For four years, tienequehaberalgomas.com worked as a content site: a podcast with close to 200 episodes, a blog, a weekly newsletter and a growing pile of free resources. It did the job it was built for, which was gathering an audience of Hispanic professionals who want more from their careers.
Then the business underneath it changed. With Sourced, my B2B talent service, the bottleneck was never demand for content. It was supply of qualified, pre-vetted candidates. I was sitting on the perfect top of funnel and running it as a magazine.
This is the story of turning that site into a talent-capture system between June and August 2026, and of the variable that made this project different from anything I had built before: I did not write the code. An AI agent did, and my job became directing, reviewing and deciding.
The asset I was sitting on
Before describing what was broken, it is worth being precise about what was working, because the whole project is an argument about conversion rather than about reach.
| Asset | Scale | Source |
|---|---|---|
| Newsletter list | 17,779 subscribers | Kit, 746 campaigns since May 2023 |
| Email engagement | 29% weighted open rate across 2026 sends | Kit, 966k deliveries YTD |
| Search visibility | 88,100 impressions per quarter | Search Console, Apr to Jul 2026 |
| Job board demand | 11,527 clicks on listings, Mar to Jul 2026 | Job board click log |
| Podcast | Close to 200 episodes, top 5% of Spotify globally | Spotify for Podcasters |
| Press | La Nación, Infobae, Telefé, El País |
None of that was reaching the business.
Challenge
Google was already showing the site 88,100 times a quarter and sending 1,020 visits. That is a 1.2% click-through rate against a 2 to 4% norm. The audience existed, the demand existed, and the machine between them did not.
Three obstacles sat underneath the numbers:
- No funnel. Resources existed, but pages did not push anywhere. Dead ends everywhere.
- A legacy build. Four years of WordPress plus Elementor had accumulated roughly 130 pages, inconsistent styles, a page builder that would not even load, and a mobile experience that ranged from mediocre to broken.
- No line to the business. The one thing Sourced needed most, CVs from senior Hispanic talent, was not being asked for on any page.
How do you convert an audience asset into a business asset without breaking the thing that made the audience in the first place?
Solution
1. Treat staging as the laboratory and production as sacred
I am technical enough so the first decision was structural: all experimentation happens on a staging copy, and production only receives verified changes.
That decision cost a week before it paid anything. Elementor would not load, the editor spun forever, and I could not edit my own site. We backed up with UpdraftPlus, built staging on Bluehost, tested in safe mode, retired Slider Revolution and replaced the abandoned podcast widget plugin. Slow, unglamorous, and the precondition for everything after it.
2. Rebuild the architecture around intent instead of format
The old site organised content the way a CMS thinks: posts here, pages there, episodes in a plugin. The new one organises it the way a visitor arrives:
- Empezá acá: four guided routes by situation (change jobs, improve your CV, work remote, stop burnout)
- A podcast hub with every episode filterable by audience: entrepreneurs, freelancers, tech, career growth
- Resources by category: guides, market reports, templates, tools
- A job board of verified remote listings, with a Sourced CTA between every few rows
- A permanent CV capture CTA: “Tu CV puede trabajar mientras dormís”
Then I did the part nobody enjoys. I went through the roughly 130 accumulated pages by hand and cut production down to 39, with redirects for what disappeared.
3. Build one visual language and let it repeat
The design system, called Amanecer, runs on a dawn metaphor: indigo into periwinkle into champagne, with orange accents. One gradient grammar repeated across every surface.
Getting there took several rounds of me rejecting things. Out of it came 130 illustrations produced in-house: 112 blog and episode covers plus 18 resource covers. 80 are live and 69 are assigned as featured images, across 18 resource pages, 11 posts and 40 episodes. Stock imagery went to zero.
4. Gate every resource with its own form
Each of the 18 resource pages keeps its own email form, wired to deliver that specific resource. A single generic form would have been faster to build and would have destroyed both deliverability and attribution, so each page carries its own form ID and its own delivery. The lead magnet library itself grew from 6 forms built in 2025 to 9 more in 2026, so the pages had to be a system rather than one-offs.
The 18 resource covers. The cover colour encodes the type before you read the title: indigo for a guide, ink for a report, periwinkle for a template, orange for a tool.
Late in the project I generalised the treatment across 15 resource pages in a single pass: rounded panel, full-width CTA, and the page’s own featured image injected above the form. The interesting part is what stopped it from being a disaster. 7 of those 15 pages already had a custom mockup, so the code checks whether an image is present and leaves those alone. A blind rollout would have stacked a second image on top of seven pages.
5. Run the audits nobody runs
With the structure live, we swept the site the way you would sweep somebody else’s:
- Conversion: all 39 published pages checked for a CTA before the footer. Result: 39 of 39, zero dead ends.
- SEO: the skeleton was healthy and the meta descriptions were empty across essentially the whole site. We wrote and applied 39 unique ones, shortened 3 truncated titles and cleaned duplicate H1s.
