Why we built our own AI Visibility tracker (and what every other tool gets wrong).
Most “AI mention” tools are screenshot scrapers wrapped in a dashboard. We rebuilt the stack from the model layer up — here's why that matters and what we learned.
When we started SEONova, the biggest internal debate was whether to build AI Visibility from scratch or wrap one of the existing third-party APIs. We ended up rebuilding the entire stack ourselves. Eighteen months in, I want to walk through why — because most of the tools in this space made the opposite choice, and I think it shows.
The screenshot-scraper problem
The dominant pattern for AI-mention tracking right now is: spin up a headless browser, log into ChatGPT (or Perplexity, or Gemini), type the prompt, screenshot the answer, OCR the screenshot, regex out brand mentions. Ship it as a dashboard.
This works, sort of. It also drifts constantly. The UI changes; the OCR misreads; the login session expires. Worse — the tracked "response" is whatever happens to render in a logged-in browser session, which is personalized to whoever's account did the scraping. The number on your dashboard is a number, but it's not the number that means what people think it means.
What we built instead
Every AI Visibility query in SEONova goes through the model provider's actual API, not a scraped browser. Account-less. History-less. No memory, no system prompt, no past behavior bias. The response you see is the response a brand-new user asking that exact question would get on that exact engine.
- ChatGPT via OpenAI API, both gpt-5 and gpt-4o, web-search enabled and disabled tracked separately.
- Perplexity via Sonar API, with full citation parsing so you can see which sources the model cited for each answer.
- Gemini via the Google AI API, both 2.5-pro and 2.5-flash.
- Claude via the Anthropic API, sonnet-4.6 and opus-4.7.
Why this matters for share-of-voice math
Personalized scraped responses produce inflated, misleading SoV numbers for whoever owns the scraping account. A clean API path produces the number that actually correlates with real-world buyer recommendations.
The lesson, generalized
If your data layer is built on screenshots, your conclusions are built on noise. Most AI-mention tools chose convenience over correctness. We chose the harder path. I'd choose it again tomorrow.
Written by
Lena Park
Co-founder & Head of Research
Built the AI Visibility engine. Spends too much time reading model release notes and yelling about prompt drift.
Like how we write? Try what we built.
Same team, same opinions, same rigor — in a platform that tracks AI visibility, rankings, content, audits, and backlinks on the same dashboard. 14-day trial, no card.