic SaaS right now?

Vertical SaaS looks unusually strong today because the best products become part of how a real industry runs, which gives AI more reasons to use them rather than bypass them.

ServiceTitan is a useful example. The company handles scheduling, dispatching, customer records, financing, payments and other workflows for trades such as HVAC, plumbing and electrical contractors. Revenue has continued growing above 20%, and its platform already handles tens of billions of dollars of transaction volume.

Veeva goes deeper into one industry. Pharmaceutical and biotechnology companies use its software across commercial operations, clinical work, safety, quality and regulatory processes. Its latest revenue reached $928 million for the quarter, up 18% year over year, even as Veeva itself pushes AI agents into those workflows.

Procore follows the same basic pattern in construction. Its customers manage projects, documents, communication and contractors inside the platform. The useful data comes from people actually running construction projects rather than from information anyone can scrape from the web.

This is the good setup. Better AI can automate quoting, documentation, scheduling or analysis, while the vertical platform keeps the customer history, permissions, integrations and operational record underneath.

For a small founder, the interesting opportunity is often much narrower than “build SaaS.” Software for one ugly workflow inside one expensive industry can be safer than a beautifully designed horizontal app with a much larger theoretical market.

Business	What customers rely on it for	Why AI has trouble bypassing it
ServiceTitan	Trade-business operations and transactions	Real customer and job workflow
Veeva	Life-sciences records and regulated processes	Compliance, permissions and audit history
Procore	Construction project operations	Project data and multi-party workflow
Toast	Restaurant software and payments	Transactions tied to physical locations

Are cybersecurity and identity businesses becoming more valuable because of AI?

Cybersecurity and identity currently look like some of the strongest AI-resistant online businesses because more autonomous software creates more things that companies need to control.

CrowdStrike's latest quarter was exceptionally strong. Revenue reached about $1.47 billion, annual recurring revenue rose 25% to $5.84 billion, and the company added a record $333 million of net new ARR. That last figure grew 51% year over year.

Those numbers are happening while companies are adopting AI faster, not before AI arrived.

The reason is practical. Every new agent that can open files, call APIs, write code, access customer systems or trigger transactions creates another identity with permissions that can be abused. AI can also make phishing, malware development and vulnerability discovery easier for attackers.

An employee asking an AI assistant to prepare a report may use fewer conventional software screens. The company still has to decide what that assistant can access, detect suspicious behavior and stop compromised credentials.

Individual security features will get copied. Some vendors will disappear as platforms consolidate. But security itself is moving closer to a mandatory layer underneath AI-heavy businesses, which makes the category fundamentally different from optional productivity software.

Are payments one of the safest online businesses from AI?

Payments are among the strongest AI-resistant online businesses because better AI can change who initiates a purchase without removing the need to authorize, settle and record the money.

Stripe says businesses running on its infrastructure processed $1.9 trillion in payments last year, up 34%, equivalent to roughly 1.6% of global GDP. More than five million businesses now use Stripe directly or through platforms.

Toast gives us the same pattern inside restaurants. Its recurring revenue reached $2.2 billion earlier this year, up 26%, while payment volume increased 22% to $51.3 billion for the quarter. The platform was already serving roughly 171,000 locations.

AI can make checkout easier. Agents may eventually compare products, choose what to buy and initiate purchases with little human involvement. Every one of those purchases still needs payment authorization, fraud controls, settlement, refunds, accounting and often tax calculation.

AI may remove a lot of the visible interface around commerce while generating more activity for the infrastructure underneath it.

For founders, that does not mean launching another generic payment processor. The more accessible opportunities sit around narrow payment problems, reconciliation, invoicing, industry-specific billing, risk, tax and money movement inside underserved workflows.

Which marketplaces are actually protected from AI?

Marketplaces built around scarce real-world supply are much safer than marketplaces whose sellers mainly produce digital output.

Airbnb shows why. AI can recommend where we should stay, compare neighborhoods and plan the trip. It cannot create an available apartment in Rome next weekend. Airbnb still has to aggregate real inventory, attract hosts, maintain reputation systems, handle the booking and process the payment.

The latest results remain strong. Airbnb's revenue grew 17% to $3.6 billion, gross booking value increased 16% to $27.2 billion, and nights and seats booked rose 10%.

Etsy has a different kind of scarce supply. Generative AI can produce an image of a custom ceramic mug in seconds. Shipping an actual handmade mug with a seller history, customer reviews and reliable fulfillment requires someone in the real world. Etsy marketplace merchandise sales have now grown year over year for three consecutive quarters after a long period of weakness.

