How to Market
Software

Every few years, someone discovers a marketing channel that works unusually well. They write about it, other people copy it, the channel gets crowded, and the platform that hosts it notices the activity and starts charging for access. Effectiveness declines and everyone moves on to the next channel. This cycle has been running since at least the invention of email marketing, and probably since someone in ancient Rome figured out that you could pay people to recommend your olive oil in the forum.

AI compressed the cycle from years to months.

In 2019, a well-written cold email with genuine personalization could get a 15-20% reply rate because most cold emails were generic and bad. Writing a good one was a cost of skill and time, which limited competition. Then AI tools made it possible to generate thousands of "personalized" emails per day, each one referencing the recipient's company, recent blog post, and job title. The output was competent. It was also identical in structure to every other AI-generated cold email, which recipients learned to recognize almost immediately. The skill premium was competed away in about eighteen months.

This same pattern played out across channels roughly at the same time. AI-generated SEO content flooded search results, which made Google tighten its quality filters, which penalized everyone including the people writing good content. AI-generated social media posts increased volume without increasing insight, which trained audiences to scroll faster. AI-generated ad creative made A/B testing cheaper, which meant everyone optimized toward the same high-performing formats, which meant the formats stopped performing because they all looked the same.

Running underneath all of this is a structural incentive that has nothing to do with AI: the platforms that connect companies to buyers are advertising businesses. Their revenue comes from selling access to the audiences they've accumulated, which means their interests are aligned with making organic access harder over time. The business model is working as designed.

Marketing has never been easier to do and has never been harder to do well. The tools are better than ever. The competition is fiercer than ever. The platforms take a larger cut than ever. And buyers, having been subjected to a decade of increasingly sophisticated marketing, have developed defenses that are themselves increasingly sophisticated. They research in private channels. They trust peers over vendors. They've learned to pattern-match on marketing tactics well enough that the tactics stop working on them roughly as fast as marketers can deploy them.

The tempting conclusion is that marketing is broken. The more accurate conclusion is that marketing in which the primary advantage was access to tools and channels is broken, because tools and channels are now commodities. What remains scarce is judgment about which problems are worth solving and expertise deep enough to be actually useful, plus the patience to build trust with an audience over a timeline longer than a quarter. These have always been the things that mattered. They just used to be obscured by the fact that you could also succeed by being early to a channel or clever with a tactic. That's the part that stopped working.

Product Hunt solved a specific problem: good products are hard to discover, so let's build a place where early adopters surface them through voting and discussion. For a while, this worked. Then it stopped working.

The current state of Product Hunt is roughly what you'd get if you designed a system to answer the question "which products are good?" and then offered large financial rewards to anyone who could make the system say their product was good. Services sell fake upvotes openly. Comment sections fill with AI-generated enthusiasm that reads like it was produced by a language model asked to "write a supportive comment about a SaaS product," which is in fact what happened. Launches are won through coordination: activating networks of supporters at the right time, in the right pattern, to trigger the algorithm's preference for rapid early engagement.

This is Goodhart's Law applied to product discovery: when a ranking becomes a target, it ceases to be a useful ranking. The same thing has played out with app store rankings, Amazon reviews, Google search results, and every other system where a numerical score determines visibility and visibility determines revenue. The people with the most to gain from manipulating the score will always invest more in manipulation than the platform invests in detection, because the manipulators capture 100% of the benefit while the platform spreads the cost of enforcement across its entire operation.

Product Hunt has responded with the standard countermeasures: algorithm adjustments, more human moderators, better bot detection. These help at the margins. They do not solve the underlying problem, because the underlying problem is structural rather than technical. As long as a high ranking on Product Hunt translates into meaningful distribution, people will spend money to achieve a high ranking on Product Hunt, and some of that spending will be on manipulation rather than product quality.

The practical question for someone launching a product is whether Product Hunt still provides enough signal through the noise to be worth the effort. The answer is probably yes for a narrow set of use cases: if your target audience is the kind of early adopter who still browses Product Hunt regularly, and if you have a genuine network willing to engage authentically, a launch can generate a useful burst of attention. But the days when a Product Hunt launch could meaningfully change a company's trajectory are largely over, and the effort spent orchestrating a "successful" launch might produce better returns if directed at almost anything else.