- Performance: initial load around 2.6s with deferred JavaScript already in place. The gap was that zero images were being served in WebP: the optimisation had been configured and never triggered.
- Job board: the board was painting all 3,815 listings at once. It now renders 25 at a time with infinite scroll.
- Design parity: production compared against staging element by element until the diff came back empty. The job board banner had a 122px white gap under the header, now 10px.
What broke: the caching plugin that ate the design
The most instructive failure of the project was infrastructure.
Midway through, the mobile site looked destroyed. The menu rendered as a raw bulleted list, backgrounds vanished, spacing exploded. We spent hours chasing design bugs that measurement said did not exist: the pages were structurally perfect and simply looked wrong.
The cause was the caching plugin’s “Unused CSS” feature deleting stylesheet rules it had decided were unused, including the entire responsive header, because those rules only activate through JavaScript and media queries that its static analysis could not see.
Two things came out of that week. Measure before you fix, because every obvious repair before the measurement was patching a symptom. And your stack has layers you do not know about: a WordPress site’s appearance lives in the page builder’s data, the theme customiser, the design system’s global tokens and whatever the optimisation plugins do to all three. They fail independently.
There was a smaller self-inflicted one worth recording. While migrating I checked whether the new URLs existed before creating the pages, and the CDN cached those 404s for four hours. Order of operations matters even for read-only checks.
Migrating to production without a full push
When staging was approved I refused to push the whole site, because production had the curated 39-page structure I had cleaned by hand and a push would have resurrected the graveyard. The migration was additive and reversible by design:
- Full backup, database and files, verified before anything was touched
- Every design change shipped as a discrete code snippet that can be disabled with one click
- Homepage content applied through the page builder’s own save mechanism, old version preserved in revision history
- Every change verified against staging immediately after applying
The four-layer lesson came back here. Copying page data was not sufficient. Typography tokens, the global palette and the custom CSS lived in separate systems and had to be migrated separately. The site looked almost right until all four matched, and almost right is the most dangerous state, because everything works except the feeling.
Impact
Qualitative impact
- The archive became navigable. Close to 200 episodes went from a plugin listing to a filterable hub
- The site reads as one brand across posts, resources and episodes for the first time
- Mobile went from broken to designed, after three deliberate passes
Strategic impact
- The content operation and the recruiting business are now connected: the same traffic, the same posts, with every path leading somewhere that matters to Sourced
- Production became safe to change. Reversible snippets and a verified backup mean speed stopped being a risk
- The design system is reusable, so new pages cost hours instead of days
- The measurement layer is now the constraint rather than the build, which is a much better problem to have
Key performance indicators
Shipped and verified
| Indicator | Value | What it measures |
|---|---|---|
| Funnel coverage | 39 / 39 pages | Every published page ends in a resource, a signup or a CV |
| Page inventory | 130 to 39 | Index bloat removed, with redirects in place |
| Resource capture | 18 forms | One dedicated form per resource, for delivery and attribution |
| Search surface | 39 meta descriptions | Written from zero, each with a value proposition and a CTA |
| Visual system | 130 illustrations, 80 live | Owned library replacing stock across the site |
| Job board render | 3,815 to 25 rows | Paint cost per page load |
| Rollback surface | 1 click per change | Every design change ships as a toggleable snippet |
What this taught me about building with an AI agent
The agent wrote the CSS and PHP, edited the page builder programmatically, ran the audits and measured every change. I supplied the two things it could not: taste, and access, because logins stayed human throughout.
What it was genuinely better at:
- Verification loops. It measured before and after and refused to trust its own changes without confirming them live. I would never sustain that discipline by hand.
- Diagnosis. The caching bug, the four-layer mismatch, a button stretched by a stray flexbox rule, all found by comparing computed values instead of staring at screens.
- Volume. Thirty-nine meta descriptions, 130 illustrations, 15 resource pages rebuilt in one pass, dozens of CSS blocks and full-site audits, in days rather than weeks.
What still needed me:
- Taste. It could confirm two pages were identical. It could not tell me the hierarchy felt inverted or a button felt like it was shouting.
- Safety. Backups verified before any production change, rollback plan in place before the first snippet shipped.
- Judgment under trade-offs. Deferring risky performance work because it could break the job board’s forms was a business call.
The rhythm that emerged: the agent proposes and verifies, I look and decide, and memory carries context between sessions. When execution gets cheap, the founder’s job concentrates into the two things that were always scarce, which are knowing what good looks like and deciding what is worth doing.
Up next
The machine is built and now it needs fuel and measurement. Meta descriptions for the 48 blog posts. Structured data so the salary survey becomes citable by AI assistants, because the AEO opportunity in the Hispanic career niche is still wide open: on a fixed set of ten test prompts, the site currently surfaces in 0 of 7 generic career questions. Conversion tracking, so the next round of decisions comes from data instead of intuition.
And two loose ends I already know about: 2 of the 18 resource pages carry forms that are not wired to the email platform and may not be capturing at all, and 12 of the 18 are still missing an H1.