Test Name: Training Data Check
Assigned to: Arda
Notes: This assumption is critical because it is a crucial component of our pivot into the GenAI space. If we are wrong, we will not be able to train our own model and therefore stay limited in our capabilities since it will just be advanced search options.
Step 1: Hypothesis
We believe that: we have sufficient training data to be able to train our own image generation models, which will generate images of quality on par with current GenAI image models.
Step 2: Test
To verify that we will: run a small-scale trial and see how results improve as we add more data. We will try to train or fine-tune a lightweight, existing model (like a smaller open-source image generator) using part of our dataset, then repeat with larger portions (25%, 50%, 75%, and 100%). If the image quality quickly levels off and stops improving, our dataset is likely large and diverse enough. But if each increase clearly improves results, you probably need more or better-balanced data. This experiment is simple, affordable, and gives a realistic sense of whether your data can support a full-sized AI model.
Step 3: Metric
And measure: whether the model trained on our dataset can consistently produce images that are realistic and relevant to the prompts that are comparable in quality to the existing models trained on larger datasets.
Some metrics could include:
- Humans rating images on a scale from 1-5 for visual realism
- Humans rating images on a scale from 1-5 for how well it matches their textual prompt
We can also use indicators like unique output rate or loss to understand how the model is doing in terms of generalization or overtraining.
Step 4: Criteria
We are right if: The model trained on our dataset can consistently generate images that are realistic, diverse, and relevant to their prompts, performing comparably to open-source baselines.
Also:
- Human evaluators rate image realism and prompt accuracy at 4/5 or higher on average.
- Model performance improves minimally with additional data, suggesting the dataset is already sufficient in scale and diversity.
Test Name: Willingness to Pay for Ethically Verified Images
Assigned to: Nayan, Reyna
Notes: We picked this assumption because if buyers won’t pay because the images are ethically verified, our core value proposition fails, so we will gain no revenue. This is a top risk, as it is high impact, but we currently have low evidence only based on anecdotes.
Step 1: Hypothesis
We believe that marketing/brand teams at enterprise companies will pay standard stock photo pricing (eg. $29/image or $99/month) if images include clear, verifiable ethical proof. Specifically, at least 5% of qualified visitors will submit payment info for a $99/month plan when the ethical verification is clearly presented at purchase.
Step 2: Test
To verify that we will make a pricing page with checkout that shows a badge on the image denoting its “Ethical” status and a link to a clickable prototype of the image detail page. We will also make a control page that is the same but without the ethical badge/proof. We will send 200-300 target buyers in brand/marketing roles to these pages and split traffic evenly. At checkout, we will collect card details. After checkout, we will ask the target buyers how big the problem is for them on a scale from 1-10.
Step 3: Metric
And measure the percentage of visitors who enter card details per page (payment intent); whether the page with the ethical badge gets a higher payment rate than the control; the number of people asking for invoice or vendor setup (buying signals); and the percentage of people rating the problem 8-10/10 (problem score).
Step 4: Criteria
We are right if at least 5% of visitors enter card details on the ethical page; the ethical page shows at least a 30% higher payment rate than the control; we get at least 2 real buying steps (invoice/vendor requests); and at least 60% rate the problem 8-10. We are wrong if fewer than 2% of visitors enter card details overall or the ethical page has little/no increase in payment rate vs. the control; we get 0 real buying steps; and most rate the problem below 8.
Test Name: Enterprises value copyright safety
Assigned to: Tuvana
Notes: I thought that enterprise buyers value copyright safety, but we have very little evidence to support it. Because of recent regulations like the EU AI Act, many top companies are being held accountable for how they use AI-generated content. This has increased pressure on marketing and legal teams to ensure every image they use is compliant and protected.
Step 1: Hypothesis
We believe that enterprise buyers will prefer tools that make copyright approval easier and faster. Platforms that clearly show usage rights will be seen as safer and less risky for legal and procurement teams.
Step 2: Test
To verify this, we will interview ten enterprise professionals, including marketing directors and legal counsel, from media, retail, and design companies. During each session, we will show participants three platform options that offer different levels of copyright protection:
- A creative tool with unclear licensing, where ownership and reuse rights are uncertain.
- A standard image generation tool that has basic content rights but no copyright guarantees.
- GSPhotoCo, which provides copyright free content
We will also include a simple pricing chart showing realistic subscription levels ($20, $35, and $50 per user per month). Participants will choose which platform they prefer, explain why, and share the price point where they would be willing to pay more or start to reconsider. This will help us understand not just interest, but real willingness to pay for copyright safety. This is something we deeply care about.
Step 3: Metric
And measure: we will measure
- The percentage of participants who choose the copyright platform.
- The average amount they are willing to pay compared to standard tools.
- The number of participants who mention brand protection or legal safety as their main reason for choosing a platform.
Together, these measures capture both brand safety reasoning and price sensitivity, providing a better view of enterprise motivation.
Step 4: Criteria
We are right if:
- At least 80% of participants choose the copyright free platform.
- At least 50% say they would pay 10% or more above standard pricing for stronger copyright protection.
If these goals are not met, we’ll know that copyright safety alone may not be a strong selling point and that we should test other motivations within copyright safety. Achieving these thresholds would confirm that copyright safety can function as a key differentiator in enterprise purchasing behavior. This will benefit by helping us focus on the most convincing factors when pitching.