The platforms that work best for discovery in 2026 are the ones where the ranking mechanism is either opaque enough that gaming is difficult or distributed enough that no single score determines visibility. Neither of these descriptions applies to Product Hunt.

Before AI writing tools, sending a good cold email was comparatively expensive. You had to research the recipient, write something specific to their situation, and make a genuine case for why your product was relevant to their problems. This took time, which meant you could only send a limited number per day, which meant the recipient's inbox contained a manageable number of cold pitches, which meant each one had a reasonable chance of being read. The friction was load-bearing. It kept volume low enough that the channel worked.

AI removed the friction. You can now generate a thousand personalized cold emails in the time it used to take to write ten. Each one references the recipient's company, their recent LinkedIn activity, their job title, and a plausible reason for reaching out. The individual emails are, in isolation, fine. The problem is that everyone else is also generating a thousand personalized cold emails, and the recipient's inbox now contains dozens of messages that are all competent, all personalized, and all structurally identical in a way that makes them immediately recognizable as AI-generated.

The result is that reply rates have collapsed. Most campaigns land somewhere between 1% and 5%. A small number of senders consistently exceed 10%, but they tend to be the ones who are doing something that AI can't easily replicate: writing from genuine expertise about a problem they deeply understand, or leveraging a real relationship, or saying something surprising enough that it couldn't have been produced by a model trained to be agreeable and professional.

The tool that was supposed to make outbound email better made it worse, because it made it better for everyone simultaneously, which is the same as making it better for no one. When every email in someone's inbox is personalized, personalization stops being a signal of effort and becomes background noise.

The trajectory of organic reach on every major social platform follows the same curve, and if you plot them all on the same graph they look like a family of functions converging on zero at different speeds.

Facebook got there first. If you run a business page on Facebook, your posts reach a low single-digit percentage of the people who chose to follow you. These people told Facebook, explicitly, "I want to see content from this page," and Facebook decided they mostly shouldn't, because showing them your content for free would reduce the incentive for you to pay to show them your content. This is the business model working as designed. Meta makes money by selling access to attention, and giving that access away would be like a landlord letting tenants live rent-free because they asked nicely.

Instagram followed the same path with a slight delay. LinkedIn is following it now, which is notable because LinkedIn was supposed to be different -- the B2B platform, the professional network, the place where company content still had organic distribution. Then LinkedIn started doing the same thing every platform does once it has enough users to monetize: reducing organic reach for company pages while building out its advertising product. Company page content now reaches a small fraction of followers.

The pattern is the same for X and for Threads and for everywhere, because the incentive is the same everywhere. A platform that has accumulated an audience has a choice: let brands reach that audience for free, or charge them for the privilege. No publicly traded company will choose the first option indefinitely, because the second option is where the revenue is.

The one consistent exception across platforms is that content from individual people still gets meaningfully more distribution than content from company pages. This is why employee advocacy and founder-led content keep coming up as strategies. They're not workarounds for a temporary glitch. They're responses to a structural feature of how every social platform allocates attention.

These platforms are fine as paid channels, priced accordingly. But building a marketing strategy around organic social distribution is building on ground that is subsiding by design, and has been for years, and will continue to, because the companies that own the ground make more money when it subsides. Planning for this is straightforward.

Pretending it isn't happening is expensive.

Advertising costs in most industries have been rising steadily, and in software they've been rising faster than in most industries. This makes sense if you think about it from the platform's side: digital ad inventory is sold by auction, there are more software companies bidding for the same audiences than there were five years ago, and auction prices go up when demand increases. The platforms have no incentive to fix this, because rising prices are how auctions are supposed to work. You are not being overcharged. You are being outbid.

Paid acquisition has a scaling problem that's easy to miss when you're looking at campaign-level metrics. Your first dollars of ad spend reach the people most likely to convert. Your next dollars reach slightly less likely prospects. At some point on this curve, your customer acquisition cost exceeds the lifetime value of the customers you're acquiring. Companies that aren't tracking this carefully don't notice when they cross the line, because the aggregate numbers still look healthy: total spend is up, total customers acquired is up, the dashboard is green. But the marginal unit economics have gone negative.

There is a game-theoretic element to this that resembles an arms race. Each company would be better off if everyone spent less, but no individual company can unilaterally reduce spending without losing share to the others. The result is an equilibrium where everyone spends more than they'd like to, the platforms capture the surplus, and the companies compete for roughly the same customers at higher and higher prices.

