The context

I was at Softonic from 2019 to 2024, first as SEO Specialist, then as Product Owner and finally as SEO Manager. Softonic is one of the largest software download platforms in the world: a portfolio of six sites (Softonic, download.com, CNET and others) with over 1 billion users a year. At that scale, a template change reaches hundreds of thousands of URLs in hours and every point of traffic is millions of visits. My job was to grow that traffic in a market where Google changes the rules every few months.

The problem

The catalog covered about 100,000 applications, a fraction of the software people search for every day. There was a huge opportunity to capture unmet demand, but writing useful content at that scale by hand was unworkable on cost, speed and quality.

The system

I designed and led a large-scale content generation program, the company’s highest-impact initiative in that period:

  • Demand detection: semantic and programmatic keyword research to map what software people searched for and couldn’t find in the catalog. A page was only created if there was real demand behind it.
  • Our own models, before ChatGPT: the program started before ChatGPT came out (November 2022), when almost nobody was producing AI content at this scale. Instead of relying on an external API, we trained our own generation models.
  • Production with quality control: every batch went through generation and review pipelines before publishing, to keep the editorial standard without people becoming the bottleneck.
  • SEO architecture: templates, internal links and structured data so every new page added authority to the whole.
  • Agile methodology: sprints, per-batch metrics and decisions made from performance data. What didn’t work in one batch wasn’t repeated in the next.

What went wrong: translations and a core update

Not everything went up. We consolidated the languages, which lived on subdomains, into the main domain. At first it worked very well and traffic grew. Then a core update hit us hard: a drop of over 20% on a site that gets millions of visits a day.

When we analysed the causes, the conclusion was that the translations, which weren’t as good as the original content, were lowering the quality of the site as a whole. So we moved the languages to their own ccTLDs, linked to each other with hreflang. With the next core update the main domain didn’t just recover: it rose above the levels it had before the consolidation. And each language, on its own ccTLD, got more traffic than it had inside the main domain.

The lesson for any large-scale content program is the same: Google looks at the quality of the whole site, and weak pages drag down the good ones. I cover it in more detail in Quality at Google: what it is, how it is measured and how long it takes to recover from a core update.

The result

In two years the program went from 0 to 20 million monthly organic sessions, and the catalog grew from 100,000 to over 1 million pages. In 2024, my last year, organic traffic for the whole portfolio (not just the program) grew 40% over 2023. And the program kept running after I left.