1. Capture the buying situation
Use requests that reflect how customers choose: whether your product is worth using, which product fits their problem and how yours compares with alternatives. Include brand-free discovery questions as well as questions that name your product.
Save the exact request, product URL, date, model configuration and original answer. A named-product recommendation does not establish that buyers will discover you unprompted.
2. Locate the barrier
Read the answer and citations before acting on a summary. Separate what the agent actually observed from what the analysis suggests.
Not found
Inspect the category and problem language in the answer and available research records. Build a useful page for that intent and link it from relevant product pages. A new page alone does not guarantee discovery.
Misunderstood
Compare the agent’s product description with your intended use case. Put a concrete description of the job, user and output near the top of the relevant page.
Not trusted
Identify the unsupported claim or missing policy. Add attributable evidence or the actual operating terms beside the claim, with a link to the source.
A rival fits better
Identify the deciding requirement. Explain your actual capability and limits in a comparison or use-case section. If you do not meet the requirement, changing copy will not fix the product gap.
3. Write an implementation-ready change
Give each task a page, a location, replacement content, supporting evidence and an acceptance check. Prefer one focused change to a general instruction to improve trust.
Instead of: make pricing clearer
On /pricing, directly below the first-test CTA, state: ‘Your first test is free. No credit card required.’ Then explain how a paid test is charged using the current billing rules.
Instead of: add more proof
Beside the outcome claim, add a result naming the measured outcome, population, time period and original source. If that evidence does not exist, narrow the claim rather than inventing support.
Acceptance check
Confirm the information is visible in the rendered page, accessible without login and reachable through a relevant internal link. Check that its wording matches the source.
4. Prioritize the decision, not the score
Start with a recurring concern that affects a buying decision, has clear evidence and can be addressed truthfully. Group repeated concerns so several answers do not become duplicate tasks.
Record whether the work needs copy, product changes, new research or third-party proof. Your website cannot supply independent validation simply by describing itself more confidently.
5. Retest and inspect the reason
After publishing, check the same requests, model configuration and completed coverage. Read whether the original concern disappeared, whether the new evidence was cited and whether the final decision changed.
Repeat observations before concluding that an edit caused an improvement. Live sources and model behavior can change. If discovery prompts or routes differ, label the result as a new baseline rather than presenting a clean before-and-after lift.
Bring the work into your coding agent
Connect through Agentzia MCP, ask the agent to inspect the report’s actions and decision records, then propose the exact page edits. Review factual claims and supporting sources before publishing.
After deployment, run a new test and ask the agent to compare the original concern with the new answer. Keep the edit and the evidence together so the next iteration starts from what you learned.