Fiverr helps us see the other side. When marketplace supply consists of simple digital tasks, AI can reduce the need to hire a seller at all.

Marketplace defensibility therefore depends heavily on what sits behind the listing.

Marketplace supply	AI resistance	What remains scarce
Accommodation	High	Real properties and availability
Physical handmade goods	High	Production, fulfillment and seller reputation
Specialized expert work	Medium-high	Context, judgment and accountability
Commodity digital gigs	Low	The deliverable can often be generated
Public-information directories	Very low	AI can increasingly do the matching elsewhere

Is e-commerce itself AI-proof?

E-commerce remains a huge business, but a generic online store has become easier to copy than ever.

Shopify's latest quarter makes clear that online commerce itself is healthy: revenue grew 34%, and the company reported growth above 30% across GMV, revenue, gross profit and free cash flow.

People have not stopped buying online because AI exists. The risk sits with the merchant whose entire advantage was finding a product, launching a storefront and buying traffic.

AI now makes product research faster, creates product photos and ads, writes landing pages, answers support tickets and helps competitors reproduce a store quickly. Those tools improve operations for good brands while stripping scarcity from mediocre ones.

The e-commerce businesses we would trust most own something outside the storefront: an exclusive product, manufacturing knowledge, a hard-to-copy supply chain, habitual repeat purchases, a trusted brand or direct access to a valuable audience.

This is why AI can simultaneously make Shopify stronger and make thousands of interchangeable Shopify stores weaker. Shopify owns infrastructure used by the whole merchant ecosystem. A commodity merchant owns much less.

Are online services becoming safer than selling software?

Outcome-based online services are becoming more attractive because AI can reduce the work required to deliver the service while the customer still pays to have the problem solved.

The split inside freelance marketplaces supports this. Fiverr is losing activity in low-value transactional work, while its $1,000-plus projects are growing. Upwork's overall marketplace has been much softer than its AI strategy, consulting and higher-touch business products.

Imagine two offers. One freelancer sells ten hours of bookkeeping. Another promises clean monthly books, reconciled accounts and a finished reporting package for a fixed fee.

AI threatens the first offer directly because hours become cheaper. The second operator can automate categorization, reconciliation checks, document extraction and reporting while keeping the same customer outcome.

The same pattern can work in compliance, recruiting, marketing operations, insurance administration, research, localization, procurement and dozens of specialized B2B processes.

There is a catch. A generic “AI agency” has almost no protection when everyone uses the same models. The stronger service builds proprietary workflow knowledge, customer relationships, templates, integrations, historical data and a reputation for actually finishing difficult work.

That can turn AI from a substitute into a margin-expansion tool.

Are paid communities safer than online courses?

A good paid community is more AI-resistant than a static course because people can generate explanations cheaply while access to specific people remains scarce.

Online courses face a straightforward challenge. Someone who wants to learn cold outreach, Excel, Python or basic marketing can ask an AI tutor unlimited follow-up questions, request personalized exercises and get feedback immediately.

A useful community sells something different. Members might pay for access to experienced operators, introductions, accountability, private deal flow, recruiting opportunities, peer feedback or a reputation inside the group.

Substack now reports more than five million paid subscriptions across its network, showing that people will still pay repeatedly for direct relationships with creators and communities despite an enormous supply of free AI-generated information.

This category still requires caution. A Discord server full of generic posts has little moat. A paid newsletter whose entire value is summarizing public news also faces increasing pressure.

The stronger membership gives members access they cannot prompt into existence.

Is proprietary data becoming a bigger moat because of AI?

Unique data is becoming more valuable as AI gets better because powerful models still need the right inputs before they can produce a useful answer.

The phrase “proprietary data” gets abused, though. A database built from public websites has limited protection when another company can scrape the same sources or an AI system can search them directly.

Operational data is much harder to recreate.

ServiceTitan observes jobs, customers and payments because contractors actually run their businesses through the product. Toast sees restaurant transactions because payments flow through its network. Veeva holds information created during real clinical, commercial and regulatory workflows. CrowdStrike continuously sees security telemetry generated by customer systems.

Those datasets improve every time customers do real work.

That is a much stronger position than owning a collection of public information.

For a small founder, the lesson is useful. A narrow product can become defensible if normal customer usage creates unique history: private benchmarks, pricing records, performance data, verified supplier behavior, maintenance histories, transaction patterns or other information competitors cannot simply download.

Does regulation actually protect an online business from AI?

