The Open Directory, Reimagined for AI Search

A Feasibility and Business Plan for an AI-Era Web Directory

I. The Opportunity

In the 1990s, the Open Directory Project (DMOZ) became an authority signal: Yahoo used it to populate a hand-curated web directory, and early Google gave real weight to sites listed there because a human editor had vouched for them. AI answer engines now play a similar gatekeeping role, and the same gap exists again.

• Citation concentration is extreme — current research on AI answer engines finds that a small handful of domains, chiefly Reddit, Wikipedia, YouTube, and LinkedIn, account for the large majority of citations across ChatGPT, Perplexity, Gemini, and Claude, leaving most legitimate businesses, nonprofits, and niche experts effectively invisible in AI-generated answers.
• There is no verified trust layer for AI the way DMOZ once was for search; volunteer editors used to vouch that a site was real, correctly categorized, and legitimate, and no equivalent structured, human-vetted registry exists for today’s answer engines.
• llms.txt is a map, not a directory — the emerging llms.txt convention only lets a site describe itself to a crawler; it carries no independent third-party authority, and as of 2026 no major AI lab has publicly committed to acting on it in production.
• GEO spending is already flowing into an unproven category, as brands react to falling click-through and rising zero-click AI answers by paying for optimization services; that budget is actively searching for a credible, structural answer.
• The underlying need is unmet: a neutral, categorized, independently verified registry of who exists, what they do, and how to confirm it is precisely the kind of third-party, structured source that answer engines are built to prefer over marketing copy.

II. How It Would Be Structured

• A hierarchical taxonomy organized the way DMOZ was — industry, sub-industry, specialty, and locality — but designed from day one around entity resolution rather than link browsing.
• Structured data comes first: every listing is published simultaneously as a human-readable page and as schema.org/JSON-LD markup, so machines can parse it without inference.
• A verified editorial layer combining paid category editors with vetted volunteer editors, mirroring DMOZ’s model, confirms accuracy, ownership, and category placement before anything goes live.
• An open, licensable feed — a free bulk data download plus a paid real-time API — explicitly invites AI labs, researchers, and startups to ingest the dataset, the same way DMOZ’s RDF dump once seeded early search engines.
• Citation-ready summaries on every listing: a short, fact-dense paragraph written specifically to be lifted cleanly into a generated answer, rather than marketing prose an AI has to rewrite.
• Visible provenance signals — last-verified date, responsible editor, and a public changelog — give both AI systems and human reviewers a confidence signal that current search results and social platforms do not offer.

III. The Business Model

• Freemium verified listings: a basic listing is free to maximize coverage, while a paid annual Verified/Premium tier adds priority review, richer schema fields, and monthly re-verification.
• Data licensing to AI companies — a paid bulk feed and API sold to AI labs, GEO tool vendors, and enterprise data platforms, comparable to how stock-media libraries license structured content to model builders.
• A citation-monitoring subscription as a SaaS add-on, showing listed businesses where and how often they are cited across ChatGPT, Perplexity, Gemini, and Claude, cross-sold to every paying listing.
• Disclosed category sponsorships placed clearly apart from organic listings, functioning like legacy directory advertising without compromising the trust the directory depends on.
• Affiliate and referral revenue from verified traffic sent to partner services in categories such as legal, financial, home services, and healthcare.

IV. Cost and Team to Launch

• A lean six-person launch team: one editor-in-chief, two category editors, two engineers, and one business-development lead is enough to build and seed the first vertical.
• A first-year budget near $650,000 covers payroll, editorial tooling, hosting, legal review of licensing terms, and hand-verification of the first 20,000 to 50,000 listings.
• A volunteer editor network modeled directly on the original Open Directory Project keeps category coverage growing without a payroll expense — the real cost is vetting time, not cash.
• Infrastructure is the cheap part: hosting, database, and API costs run a few thousand dollars a month at launch scale, since the build is closer to a structured content system than a search index.
• Legal costs are not optional: budget real money for licensing agreements, data-use terms, and defamation and liability review, because deciding who counts as authoritative is inherently a publishing decision with legal exposure.

V. Revenue Potential

• Premium listings at scale: 25,000 premium listings at $300 to $600 a year is $7.5 million to $15 million in annual recurring revenue once the directory reaches meaningful scale, realistically in year two or three.
• Data licensing is the bigger prize: a small number of AI labs and GEO platforms paying five to six figures a year each for a clean, licensed feed could match or exceed listing revenue with far fewer accounts to manage.
• Citation monitoring compounds revenue: cross-selling a $50 to $150 monthly tracking product to even 10% of paying listings adds meaningful recurring revenue with almost no incremental acquisition cost.
• Year-one revenue will be modest, realistically low six figures while trust and coverage are being built — this is a multi-year asset play, not a quick return.
• The real value is the dataset itself: the most likely path to a large payoff is acquisition of the verified corpus by an AI lab, data broker, or search company, not subscription revenue alone.

VI. Where to Start

• Pick one vertical first — launch inside a single underserved, high-intent category such as licensed contractors, medical practices, or B2B software vendors, rather than attempting broad coverage the way the original ODP did.
• Build the schema before the front end: get the JSON-LD data model and verification workflow right first, since the machine-facing feed matters more at launch than a polished consumer website.
• Recruit founding editors publicly through open applications, transparent editorial guidelines, and public credit — exactly as DMOZ did — to build a volunteer base cheaply and legitimately.
• Seed the corpus by hand: personally verify the first 1,000 to 2,000 listings before any self-serve submission opens, to set an unmistakable quality bar from day one.
• Publish the feed immediately once a defensible core dataset exists — release a free bulk download or lightweight API right away so AI crawlers and researchers can start pulling from and citing it, instead of waiting for a “finished” product that never arrives.