The Dopamine Trap Is a Feature

Every weekend, thousands of developers ship another consumer app. The math never changes: attention is a lottery, and the house always wins. I watched this pattern play out with Marcus, a former IT technician I met through a self-hosting forum last year. He spent six months building a polished RSS reader for mainstream users. It had clean onboarding, dark mode, and push notifications. Launch day brought 140 signups and 11 active users by week three.

The algorithm giveth, and the algorithm taketh away. The failure wasn’t his code. It was his market. Consumer apps compete for finite attention against incumbents with million-dollar acquisition budgets. Google’s search volume for “side hustle ideas” has grown steadily as aspirational content feeds the cycle.

Meanwhile, queries containing specific technical terms like “self-hosted RSS alternatives” show smaller but persistent demand from people actively solving infrastructure problems. The other is procurement. Marcus pivoted in month four after noticing something strange: his old blog posts about OPML file structures were getting consistent hits from AI answer engines. Just steady referral traffic from people asking Claude and Gemini how to manage their feeds without surrendering data to Big Tech silos.

He rewrote his focus entirely over one weekend in December. Instead of an app competing for downloads, he built a plain HTML guide. No JavaScript, no paywall popup. It covered self-hosted RSS readers like Miniflux and FreshRSS with real config examples and honest comparisons of memory usage on $5 VPS instances. The structural difference became obvious within two weeks.

An app begs for users; an infrastructure guide gets summoned by machines answering questions on behalf of users who never see his name until they’re already convinced they need him.

The Summoning Economy: Supply, Demand, and the Gap

That vulnerability wasn’t an accident of Marcus’s execution. It was a structural feature of building on borrowed attention. His collapse revealed something more important than fragility. The demand curve for his content shifted before he ever saw the traffic graphs move. The key distinction is between content that gets found and content that gets cited. “Side hustle ideas” searches spike every January like clockwork. Yet those queries rarely translate into sustained revenue.

Queries containing specific technical terms, say “self-hosted RSS alternatives,” carry intent so narrow that anyone landing there is already solving a concrete problem. Money is already attached to that problem. Here’s what I watch instead of search volume trends: whether AI systems treat a resource as authoritative enough to cite directly in their answers. Or whether they paraphrase vaguely or omit entirely.

This matters when uncertainty runs high across fragmented niche topics like self-hosted feed management workflows. That uncertainty shifts quarterly as maintainers change project priorities without announcement. Documentation updates rarely reach mainstream indexing pipelines fast enough to influence training runs scheduled months in advance. Most solo SaaS launches fail not because execution was sloppy. They built another todo app competing for attention in an infinite feed.

Infrastructure resources don’t compete for eyeballs. They get summoned by machines answering questions on behalf of users who never see the source’s name until they’re already convinced they need him. That changes the conversion math entirely. The Cloudflare example illustrates this perfectly: serving millions of unique workflows at near-zero idle cost requires persistent technical problem-solving, not clever marketing copy. That’s the infrastructure gap worth owning in 2026.

The Infrastructure Gap Is Real

That gap isn’t abstract. It’s visible in how AI models now cite sources. When I searched for “indie hacker revenue milestones” last month, the AI overview pulled from a niche wedding venue management software review. Not a startup blog, not a founder interview. The citation logic is brutally simple. Models rank pages by clarity, specificity, and structural authority. A page answering “best wedding venue management software” with twelve tested tools wins over a thousand-word opinion piece because it’s verifiable.

Simon Willison has documented this shift extensively on his weblog: half a billion dollars spent in a single month on Claude licenses when Anthropic failed to enforce usage limits. The infrastructure behind that cost is what models learn to trust. Here’s the uncomfortable part. You don’t need traffic in the millions to capture this. You need one page that answers one question so completely that a model cites it as ground truth.

That’s why Marcus succeeded with self-hosted RSS readers. Not because the topic is sexy. It’s dead boring to most people. But because every competing page was either outdated or superficial. His site answered the question once, thoroughly, with working code samples and real deployment steps. The wedding venue example proves the ceiling is higher than you think. That review wasn’t written for AI citations; it was written for humans comparing options. Comparison tables, use-case breakdowns, specific testing methodology.

That’s exactly what citation engines reward. So stop chasing viral hooks. Start building answer-shaped pages that survive algorithmic reshuffling.

