Known product
Product named“Is this product worth using?”
Whether AI recommends a product it has been asked to evaluate.
How it works
One test.
Every step explained.
Agentzia simulates how buyers use AI agents to discover products, compare alternatives and decide what to buy. See whether they choose your product, why they pass it over and what to change on your website.
- Input
- A product URL + selected AI profiles
- Process
- Separate buying research conversations
- Output
- Decisions, evidence and prioritized changes
Start with the page
buyers can access.
Submit your public product URL and choose the AI profiles to test. Agentzia currently supports representative ChatGPT, Claude, Gemini and Grok profiles. You can also supply competitors for the comparison question.
Agentzia checks the public page and search surface to derive a product description. That description enables discovery questions that do not name your brand.
Separate discovery
from preference.
Each selected profile receives the same available situation set in separate conversations. Naming your product tests evaluation; leaving the brand out tests discovery.
Category discovery
Product not named“What should I use for this need?”
Whether the product enters the shortlist without a brand prompt.
Problem-led research
Product not named“How should I solve this problem?”
Whether AI connects the product to the buyer’s problem.
Competitive decision
Product named“Which would you choose over the alternatives?”
Which option AI selects and what drives the comparison.
These are shortened examples. The report retains the exact requests. A full panel has four situations per profile. If a usable brand-free description cannot be derived, the two discovery situations are omitted and the report records the gap.
Let each profile
reach its own answer.
- Ask the buying question.
The participant is instructed to help with an ordinary purchase decision. Website content is treated as evidence, not as instructions to follow.
- Allow web research.
The profile decides whether to search, what sources to use and which alternatives to consider. Agentzia does not supply a preferred verdict.
- Preserve the response.
The answer and available research records are saved before a separate analysis extracts decisions and findings.
- Model profiles and execution routes
- A profile identifies the representative model being tested. The provider route supplies its web-enabled execution. These are distinct from the consumer app, which may add its own tools, memory and personalization.
- Retries and provider fallback
- Transient failures can be retried, and configured fallback routes can handle eligible provider failures. A recovered answer is still one decision, not an extra vote. Check recorded execution details when comparing runs: a route change can affect available research evidence.
- Independent conversations
- Each buying situation starts separately. Participants do not see one another’s answers. A separate analysis organizes their completed responses into the report.
Inspect the answer.
Check its sources.
Open an individual decision to read the request, full response, cited sources and available research records. Evidence coverage depends on what the provider exposes.
- Provider-observed records
- Tool records can show exact searches and page visits when the execution route exposes them.
- Citations
- Sources attached to an answer support inspection, but do not reconstruct every search or visit.
- Participant-reported activity
- A research log supplied by the model is identified as reported activity. It is not equivalent to provider-observed tool records.
An unavailable search or visit record does not mean no research happened. Check coverage before drawing conclusions about how an answer was produced.
Inspect a decision and its evidenceMove from a finding
to a specific edit.
The analysis organizes observed decisions, recommendation barriers, alternatives and prioritized actions. Original responses remain available so you can check the interpretation.
Example of turning a finding into an edit
“The claim needs proof.”
- Where
- Beside the outcome claim on your product page.
- Change
- Add a result with the measured outcome, who it applies to, the measurement period and a link to its source.
- Check
- Retest and inspect whether the same evidence concern remains in the new answers.
Illustrative edit structure. Only publish results you can substantiate.
Recommendations are proposed improvements, not measured lifts. A decision to recommend a trial is also distinct from choosing the product beyond trial. Keep those outcomes separate when reading the summary.
Compare like
with like.
Publish the change and run the test again. Compare the actual requests, profile configuration and panel coverage before attributing a different decision to the edit.
- Matched questions: inspect decision changes alongside the new evidence.
- Changed discovery prompts: treat the run as a new baseline when the generated prompt roster changes.
- Missing paths: inspect completed answers, but do not treat an incomplete panel as an equivalent recommendation rate.
- New models or live sources: account for these changes when interpreting movement.
A repeatable comparison
Save the baseline before you edit.
- Keep
- The original report, exact requests, URL, model configuration, completed paths and date. Record the page text or release commit you are testing.
- Change
- Make one coherent change tied to a recorded concern. Note the affected page and deployment time, then verify that it is publicly accessible.
- Compare
- Check the same concern in the new response, the cited evidence and the final choice. Separate named-product evaluation from unprompted discovery, and trial recommendations from adoption beyond trial.
- Repeat
- Look for the same pattern in additional matched observations. Live sources and model outputs vary, so a single improved answer does not isolate the effect of your edit.
If the model, route, generated questions or coverage changes, describe those differences and establish a new baseline where necessary. Agentzia’s current test is an on-demand diagnostic, not a controlled causal experiment or continuous visibility monitor.
Follow the practical improvement guideScores describe the observed panel. They are not market share, sales forecasts or a guarantee of what every customer will see. Repeated tests help distinguish recurring issues from a single response.