Regulated workflow businesses look unusually durable because customers still need proof that the correct process happened even when AI performs much of the work.

Veeva is our clearest large-scale example. Its software sits inside clinical, quality, safety and regulatory workflows in life sciences, and the company is still growing revenue 18% while actively adding AI agents.

A pharmaceutical company can absolutely use AI to draft documents, summarize evidence and assist with submissions. It still needs controlled records, approval history, permissions, validated processes and an audit trail showing what happened.

The same logic appears in payroll, tax, healthcare administration, insurance, financial reporting, import compliance and employment regulation.

That creates two very different online businesses. A website explaining tax rules faces heavy AI pressure because the model can explain those rules too. A system that collects the customer's information, applies the rules, creates the filing, tracks deadlines and records completion has a much stronger position.

Regulation works best as a moat when the business owns execution rather than information about the regulation.

What should a small founder actually build if they want an AI-resistant online business?

For a small founder today, the best AI-resistant opportunities usually sit inside narrow, expensive workflows where somebody still has to complete a real transaction or take responsibility for the result.

We would look closely at vertical software with payments, specialized compliance services, managed B2B operations, proprietary-data products, industry marketplaces, software tied to physical businesses, and memberships built around valuable access.

The size of the initial niche matters less than many founders assume. Software that handles quoting, scheduling and payment for one category of specialist contractor can develop stronger retention than a horizontal productivity app aimed at millions of theoretical users. A compliance service for one annoying filing can become valuable if customers repeatedly need the filing completed correctly. A marketplace can start small if the supply is genuinely scarce.

AI should sit underneath these businesses wherever it lowers cost. Let it classify documents, answer routine questions, draft work, reconcile records, qualify leads and automate customer support.

We would be much more cautious when the customer is paying mainly for the AI-generated output itself. General models keep improving, competitors can access similar capabilities, and the model providers can move up the stack.

The best small-business version of AI-proofing is practical: own the customer, own the workflow and gradually accumulate information or relationships that do not come bundled with the model.

Which online businesses are most AI-proof now?

The strongest online businesses today are vertical workflow platforms, payments and financial infrastructure, cybersecurity and identity, regulated execution, marketplaces with scarce real-world supply, proprietary operational data and outcome-based services.

The latest evidence is unusually consistent. CrowdStrike's ARR is growing 25%. Veeva's revenue is up 18%. Shopify's revenue is up 34%. Airbnb's booking value is up 16%. Stripe's annual payment volume grew 34%. These companies occupy very different markets, yet AI still needs the assets they control: permissions, records, inventory, transactions, compliance processes and merchant infrastructure.

At the vulnerable end, Chegg's quarterly revenue is down 51%, Fiverr is openly reporting AI pressure in low-value transactional categories, and Ahrefs' latest 300,000-keyword study found a 58% reduction in modeled click-through rate when Google AI Overviews appeared.

Fiverr adds an especially useful nuance. Higher-value $1,000-plus projects grew even while active buyers fell sharply. Customers are becoming less willing to pay humans for simple digital production while continuing to spend on work where context and accountability matter.

That gives us a fairly clear ranking today.

Rank	Online business model	AI resistance	Why it holds up
1	Vertical software tied to transactions	Very high	Owns the workflow, records and money movement
2	Cybersecurity and identity	Very high	More AI creates more permissions and attack surface
3	Payments and financial infrastructure	Very high	Every automated purchase still needs financial rails
4	Regulated execution and compliance	Very high	Customers need completion, records and auditability
5	Scarce-supply marketplaces	High	AI cannot generate real inventory or established reputation
6	Proprietary operational data	High	Unique inputs become more useful with stronger models
7	Outcome-based managed services	High	AI lowers delivery cost while responsibility remains valuable
8	Strong physical-product e-commerce brands	Medium-high	Product, supply and customer trust remain scarce
9	Paid expert communities	Medium-high	Valuable human access cannot be generated on demand
10	Generic horizontal SaaS	Medium-low	Agents can bypass interfaces and reduce seat demand
11	Online courses and generic information products	Low	AI can increasingly teach the same material interactively
12	Commodity freelance production	Low	Customers can automate much of the deliverable
13	SEO-dependent information sites	Very low	AI pressures both production economics and search traffic
14	Thin AI wrappers	Very low	Similar model capabilities are widely available

The pattern underneath the ranking is stronger than any individual category. Businesses become safer as they move closer to money, permissions, private records, physical supply, regulatory responsibility and completion of an important outcome.

Those are the areas where cheap intelligence helps the operator without making the operator unnecessary.

That is where we would build now.