How we research, build, launch and rent local lead generation sites

The average job has to be large enough that one lead is genuinely valuable to a contractor. If a job is worth a few hundred dollars, no contractor will pay a meaningful amount per call, and the rent will never justify the build.

This one filter alone kills most of the obvious ideas: gutters, fencing, drywall, railings, gates. Clean search results, no competition, and completely unrentable.

The searcher has to be a homeowner ready to hire, not a researcher and not a commercial buyer. We use advertising cost per click as a proxy: if advertisers are not paying real money for the keyword, the intent is weaker than it looks.

This filter is what separates a remodeling niche from handyman territory. Handyman work has volume but no urgency and no ticket.

The most important filter, and the one most people skip. Several home service categories have been rolled up by private equity or covered by national franchise networks. When that has happened, the top results in every city belong to the same brand, backed by an authority profile you cannot outrank as a new site.

The test is simple: search the niche in three to five different mid size cities and see whether the same national name keeps appearing near the top. If it does, drop the niche. Whole categories are effectively closed this way.

Watch for the franchise variant too, which is subtler. Some networks build genuinely local, content rich city pages maintained by a real franchise owner. Those behave like local competitors, not like generic corporate pages, so they cannot be dismissed.

A fast manual look at the search results before running any tooling. What you want to see is small local businesses with amateur websites. What you do not want is directories, aggregators and national brands filling the page.

If searching the niche in your target city returns results from a different city or state, the local search volume is too thin. Google is padding the page because it does not have enough local supply. You can rank in a market like that and still get no calls.

A large share of AI citations come from the first portion of a page's text. Clear heading hierarchy, question shaped subheadings, direct answers at the top. Structure is doing more work here than word choice.

The most useful and most uncomfortable finding: AI search shows a systematic preference for third party sources over brand owned content, plus a bias toward large recognised brands. A brand is substantially more likely to be cited through someone else's page than its own.

For rank and rent that is significant, because it means the site alone will not get you cited. You need a consistent off site footprint saying the same things: directory listings, reviews, forum presence, video, industry mentions. Our own directory listing sprint across the network was done for exactly this reason, not for the referral traffic.

Recently updated pages get cited materially more often than stale ones, and pages left untouched for long periods lose citations they previously held. A visible update date plus a real refresh cadence is a cheap lever most local sites ignore.

We put a single JSON-LD block on every page describing the business: name, address, phone, hours, coordinates verified from the map rather than guessed, price range, service area, and a service catalogue. Organization and breadcrumb markup support entity understanding.

The correction worth passing on: FAQ and how-to markup no longer produce rich results in Google. They are still valid markup and may still assist AI extraction, so we keep them where the content genuinely is question shaped. But if you read a 2025 guide telling you FAQ markup is a top tier tactic, that guide is out of date.

Non negotiable rule: structured data must match what is visible on the page. Mismatches invite trouble.

Check robots.txt actually permits GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Google-Extended. Trivial to verify, easy to get wrong, and it silently costs you the entire generative surface.

On llms.txt: the evidence that it affects citations today is weak. Adoption is low and no major provider has confirmed using it. It costs minutes, so it is harmless to add, but treat it as speculative rather than as a tactic.

Loading, responsiveness and layout stability, measured on real mobile traffic. A page group is judged by its worst metric. Treat this as a tie breaker rather than a primary lever: worth passing, not worth obsessing over. Static single file HTML passes comfortably without any special effort, which is one of the quiet advantages of the architecture.

Defensible as a decision, expensive as an outcome. We have ten built assets and one signed tenant. The ranking and monetization work is harder and slower than the build work, and we discovered that after committing all the build effort. If starting again: two sites, one tenant, then scale.

Without a verified business profile a site cannot enter the map pack, which is a large share of local clicks, and cannot collect reviews. We built the whole network knowing this abstractly and not weighting it properly. There is no honest workaround: the route runs through a real contractor verifying a real profile, which happens after you have a tenant.

The largest single blocker across the whole project has been business verification with the telephony provider, which gates message delivery and, at times, outbound calling. It is a regulatory requirement, not a support ticket, and it ultimately depends on having a registered business entity in the right jurisdiction. Solve this early, before it silently breaks something you have already promised a client.

