What we tested
Each product received the same three prompt types on the available consumer web versions of Gemini and Perplexity. We preserved the answer and checked whether the target and direct competitors appeared.
| Prompt | Buyer question | Signal |
|---|---|---|
| P1 | Open category discovery for a static-site contact form | Can the product enter an unprompted shortlist? |
| P2 | Feature-led selection: spam protection, setup, free tier | Can requirements retrieve the product? |
| P3 | Named comparison against four direct alternatives | Can the model understand and position it? |
Open discovery results
The chart shows appearances across the four open P1/P2 answers. A zero does not mean the product is unknown. It means the product did not enter these small buyer shortlists before being named.
The gap appeared before comparison.
Form.taxi, Basin, and Formspark were absent from all four open answers even as Formspree, Web3Forms, FormBold, FormSubmit, Netlify Forms, and other alternatives recurred.
Named understanding was different
When the comparison prompt explicitly named the products, Gemini and Perplexity both recognized Form.taxi, Basin, Formspark, and FormBackend. FormBold was recognized by Gemini but omitted by Perplexity even in the named comparison.
| Product | Open answers | Named comparison | Observed pattern |
|---|---|---|---|
| Form.taxi | 0 / 4 | 2 / 2 | Understood, but not retrieved by default |
| Basin | 0 / 4 | 2 / 2 | Understood, but not retrieved by default |
| Formspark | 0 / 4 | 2 / 2 | Understood, but not retrieved by default |
| FormBackend | 3 / 4 | 2 / 2 | Strongest open discovery in this sample |
| FormBold | 1 / 4 | 1 / 2 | Uneven discovery and named recall |
What this does and does not show
- It shows a repeated omission pattern in one dated, controlled sample.
- It does not prove a stable AI ranking or lost revenue.
- It does not measure product quality.
- It does show why named-product tests alone can create false confidence.
- It provides a concrete starting point for reviewing category pages, use-case facts, comparisons, FAQs, and public evidence.
Study limits
AI answers vary by date, location, model version, interface state, and prompt wording. This study used two accessible consumer web surfaces and three prompts per product. It should be treated as a field observation, not a population estimate.
No company named here commissioned the study. Product names are used only to report observed public AI outputs. AgentRank has not claimed that any omission was caused by a specific website defect.
Run the same check for your product.
The USD 99 Audit Lite covers five buyer prompts, up to four available AI surfaces, competitor evidence, and a prioritized repair roadmap.