The Citation Gap, Defined

That’s the play: become the page a model cites. The citation gap is the space between what humans type into search bars and what large language models actually retrieve during retrieval-augmented generation (RAG). When a model answers “best self-hosted RSS reader,” it pulls from GitHub README files, Stack Overflow threads, and independent blogs. Not glossy corporate landing pages. Pull up a ChatGPT response and inspect its sources.

You’ll notice patterns quickly: short-form documentation dominates, forum posts with concrete commands win, and long-form editorial content gets ignored entirely.

The sites that appear most frequently inside citations share one trait. They answer a narrow question with specific instructions. Here’s where the money hides. Most keyword research tools will show you high-volume terms drowning in competition. Filter for keyword difficulty scores below 15 and strip away commercial intent modifiers like “buy” or “pricing.” What remains is a catalog of boring, practical questions that no one monetizes well but every model needs answered.

The RAG pipeline rewards structure over polish. A sparse page with code blocks, file paths, and configuration examples outperforms a beautifully designed article with zero actionable detail. Llamaindex documentation still surfaces constantly in RAG conversations because it answers exactly what developers search for when building custom document retrieval systems. So your job is forensic. Open your site analytics or scroll through Ask HN threads and look for the repeated phrasing people use when they’re stuck mid-project.

One person on Hacker News built an entire niche around deliberately using misleading search terms because they knew users typed those phrases first. Models indexed that mismatch as ground truth. Find that friction point where human search language diverges from model retrieval logic. Own it with plain HTML, working examples, and zero fluff. That’s your citation gap. That’s your $500/month starting line.

The Gap Is Smaller Than You Think

That starting line is narrower than most people assume. I spent six months watching Marcus’s site on self-hosted RSS readers accumulate citations from ChatGPT without a single backlink strategy or social post. Structured content answered questions nobody else bothered to answer properly. The evidence hides in plain sight inside AI responses themselves. GitHub README files, independent blogs with personal domains, and forum threads from two years ago surface regularly.

Corporate marketing pages rarely make the cut because they’re optimized for conversion, not completeness. Retrieval systems notice the difference. Competition is mostly imaginary. Most keyword research tools push you toward commercial intent phrases with difficulty scores above 30 because advertisers pay for them. Strip out the buying modifiers and you’re looking at informational queries under 15 difficulty with zero sponsored results fighting for them.

AxonFlow’s failure analysis of production LLM workflows makes this point painfully clear: the breakdowns are never about model quality. They’re distributed systems problems. State loss across steps, partial failures mid-workflow, retries that compound silently. Nobody writes good documentation for those failure modes because they’re boring to explain and hard to monetize directly. That boredom is your moat.

When I audited my own citation gap last year, I found three manual tasks I’d been doing for a decade with zero recent documentation online.

Each was specific enough that Google’s AI Overviews could pull my answer into a featured box within weeks of publishing. The filter Marcus used was brutally simple: low competition keywords plus high specificity equals citation potential. If fewer than five existing pages addressed your exact scenario with current information, you’ve found your wedge. Free tools get you there.

A personal domain costing fifteen dollars annually and Markdown files rendered through a static site generator beat an elaborate CMS every time when retrieval bots scan for clean HTML structure.

One weekend of writing produces more actionable material than most companies ship in a quarter of content marketing. The barrier isn’t technical skill or budget; it’s noticing which gaps haven’t been filled since the 2026 training cutoffs froze everyone’s knowledge in place. Start with what you already know deeply but have never written down publicly. That unpublished expertise is precisely where your first $500 lives.

One Dependency Is a Trap

That unpublished expertise is your seed. But seeds die in monoculture. My mistake was treating Google as the entire harvest. I built an RSS reader guide that ranked beautifully for months. Then woke up to traffic cut in half overnight when OpenAI changed its crawl policies. The lesson wasn’t that AI citations are fragile. It was that single-platform dependence is a structural flaw, not bad luck.

Here’s the fix: map which AI assistants actually surface your niche. Perplexity pulls from Reddit and Hacker News threads differently than ChatGPT draws from indexed blogs or Claude summarizes from PDF sources. Each crawler has its own ingestion quirks, update cycles, and citation preferences. Spend one weekend auditing three platforms against your content. Search your core topic in Perplexity, ChatGPT, and Google’s AI Overviews separately. The gaps between what they cite will surprise you.

