Illustrative photograph. Photo: Laptop on wood (Unsplash).jpg by Markus Spiske, via Wikimedia Commons — CC0. Downloaded Wikimedia version; no custom edits.

You improve three pages and ChatGPT referral traffic rises the following month. Did the edits help?

Possibly. But the same pattern could reflect greater platform usage, seasonality, another campaign, or ordinary variation. A useful experiment makes those competing explanations easier to assess.

This is a suggested evaluation method, not a claim that AI visibility can be measured perfectly.

State one change and one expected outcome

“Improve AI visibility” is too broad to test cleanly.

Try: “Clarifying our integration’s supported actions will reduce incorrect descriptions in a fixed sample of integration-related answers.”

The intervention is specific. The primary outcome is factual accuracy. Citations, referrals, and leads can be secondary observations, rather than being substituted after you see which number improved.

Write down the hypothesis before editing the page.

Establish a baseline under recorded conditions

Choose a fixed set of relevant questions. Record the exact wording, model or mode if visible, search setting, location context, and personalization conditions.

Run the set on multiple baseline dates where practical. Retain full answers and cited URLs. A screenshot of the best result is not a baseline.

For manual scoring, define the rubric in advance. For the integration example, mark each answer correct, partially correct, incorrect, or unassessable, with a reason. Keep “not mentioned” separate from an incorrect description.

Change the smallest useful unit

Update the integration page without simultaneously redesigning the site, changing the product, and launching a large campaign if you can avoid it.

Log the exact edits and publication time. Confirm that the live page serves the approved content.

You do not need to preserve a harmful error for the sake of experimental purity. Correct important mistakes promptly; simply acknowledge when the circumstances prevent a clean test.

Use a comparison where it makes sense

Choose similar pages or question groups that did not receive the intervention. They can help reveal a broad change affecting both groups.

The comparison is imperfect. Pages differ in demand, competition, history, and how they are retrieved. Document those differences rather than presenting the groups as equivalent by default.

Avoid comparing a popular product page with an obscure legal page merely because both are easy to track.

Set review dates and stopping rules

Allow time for the changed page to be discoverable, while acknowledging that retrieval timing is not under your control. Choose review dates in advance instead of repeatedly checking until you find a favorable answer.

If there are too few observations, report that limitation. No universal minimum number of days or prompt runs can guarantee a reliable conclusion for every business.

If the visible product, model, or search conditions change during the test, record the change and consider restarting the comparison.

Report what happened and what you can infer

A useful report includes the intervention, sample, baseline, later observations, comparison behavior, business outcomes, and limitations.

For a fictional result: “Incorrect integration descriptions fell from six of thirty assessed answers to two of thirty; the unchanged question group also improved slightly.” That supports further investigation. It does not establish a universal effect or a platform-wide exposure increase.

Choose the next decision from the evidence: keep the clarification, extend the approach to similar pages, collect more observations, or stop an ineffective tactic. Publishing a useful correction can be worthwhile even when its effect on AI discovery remains uncertain.

Next read: Put the work into a 30-day plan.