We verified mobile layouts three times through code review and live page fetches and passed each time, while real bugs remained. Entire layout columns were missing from the markup, so there was no broken image to find. Render the page at real device dimensions and look at it. We eventually built the tooling to do that, and it caught a bug on its first run.

Our task list has repeatedly told us a blocker was cleared when it was not, and that a market was unworked when it had been worked. Twice we acted on the wrong version. If you keep a written system of record, treat correcting it as real work, and verify status against the live system rather than against your own notes.

The subtlest one. Two capabilities were recorded as proven based on tests that never left our own infrastructure: a call from one of our numbers to another of our numbers, and a message from one of our numbers to another. Both passed. Both real world equivalents failed. A test that does not cross the boundary you are worried about is not a test.

Ten sites of essentially unedited AI copy went live, and the cleanup took longer than the writing did. Thousands of stylistic tells, near duplicate area pages, filler vocabulary throughout. Write it generated, ship it edited.

By elimination, against five pre-screening filters: ticket size, genuine homeowner buying intent, absence of national franchise or private equity consolidation, cleanliness of the search results, and real market depth. Failing any one drops the niche. Around sixteen niches went through this and three survived. The franchise consolidation check is the one most people skip and the one that most often decides the answer. See section 02.

Two stages. First a demographic shortlist: population band that has volume without saturation, household income high enough to afford the work, owner occupancy and single family housing share high enough that the service is even relevant, and a preference for building geographic clusters rather than scattering. Then the search results decide, on four questions: is there an authority ceiling we cannot pass, which top results are genuine competitors versus pages that can be outranked, does someone already own the keyword domain, and would second or third place still be a viable business. We look for four viable cities before committing to a niche. See section 03.

Content is a homepage carrying the main city keyword, ten service pages, ten surrounding area pages, and informational articles for topical authority, all internally linked so the homepage accumulates the authority. Generated with AI, then edited hard: no em dashes, no filler vocabulary, and every area page rewritten to open with something verifiably local, because ten pages differing only by town name is a doorway pattern.

On the AI answer layer, four things earn their place: front load the answer because citations concentrate in the top portion of a page, keep structured data clean and matching the visible content, keep content genuinely fresh, and build a consistent off site footprint. That last one is the important one, because AI search demonstrably prefers third party sources over your own website. Also worth knowing: FAQ markup no longer produces rich results in Google, so older guides overrate it. See sections 06 and 07.

Claude for building and writing and auditing. Obsidian as the knowledge and state layer, which matters more than it sounds. Cloudflare for domains, DNS, hosting, inbound email routing and bot protection. Twilio for a tracked local number per site with voicemail, recording, transcription and forwarding. Brevo for transactional email. Google Search Console for the data that decides when a site is ready. Ahrefs for site audits. A paid search results tool for market validation. Playwright plus a custom shell script for QA. Full table with costs in section 08.

The honest answer to "what makes this work" is not any of those tools. It is that the process is written down as playbooks and checklists that get read at the start of every session, by a human and by an AI assistant. That is the actual compounding asset.

The first site in a new niche: two to four weeks, most of it infrastructure and template design rather than pages. Later sites in the same niche: a few days, dropping to a single focused session once the template is mature. Sites nine and ten in our network, about twenty five pages each with their own phone numbers, were built and deployed in one session.

Then be honest about the second number: from live to genuinely finished, add several more weeks whatever the build took. Content de-templating, mobile bugs, duplicate URLs, structured data repairs, directory listings, indexing chases. Sites that went live in June were not really finished until well into July. See section 13.

One template family per niche, built from scratch once, then replicated within that niche. Three niches, three visually and structurally distinct templates. Copied nearly unchanged: the serverless script, deployment config, CSS system, page skeletons, structured data pattern, QA script. Rewritten every time: all copy, service and area lists, coordinates, images, brand, phone, address, every meta tag.

Deliberately not one template for all ten, because near identical sites across cities is exactly the pattern Google's spam guidance names. And the price of cloning is that every template bug becomes a bug in every copy. We have fixed the same defect across ten sites more often than I would like to admit. See section 14.

The process, yes. The sequencing, no. Building the full network before proving a single site could be rented put all the effort into the half of the model that turned out to be easier. The build is mechanical. Ranking is slow. Persuading a contractor that a stranger offering free leads is not running a scam is the actual hard problem, and it is the one you should test first, with two sites rather than ten.