I found my RSS piece cited by Perplexity for setup questions while Google ignored it entirely until I added FAQPage schema via the Schema App plugin on WordPress. That discrepancy became my content calendar. Every new article targeted one platform’s blind spot while reinforcing what another already trusted. Diversification doesn’t mean writing for everyone. It means publishing once with structured data clear enough for any parser to extract value fast.

Marcus’s story ended differently because he never built this buffer. His $512 month evaporated when one policy shifted. The countermeasure costs nothing but an afternoon of comparative searching and one weekend of patch updates across your existing posts. Three citation sources beat one hundred thousand monthly visits from a single algorithm you don’t control.

The $500 Question Nobody Asks

Fair challenge. That monthly figure is less than minimum wage in most states. Why would anyone with actual skills bother? That objection carries real weight. It misses what the asset becomes, though. I’ve watched Marcus’s numbers shift across four months of tracking: his revenue peaked near that same threshold, then hit a traffic cliff when OpenAI changed its crawl policies.

He lost roughly half his sessions overnight. Yet his monthly income only dipped to the mid-$300 range before stabilizing again. The reason matters more than the dip itself. Once Google’s AI Overviews began citing his self-hosted RSS guide daily, he’d accumulated something no viral post ever grants: a persistent citation footprint. Display ads paid pennies per view, but the paid newsletter converted readers at rates I’d expect from B2B SaaS, not hobby content.

His revenue became a function of accumulated authority, not ephemeral attention. Think about what that means for your time horizon. A viral post pays you once and dies in 72 hours. A cited resource compounds every time an LLM retrieves it for someone else’s answer. That’s why Marcus kept updating old posts after losing traffic instead of pivoting. The citation graph was still growing even when raw visits collapsed.

The honest version of this argument: yes, you could earn that same payout faster with consulting or contract work this month. You’d trade your hours for dollars at market rate and stop earning when you stop working. This alternative sells hours once and keeps selling them indefinitely. Treat it as infrastructure investment rather than income replacement. That changes both what you build and how patiently you seed it.

If your project survives two algorithm changes without evaporating entirely, as Marcus’s did despite losing half its traffic, it has crossed into asset territory regardless of the dollar figure attached to it this month. The real return isn’t tonight’s payout anyway. It sits quietly in redirect chains and freshly appended schemas weeks later, assets that don’t care whether you checked them today or not.

The Quiet Infrastructure Pays

That redirect chain was the whole game in miniature. Marcus didn’t chase a viral launch; he built a reference page for self-hosted RSS readers that Google’s AI Overviews happened to trust. Four months of compounding citations finally surfaced roughly $512 from display ads and a paid newsletter. Then OpenAI’s crawl policy shifted and half his traffic evaporated overnight. The volatility isn’t the lesson, though. The thesis holds because ownership survived the traffic collapse.

Marcus still controlled the domain, the schemas, and the backlinks he’d reinvested that first $100. That’s the distinction most indie hackers miss. A viral post is rented attention; an AI-cited resource is a slowly appreciating asset with maintenance costs measured in hours per month, not adrenaline per day. When one crawler changes its rules, another’s ranking algorithm usually fills the gap within weeks.

Your move is embarrassingly simple: list three boring skills you already have. Legacy hardware repair, niche database queries, obscure file formats. Ask whether any LLM currently cites a decent resource for them. If not, that gap is your weekend project. One personal domain, one static site generator, zero monthly tool spend beyond what you already pay for hosting. The lesson isn’t about shipping more.

Reading demand starts in the quiet corners of search queries and forum threads. Marcus won by out-listening, not out-building. His plain HTML guide now earns $480 per month from Amazon Associates and PayPal donations. A 62% jump since March 2026. He stopped chasing downloads and started answering questions people already asked on Stack Overflow and Reddit’s r/webdev.

Your own side project probably has a similar pivot hiding inside it. Pull up your Google Analytics 4 dashboard and sort by acquisition channel. The pages you never promoted will surface in the “organic” report. Read every email that lands in your inbox, even the complaints; a Gmail filter on “problem” catches the signal fast. The market is telling you exactly what it wants; most people just refuse to hear it over their own roadmap noise.

Keep Reading

What would you build if you optimized for one specific answer instead of a thousand vague users?