The exit from this cycle is to reduce your dependence on paid ads by building acquisition channels -- organic content, community, product-led growth, word of mouth -- that don't have the same diminishing returns curve. Paid acquisition works best as an accelerant on top of organic demand that already exists, not as the primary engine.

The economics of content production changed faster than the economics of content quality, and the result is, broadly, a wasteland.

When writing a 2,000-word article took a human writer four to eight hours, the cost of publishing imposed a natural quality floor. Companies that published SEO content before AI tools had to be at least somewhat selective about which keywords to target, because each article represented real time and money.

AI tools reduced the production cost to approximately zero, which removed the filter. A site that previously published twenty articles a month could now publish two hundred, covering every conceivable keyword variation, without a proportional increase in budget. Many did. The result is a search environment where the same basic answer to the same basic question appears across dozens of sites, each one generated by a similar model drawing on similar training data, producing content that is technically correct and functionally useless. The articles answer the query in the same way that a mirror answers a question about your appearance: the information is there, but nothing has been added.

SEO content still works, but the definition of "works" has narrowed. Content that ranks in 2026 is content that offers something a language model can't generate from its training data: original research, proprietary data, expertise applied to a specific problem, or a perspective unusual enough that it couldn't have been produced by averaging the existing corpus.

The mental model most companies use for B2B sales is roughly: a buyer becomes aware of a need, discovers your product, evaluates it through your content and sales team, and decides to purchase. The funnel. Marketing handles the top, sales handles the bottom, everyone has a dashboard.

The mental model most buyers actually use is: they have a problem, they ask people they trust what to do about it, they read third-party reviews and community discussions, they test free tiers or watch someone else use the product, they form a strong preliminary opinion, and then, if you're lucky, they contact your sales team to confirm the decision they've already mostly made. By the time a buyer fills out your "request a demo" form, the evaluation is largely over. You are not being auditioned. You are being verified.

This means the majority of the selling process happens in places your sales team cannot see and your analytics cannot track. The buyer's colleague mentioned your product in a team meeting. The buyer read a Reddit thread comparing you to two competitors. The buyer asked in a private Slack group whether anyone had experience with your onboarding process and got three replies, two positive and one cautionary. None of this will ever appear in your CRM. The "source" field will say "direct" or "organic search" because the buyer typed your URL into their browser after deciding to check you out.

The implication that makes salespeople uncomfortable is that buyers don't particularly trust vendors as information sources. This is rational behavior, not a character flaw. Vendors have an obvious incentive to present their product favorably, and buyers know this, and they discount vendor-provided information accordingly.

The companies that have adapted to this invest less in controlling the buyer's journey and more in being present, usefully, in the places where buyers actually do their research. This means producing content that's actually informative rather than optimized for lead capture, participating in communities without a visible sales agenda, and making the product easy to evaluate without requiring a conversation with a human -- and accepting that you will never have full visibility into why people buy from you, because the most important moments in the decision process happen in contexts that are private by design.

The traditional way to sell software is to put a wall between the product and the customer and staff the wall with salespeople. The customer sees a marketing site, books a demo, watches someone else use the product for thirty minutes, waits for a follow-up email, negotiates pricing, signs a contract, gets onboarded by a different team, and finally, weeks or months later, finds out whether the product actually solves their problem.

Product-led growth is what happens when you remove the wall. You let people use the product, or a meaningful version of it, before they pay. They experience the value directly rather than hearing about it secondhand from a salesperson whose incentives are not perfectly aligned with theirs. If the product is good, some percentage of free users convert to paying customers without anyone from your sales team being involved.

The speed at which users reach the "aha moment" determines almost everything about the economics. If it takes five minutes, you have a self-sustaining acquisition engine. If it takes five weeks, you have a free trial with a high abandonment rate.

This means product-led growth is less a marketing strategy than a product design constraint. The question is: "can we restructure the product so that a new user, with no help from us, reaches the moment of value fast enough that they decide to stay?" This is a hard engineering and design problem, and companies that treat PLG as a go-to-market tactic bolted onto an existing product tend to end up with a free tier that generates support tickets rather than revenue.

