USAgentLeads is a contact database of licensed US real estate agents, sold as per-state and national CSV exports with an API. It reached the same monthly revenue as Foreclosure Data Hub on roughly 130 indexable URLs rather than 2,100, which is the most useful single fact about it: page volume was not what produced the outcome.
The site is still live at usagentleads.com, so everything described below can be checked against the running implementation.
Each section names the checklist group it belongs to. All 112 checks are free to read, with a verification command for every one.
A free searchable directory lived at one path per state, and the commercial page lived at another. They targeted different intents and did not compete. The directory earned links and ranked for informational queries; the commercial page converted.
Collapsing the two into one URL is the obvious thing to do and it would have forced one page to serve two intents. It would have lost both.
The 50 state pages carried over 1,800 lines of hand-written per-state content. Each state had a unique market description, five named cities with a note on why those markets get segmented first, the real state licensing body deep-linked to its official licence search, a state-specific note describing the verification and dedup process, outreach angles, and extra FAQs.
Every field had a fallback generator, so a state added without hand-written copy still rendered coherent, non-empty sections rather than blank blocks. This is the detail most implementations skip, and then they ship a page with three empty sections on it.
Templates supply structure. A data file supplies substance. If the only variable in the template is the place name, the result is duplicate content with extra steps.
Rather than repeating the head term in every heading, each section targeted a different long-tail variant: the broader non-branded phrasing, the market description with the licensing authority, the sourcing methodology, the city-level breakdown, the persona question, the outreach angles, and the field-level specifics.
Keyword variation across headings did more here than keyword density ever did.
A version-controlled keyword map recorded one row per important URL: primary keyword, secondary keywords, search intent, and internal-link role. Each URL owned exactly one primary target.
It also recorded explicit cannibalisation rules about which phrases belonged to commercial pages and which to editorial, and it fixed the anchor text for recurring internal links so every page linked with the same phrasing.
This is the cheapest artefact in either repository and probably the highest return. It prevents the most common cause of stalled growth on content-heavy sites: two of your own pages splitting the signal for one query. No crawler will ever detect it for you.
Twelve schema types against Foreclosure Data Hub's forty-two, at no measurable cost. What mattered was completeness on the types that actually carry rich results.
Every offer carried the validity date, availability, seller, shipping details and return policy. For a digital product these have to be declared explicitly as zero cost and zero days, and they are exactly the fields Search Console flags as missing on product results. The flag is easy to dismiss as noise until the result stops showing.
Live counts were passed into the schema factories, so the structured data and the visible text could never disagree.
A visually distinct answer block near the top of every state, comparison and guide page, with a documented contract: two to four factual sentences that stand alone without the surrounding page, so a featured snippet or a language model can lift them verbatim.
A state page's answer block packed the entire buying decision into one liftable paragraph: record count, email count, phone count, price, format, columns, guarantee and refresh date, all from live data.
This is the highest-return answer-engine change available to most sites and it is an afternoon of work.
Four constants drove every date and count on the site: the current year injected into titles, a data-refresh label evaluated server side, a manually bumped content review date feeding the sitemap, and a per-post updated date from frontmatter.
The review date was deliberately manual. An auto-updating review date is a lie with a timestamp on it.
Titles said the current year without anyone editing 50 files each January.
The footer linked all 50 state pages from every page on the site, so no commercial page was more than one hop from anywhere.
A related-states block was driven by a hand-built adjacency map of real geographic borders rather than an alphabetical fallback. Geographic relatedness is a genuine relevance signal and it matches how buyers actually expand territory.
Archive and taxonomy pages were set to no-index but still followed, which keeps the internal link graph that helps discovery while removing near-duplicate thin pages from the index.
Published on purpose. A case study that lists only wins is an advertisement, and these are the traps that recur on every site.
Everything above is described in the 112-check playbook, all of which is free to read. The agent skill is the version your coding agent runs: it audits your site against the same checks, reports what is missing, and implements the gaps one at a time.
Product analytics are optional. On the skill page, X advertising measurement is also optional and shares your visit and ad identifiers with X. Both are off until you allow them. Privacy policy.