Vidima AI

Vidima tracks your AI visibility across ChatGPT, Perplexity, AI Overview and Gemini. Then it hands you a plain-language fix list to show up more, in any language your buyers ask in.

Comment

Inspiration

I kept telling friends who run small businesses that they needed to think about AI search. They usually shrugged, most of their customers came through word of mouth.

Then I tested the questions their buyers might ask:

„Кой е най-добрият доставчик на рекламни материали в България?“
“Who is the best supplier of promotional products in Bulgaria?”

Sometimes the AI recommended their business. More often, it recommended a competitor. The owners had no idea this was happening.

There is no bounce rate for a recommendation that never happened, no customer complaint, and no obvious metric only quiet, invisible lost demand.

That was the inspiration for Vidima. Buyers are beginning to move from typing keywords into search engines to asking AI for recommendations in their own language. Yet existing AI visibility tools are primarily designed for large English language marketing teams and provide limited coverage of smaller European markets.

I wanted to make this invisible loss measurable and more importantly, give business owners a practical way to respond.

What it does

Vidima runs the real, local-language questions a business’s buyers might ask across ChatGPT, Gemini, Perplexity, and Google’s AI answers.

Because AI answers vary, Vidima samples each question repeatedly. It measures how often the business is recommended, how prominently it appears, which competitors are winning, and which sources influence the answers.

It then turns that evidence into a prioritised action plan: pages to improve or create, FAQs to add, topics to cover, and relevant third-party websites to approach. After changes are implemented, Vidima re-scans the same questions to measure whether visibility, mentions, citations, or competitive position improved.

How I built it

The first technical challenge was measurement. Standard model APIs do not always reproduce what buyers see in consumer AI products, so I built browser-based agents that run repeatable, logged-out sessions across the four answer engines.

The next challenge was converting unstructured AI answers into reliable competitive data. An extraction model identifies which businesses were recommended, their order, and the evidence supporting each result. Deterministic guards reject unsupported candidates, while deterministic code calculates the final scores. The models interpret language, but they never invent the numbers.

The largest step was moving from measurement to action. Once a scan finishes, an event-driven recommendation workflow analyses where the customer is losing, inspects the customer’s website, and produces a prioritised implementation brief. It can recommend a new page, an on-page refresh, a comparison page, an article, or an FAQ.

Vidima also discovers influential third-party sources and classifies each one before drafting outreach. Directories, publishers, associations, news sites, and competitors are treated differently, and deterministic eligibility rules prevent the system from pitching unsuitable targets. Sending remains under human control.

I also built the leaderboard pipeline on Google Cloud. Cloud Scheduler starts a weekly Cloud Run job, which submits containerised engine scans to Cloud Batch. Results are stored in Cloud SQL and exposed through an API running on Cloud Run, with operational evidence captured in Cloud Logging.

Together, these systems create an AI-operated workflow: scheduled agents collect evidence, models make semantic decisions, deterministic code verifies and scores the results, and event-driven workers produce the next actions.

Challenges I ran into

The first challenge was small language blindness. Many keyword and AI-visibility datasets provide little or no useful Bulgarian coverage. The same market gap that makes Vidima valuable also made research and validation harder.

Bulgarian AI answers also transliterate brand names, for example, “Импрес (Impress)” which initially fragmented competitor matching. I had to normalise multilingual brand variants before scoring them.

The second challenge was noise. Extraction models could overcount businesses, misread refusals, or interpret an incidental mention as a recommendation. This pushed me to place deterministic evidence requirements around every model decision.

The hardest challenge was commercial, selling an invisible problem. A visibility score alone does not create urgency. The product needed to show concrete proof of loss, the questions where a competitor is recommended and the customer is absent, followed by a practical next step.

Accomplishments I’m proud of

2 business are in a trial period and will convert to paid users shortly after the competition.

During the hackathon, Vidima completed more than 500 scans and generated over 100 recommendations. The workflows run through production schedules and event-driven triggers rather than requiring me to start every stage manually.

Vidima is also its own customer. I use it to measure Vidima’s AI visibility, identify content gaps, and produce recommendations and drafts for improving the product’s own SEO and GEO presence.

What I learned

AI works best as the workforce, while the human remains responsible for judgment. As a solo founder, I can operate workflows that would traditionally require engineers, analysts, and agency staff. My role is to decide what should be built, what customers can trust, and which consequential actions should remain human-approved.

I also learned to use models for language judgment and deterministic code for measurement. A visibility number is only credible if it cannot be hallucinated.

Finally, measurement creates awareness, prescription creates action, and competitor monitoring creates recurring value. Customers need to understand why they are losing, what to do next, and when the market changes again.

Who it lifts

Vidima is built for small, local, non English language businesses: manufacturers, specialist retailers, printers, wholesalers, food producers, e-commerce companies, and other firms that win through relationships but risk being overlooked in AI-mediated discovery.

When one of these businesses becomes recommended where it was previously invisible, potential demand flows back to a real company and the people it employs.

What’s next

My next step is to prove retention and measurable customer outcomes in Bulgaria before expanding the same system into additional European language markets.

Over time, I want more of the research, monitoring, and growth loop to run autonomously while keeping evidence, customer trust, and consequential external actions under deliberate human control.

Share this project:

Updates