The failure mode worth watching for is the assumption that PLG eliminates the need for sales. It doesn't. It changes what the sales team does. In a PLG model, sales talks to people who are already using the product and have already experienced its value. The sales team's job shifts from generating interest to expanding accounts. Companies that fire their sales team because they've "gone product-led" tend to discover, painfully, that free users are good at staying free and that converting them to enterprise contracts still requires a human being.

Human brains are built to trust people and to be suspicious of institutions, and no amount of brand strategy can overcome a cognitive bias that's been adaptive for a few hundred thousand years.

Corporate accounts post content and people scroll past it. The same content, posted by a person with a name and a face and a history of saying things that are sometimes surprising, gets read. The gap has widened to the point where it's difficult to justify spending on brand-channel content if you have a founder who's willing to write.

It's not really about "authenticity," even if that's the word everyone uses. It's information content in the signal-theory sense. When a company account posts "we're excited to announce," you learn nothing, because a company account would say that regardless of whether the thing is exciting. The statement carries zero bits of information. When a founder posts something specific about a problem they encountered, a decision they made, or a belief they hold that might be wrong, you learn something about how they think, which is also information about how their company operates, which is also information about whether their product is likely to be good.

The practical implication is that founder-led content works best when it's actually led by the founder. What most companies mean by "founder-led content" is "content written by a marketing team and posted under the founder's name." This works for approximately two weeks, until the audience notices that the founder's posts read like everything else in their feed.

The constraint is that this requires a founder who can write, who's willing to be visible, and who has enough intellectual surface area to sustain an audience's interest over years rather than months. Not every founder is this person. The ones who aren't should not be coerced into performing a version of public presence that doesn't come naturally. In those cases, someone else at the company -- a cofounder, an early engineer, a head of product who has genuine opinions -- can fill the role. The important thing is that whoever does it is actually the person thinking and writing.

The conversations where buyers actually form opinions about which products to use have migrated into spaces that are invisible to most sales and marketing teams: private Slack groups, Discord servers, invite-only communities. A VP of Engineering asks their peer group "has anyone used X for Y?" and gets six replies from people they trust, and that conversation matters more than any case study or demo your marketing team has ever produced. You cannot buy access to that conversation. You cannot optimize for it. You probably don't know it happened.

This is sometimes called "dark social" because the activity doesn't show up in attribution models. Your analytics will record that the customer arrived via direct traffic or a Google search, and you will credit your SEO team, when what actually happened is that someone recommended you in a private channel and the buyer typed your name into their browser.

The question is how to become the kind of company that gets recommended in those rooms. You participate in the communities where your buyers spend time, not as a sales channel but as a member. You answer questions without a pitch attached. You share expertise that's useful whether or not anyone buys your product. You do this for long enough that people in the community develop a real opinion about your competence and trustworthiness.

This cannot be faked at scale, which is what makes it worth anything. The timeline is months, not weeks. The attribution is invisible. The ROI calculation will never close cleanly in a spreadsheet. And it is, for many companies, the highest-converting acquisition channel they have, which they cannot prove because the proof would require instrumenting private conversations that exist specifically because they're not instrumented.

Building the community yourself is the more interesting option. The critical design decision is whether the community is about your product or about your members' problems. Product communities tend to devolve into support forums. Nobody wakes up excited to participate in a support forum. Problem communities, where the shared identity is "we are people who deal with X" rather than "we are people who use Y," attract members who aren't yet customers, which is the whole point.

Traditional outbound works like this: you compile a list of companies that match your ICP, email all of them, and hope some percentage happen to be thinking about the problem you solve at the moment your message arrives. If 3% are actively evaluating, 97% of your outreach is noise. You're paying your sales team to generate irritation in people who aren't buying.

Intent data is an attempt to solve the timing problem. Instead of guessing which companies might be interested, you look for behavioral signals that suggest they already are. A company whose employees are reading comparison articles about your product category, visiting competitor pricing pages, and searching for terms related to the problem you solve is probably closer to a purchase decision than a company that isn't doing any of those things.

But signal quality varies enormously. Some intent data is specific and reliable: a named person at a target account visited your pricing page three times this week. Some is vague and inferential: an IP address associated with a company resolved to a page about a topic tangentially related to your category. Treating them the same way reproduces the spray-and-pray problem with extra steps and a higher software bill.

