Product strategy
Any advertiser can put an air-ready TV spot on screen and know what it caused.
Spotcheck is a prototype for an ad platform's creative and measurement layer. It tests two bets: that AI plus a rules gate can produce spots that pass publisher specs and brand review with little rework, and that experiments, not attribution, should be the default answer to "what did my TV spend do?"
Evidence status. Personas, problems and metrics on this page are working assumptions from public job descriptions and industry knowledge. They have not yet been tested with users. The validation plan is at the end.
Vision
Any advertiser, not just the ones with agencies and measurement vendors, can run TV-quality campaigns and trust the answer to what they caused.
Mission
Make creative production and measurement trustworthy by default: AI for judgment, code for anything that must be exact, and experiments for anything that claims cause.
Goal of this prototype
Find out, with numbers, how often an AI-drafted spot is ready to air on the first try and after fixes, and show how far attribution drifts from what ads actually cause.
Problem statements
Mid-market advertisers who already run display and social often can't run CTV, because producing a spec-compliant video spot takes money, time and skills they don't have in-house. Budget that could go to TV stays in channels they already have creative for.
Designers spend hours resizing one idea for TV, phone, billboard and audio, checking safe areas, legibility and legal-line timing by eye. Mistakes are caught by publisher QA after submission, which means rejections and missed flight dates.
Generative tools write fluent copy that can state things the brand never said, like rankings, ingredients or prices. Brand and legal reviewers have no fast way to see which lines are backed by approved source material.
Last-touch and data-driven reports reward channels that sit close to purchase and punish ones that can't be clicked or matched, like CTV. Advertisers cut what works, keep what doesn't, and stop trusting the platform's numbers.
Proto-personas
Composites to be replaced by interview findings. Names are placeholders.
Maya, marketing manager
Mid-market consumer brand, team of three, no video team
- Wants
- To test CTV without an agency retainer, and to show her boss it paid off.
- Blocked by
- No TV-ready spot; unsure what specs publishers require; attribution that shows TV as weak.
- Success looks like
- A spot live this month and a lift number she can defend.
Dev, creative studio designer
Platform's in-house studio, serves many advertisers at once
- Wants
- To spend time on ideas, not on resizing and spec-checking the same spot four ways.
- Blocked by
- Manual versioning, spec sheets that differ by publisher, late rejections.
- Success looks like
- One approved concept becomes every format, and nothing bounces at QA.
Sam, brand and legal approver
On the advertiser side; signs off anything that airs
- Wants
- Certainty that every claim on screen is one the brand can substantiate.
- Blocked by
- AI copy that sounds right but can't be traced to approved source material.
- Success looks like
- Each claim shown with its source, and nothing approved while a claim is open.
Priya, measurement lead
Agency or large advertiser; owns the numbers the client sees
- Wants
- Numbers that survive scrutiny from the client's own analysts and MMM vendor.
- Blocked by
- Platform reports that grade their own homework; no easy way to run and size holdouts.
- Success looks like
- Incrementality by channel, with honest uncertainty, and a budget move she can justify.
Solution
One loop: build a spot, gate it on rules and grounded review, measure what it caused with experiments, and evaluate the AI itself so quality is a number, not a demo.
| Problem | What Spotcheck does | For |
|---|---|---|
| Creative is the gate to CTV | Brief in, spot out in four formats: TV, phone, billboard and audio. Uses the brand's own images or clips. | Marketing manager |
| Manual versioning, late errors | Code retimes to the frame, reframes around the subject, and checks safe areas, legibility, contrast, broadcast colours and legal-line timing before anything is submitted. | Studio designer |
| Claim risk | A separate AI reviewer must cite the brand source for every claim; code confirms each citation exists; approval is locked while any claim is open or the review is out of date. | Brand and legal approver |
| Attribution credits the wrong channels | Experiments first: holdouts, geo tests and a sized test plan, with attribution shown alongside so the gap is visible. | Measurement lead, marketing manager |
Principles
- AI decides, code verifies. Nothing the model says is trusted until a deterministic check passes.
- Cause needs an experiment. Attribution is a view of who converted, never presented as what the ads did.
- Show the uncertainty. Every measured number carries its range, and unmeasurable is a valid answer.
- Humans approve. The system can block; only a person can approve.
Scope
Ship first
- TV 16:9 and phone 9:16 from brand assets
- Rules gate with editable publisher spec profiles
- Grounded claim review and approval lock
- CTV holdout report with a test-size planner
Next
- Billboard and audio variants
- Geo tests for channels that can't be held out per home
- Multi-user roles and comments
- Experiment-calibrated media-mix model
Not doing
- Training a video model; use partners
- Replacing agencies for top-tier brands
- Selling attribution as proof of cause
Success metrics
Targets are hypotheses to set after a baseline, not results.
| Metric | Type | Why it matters |
|---|---|---|
| Share of advertisers launching CTV with in-platform creative | North star | Direct measure of removing the creative gate. |
| Time from brief to first air-ready spot | Input | The speed promise of the product. |
| First-draft pass rate, and pass rate after AI fixes | Input | Measured in the Evaluate tab. |
| Human edits per approved spot | Input | Rework the AI didn't save. |
| Publisher rejection rate | Input | Whether the rules gate matches real specs. |
| Share of CTV spend with an incrementality readout | Input | How often advertisers get a causal answer. |
| Unsupported claims aired | Guardrail | Must be zero. Reviewer recall on planted claims is the leading indicator. |
| Creative quality rating (blind human panel) | Guardrail | Rules can pass a dull ad; people catch it. |
| AI cost and time per approved spot | Guardrail | Keeps unit economics viable for mid-market. |
Risks and mitigations
| Risk | Mitigation | What's still open |
|---|---|---|
| AI airs a claim the brand can't substantiate | Separate reviewer, source citations checked by code, human approval lock | Code checks a citation exists, not that it supports the claim |
| Creative passes every rule but looks generic | Brand kits, studio-built templates, blind quality rating in evaluation | No quality panel yet |
| Rules drift from real publisher specs | Editable spec profiles per publisher, owned by ad ops | Defaults are generic, not verified |
| Clients distrust results that contradict attribution | Show both side by side, with ranges and a plain-language summary checked by code | Needs enablement and a default-view decision |
| Holdouts withhold paid reach | Size tests with the planner; geo tests where household holdouts don't work | Small channels may never be measurable per campaign |
| Identity signal loss weakens CTV matching | Experiments don't depend on matching; false-match rate modelled | Clean-room workflows not covered |
| AI cost and latency per spot | Fast model tier for rewrites, limits on fix rounds, caching | No cost tracking in the prototype |
| The simulated world isn't the real one | Presets that include a world where attribution is right; editable true effects | Needs a pilot with real campaign data |
Assumptions to validate
- Mid-market advertisers skip CTV mainly because of creative, not budget or targeting. Test: 5 interviews, plus funnel data on CTV starts vs launches.
- Designers lose meaningful time to versioning and spec checks. Test: time a studio designer on one spot, manual vs Spotcheck.
- Approvers will trust cited sources over their own read. Test: 3 approvers review 10 spots with and without citations.
- Advertisers accept holdouts that withhold some reach. Test: offer a sized holdout to pilot accounts and track opt-in.
Roadmap
- Now
- Run the evaluation, including the planted-claim probe, and publish baseline numbers.
- Next
- User interviews; one studio designer trial; a real publisher spec profile.
- Later
- Pilot with live campaigns to replace simulated effects with measured ones.
Open questions
- Should the default report lead with experiments even when none has run?
- Who owns spec profiles: ad ops, publishers, or the platform?
- How much human editing is acceptable before "AI-first" stops being true?