How Softonic fixed a decade-old AI reputation problem in 4 months

Who should read this: brand, SEO and communications teams at companies with a long history online, where AI answers still lean on things that stopped being true years ago.

TL;DR

Softonic was being flagged by AI models as a risky download source, even though the business had spent years cleaning up its catalog and trust signals. Within four months of adopting LLM Pulse, they identified the specific prompts driving that perception, built a targeted content and PR plan around them, and lifted positive sentiment by 7pp across the tracked non-brand prompts.

About Softonic

Softonic is one of the largest software discovery platforms in the world, serving more than 100 million users a year across 20+ languages. Their catalog covers everything from open-source utilities to commercial productivity apps, with editorial reviews and safety scans behind each listing.

The problem: an AI sentiment gap they could not see

Traditional SEO and brand monitoring tools were telling Softonic everything was fine. Organic traffic was healthy, branded search was stable, and review sites showed strong scores.

Then the team started spot-checking ChatGPT, Perplexity and Gemini, and the picture was very different:

• "Is Softonic safe?" returned hedged, cautious answers in 3 out of 5 models.
• Long-form prompts about software recommendations frequently included disclaimers about historical issues from a decade ago.
• Competitors with smaller catalogs were being recommended over Softonic in 47% of head-to-head prompts.

They needed a way to quantify that gap, decide what to fix first, and measure whether the fixes were working.

What they did with LLM Pulse

Step 1: Map the prompts that actually matter

Using Prompt Tracking, the team set up 450 high-intent prompts in English and Spanish, the two languages driving most of their AI search risk. Each prompt runs weekly across ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews.

After the first execution cycle they had a ranked list of the prompts where Softonic had the lowest visibility, the most negative sentiment, or both.

Step 2: Use sentiment tracking to find the narrative

Sentiment Tracking broke each mention of Softonic into three things: polarity (positive, neutral or negative), the topics behind it (safety, ads, bundling, app quality), and the sources the model was citing as evidence.

That is where it clicked. 78% of negative sentiment came from old forum threads and a single 2014 incident report that AI models kept retrieving through outdated citations.

Step 3: Act on the citations and the content recommendations

From there the playbook wrote itself:

• PR and outreach to refresh the cited sources where possible. More than 95% of them were over 10 years old.
• Content rewrites of the 22 highest-impact pages flagged by Content Recommendations, focused on safety, transparency and trust signals.
• New canonical landing pages for the questions where AI models had nothing current to cite.
• GEO Testing to A/B test layout and trust-signal placement.

Step 4: Measure it every week

Every Monday the team reviews visibility per prompt cluster (Brand Visibility and AI Visibility Score), the sentiment shift week over week, new citations earned by the rewritten pages, and Share of Voice against their top four competitors.

The results after four months

Net Sentiment Score improved by 7pp in four months. Someone asking an AI model about Softonic now gets a picture of the company as it is today, instead of one shaped by a decade-old incident report. The team runs the same loop every week and is still working on it.

"Understanding how LLMs perceive our brand is no longer optional. It is the next frontier of a modern SEO strategy. LLM Pulse provides the deep insight we need to see exactly what influences our AI visibility. It has turned our feelings into a roadmap of actionable steps, ensuring our content is perceived accurately across every major model."

Ferran Gavin, Director of Catalog and Traffic, Softonic

Why it worked

• Multi-model coverage. Softonic stopped optimizing for one model and started optimizing for the answer they want users to get, whichever AI is asked.
• Prompts as the unit of work. Instead of a vague "improve brand perception" goal, they had a finite list of prompts and a clear definition of done for each one.
• Sentiment and citations together. Polarity alone would only have confirmed there was a problem. The citations told them which sources to go and fix.

Run the same play

If your brand has been online for a decade, AI models may still be answering with sources you stopped recognizing years ago. Start a free trial to see which prompts and which sources shape your answers today, or talk to us if you would rather walk through it with someone first.