Two companies are trying to reach the same buyer. Company A spends $200 per lead on paid acquisition, nurtures each lead through a fourteen-email sequence, and converts at 2%. Company B builds an integration with a product the buyer already uses, gets listed in that product's marketplace, and acquires customers who arrive pre-sold because a tool they already trust is telling them "this works with us." Company B's acquisition cost is lower, their sales cycle is shorter, and their churn is probably lower too.

This is the basic case for partner ecosystems. But it understates the more interesting part, which is that integrations compound and advertising doesn't. When you build a deep integration with another product, you create switching costs that didn't exist before. A customer whose workflow depends on data flowing between your tool and three others has to untangle all of those connections to leave. Each new integration makes the existing ones stickier. This is a moat, and unlike most things that get called moats, it actually functions like one.

The less-discussed failure mode is that partnerships are operationally expensive in ways that don't show up in the pitch. Co-selling means aligning two sales teams with different incentives. Joint campaigns means coordinating two marketing teams with different priorities. Every partnership that works well is a relationship maintained through ongoing effort, and every partnership that stops getting that effort decays into a logo on a landing page that doesn't actually do anything.

There are two ways to build software. You can build a tool that does one thing for everyone, or you can build a tool that does everything for someone. The first approach gives you Slack: a product so general that every company uses it and no company loves it. The second gives you something like a practice management system for veterinary clinics that handles appointment scheduling, vaccine tracking, insurance billing, and the specific regulatory paperwork that veterinary practices deal with and that no horizontal tool will ever get right.

When you build for a specific industry, you absorb domain knowledge that becomes a compounding advantage. Each detail is individually small. Collectively, they constitute a moat that horizontal competitors would need years of industry immersion to replicate, and horizontal competitors will never bother, because the market looks small from the outside. It is small. That's the point.

Vertical SaaS companies command 2-3x higher valuations than horizontal counterparts because niche expertise creates defensible moats. The vertical market for SaaS is projected to exceed $720 billion by 2030. When you own a niche, your marketing becomes dramatically more efficient -- you know exactly who to reach, where they gather, and what language they use.

The risk: you are betting that your chosen niche is big enough to sustain a business and stable enough to exist in ten years. Niche selection is load-bearing. But conditional on picking a niche that actually exists and persists, the strategy of knowing it better than anyone is one of the more reliable ways to build something durable.

For about twenty-five years, "be findable on the internet" meant "rank well on Google." That era is ending. A growing share of product discovery now starts in AI interfaces rather than search engines. Someone asks ChatGPT "what's the best project management tool for a remote team of twelve" and gets an answer that never touches Google, never loads your website, and never enters your analytics. The query happened, the evaluation happened, and you were either in the answer or you weren't. If you weren't, you don't even know you lost.

Combined AI search now drives approximately 6 billion monthly visits and is growing 15-20% per quarter. Being cited by AI models is becoming as important as ranking on Google.

The tricky part is that "the AI" is not one thing. It's several systems with different preferences. The meta-strategy is less "hit these specific signals" and more "produce content that is authoritative enough that multiple systems with different criteria all converge on citing you." This is, conveniently, the same thing that worked for traditional SEO over long time horizons: be the best actual source on your topic.

The pattern you've likely been living through: a company tries content marketing for four months, gets impatient, pivots to outbound sales, runs that for three months, gets impatient, adds a product-led growth motion, gets impatient, hires a growth hacker who suggests paid acquisition, gets impatient, and eventually runs out of money while doing five things badly instead of one thing well.

Each individual pivot feels rational at the time. Content isn't converting yet, so content must not work. Outbound isn't scaling yet, so outbound must not work. The company interprets a lack of early results as a signal about strategy when it is almost always a signal about duration. They are pulling plants out of the ground every few months to check whether the roots are growing.

The uncomfortable truth is that most go-to-market motions take longer to work than the average company's patience can last. This creates a systematic bias: companies under-invest in strategies that compound slowly and over-invest in strategies that promise fast results, which are disproportionately the strategies that don't compound at all.

One motion, executed with enough depth and consistency to actually reach the compounding phase, will outperform five motions that each get abandoned during the plateau. Everybody already knows this. The reason it doesn't get followed is that it means sitting with the discomfort of doing one thing for a long time without clear evidence that it's working, and most organizations are not built to tolerate that kind of ambiguity.