The challenge.
Sponsors pay for specific placements: at the venue, on screen and on players. What a placement is worth depends on where the logo sits as much as on the brand.
YouGov Sport's existing system could spot a logo, but it could not reliably tell which sponsorship asset the logo was on, so analysts fixed results by hand. Three gaps sat behind that.
Same brand, different value.
The same logo on two different placements is worth different amounts to the sponsor, and close-up camera shots make the two hard to tell apart.
No memory of the scene.
The system judged each frame on its own and forgot which hole was being played as soon as the on-screen hole number disappeared. Sponsor names inside leaderboards, captions and broadcast overlays went uncounted, and commentary that named the hole, the player and the sponsor was never used.
Slow to change, and it had to stay consistent.
Every new sponsorship asset or event meant manual setup and retraining. And because YouGov Sport's clients compare numbers season to season, no improvement could shift its established scoring.
What we built.
We gave YouGov Sport's platform the ability to understand the whole broadcast, combining what is seen, what is shown on screen and what is said. The work drew on our computer vision services for detection and tracking, and on generative AI for the cases a detector alone cannot settle.
Sponsorship asset detection (97%+ accuracy).
Identifies the exact sponsorship asset a logo sits on, across venue signage, on-screen placements and player apparel.
- Reads the board or placement around the logo as well as the logo itself, so small logos on large boards are placed correctly.
- Recognises where on a player's clothing a logo appears.
Scene tracking.
Follows players, objects and logos from frame to frame (using ByteTrack), so context isn't lost when the camera zooms in.
- Knows which hole is being played and carries that forward across the broadcast.
- Uses camera angles and course features to place assets correctly.
Screen overlays.
Reads text in leaderboards, scorecards, captions and broadcast graphics, in English and Japanese.
- Tells broadcast graphics apart from real signage at the venue.
- Picks out sponsor names, including multi-word brands, and ignores ordinary words that happen to share a brand's name.
Audio commentary analysis.
Converts commentary to text with speech-to-text, then picks up mentions of holes, players and sponsors and uses them to confirm which hole or player is on screen.
- Captures verbal sponsor mentions and reports them to YouGov Sport's clients as an additional exposure signal.
Vision language models (VLMs).
AI that can look at a frame and reason about it in plain language, the way a human analyst would.
- Steps in on the hard cases, such as an ambiguous close-up, to confirm the asset and the hole.
- Speeds up onboarding of new events and graphic styles with little or no manual labelling.
Feedback loop for new assets.
YouGov's own team reviews results, corrects them, and adds new assets or sponsors.
- Corrections feed straight back into the machine learning models, so accuracy improves with every event.
Client viewing portal.
YouGov Sport's clients log in to see their footage with brand boxes highlighted and how long each brand was visible.
- Built into YouGov's existing customer platform, and each client sees only its own content.
How it works, in plain terms.
Each broadcast passes through five steps, and the platform gets smarter every time YouGov's team reviews the results.
Each release is tested against YouGov Sport's established scores before it reaches a report.
Watch, understand, place, check, learn: one pass per broadcast.
Results.
Asset-level measurement is now accurate enough to report without a correction pass on every event.
- Sponsors get exposure measured per asset, which is what they actually pay for.
- Analysts spend their time on insight rather than correcting data.
- YouGov Sport can take on more events and more sports without adding headcount.
How we worked with YouGov.
We built around YouGov Sport's existing system rather than replacing it, so its clients saw better data with no disruption.
Learned from YouGov's own analysts.
Past manual corrections became training data, so the AI learned YouGov's standards from day one.
Protected the numbers clients rely on.
Every model change is tested against YouGov Sport's established scores before it goes live.
YouGov owns everything.
All code, models and data stay in YouGov's own cloud.
Built for independence.
YouGov's team can add assets, retrain and release models without depending on Algoscale.
Ready to scale.
The same approach extends from golf to Formula 1 and football.
Where it applies.
The pattern here is not specific to golf. Any business that has to measure what appears in video, from retail shelf audits to media monitoring, can combine detection, scene context and a reasoning model the same way. The model that checks hard cases before they reach a report is the same building block we use in AI agents.
If you are working out where computer vision or a VLM fits in your own measurement, our artificial intelligence consulting services start with your footage, your existing scores and the decisions they feed.
