How to tell whether AI search is bringing people to your website

Ask three separate questions.

After improving your website, it is reasonable to ask whether the work helped. Start by separating three things: did the website issue change, did a platform report more appearances, and did useful visits or enquiries follow?

Your ScanMySEO report helps with the first question. It is not a live monitor of answers from ChatGPT, Google or other services. A clearer service page can be a worthwhile improvement even before there is enough data to assess discovery.

Keep each measure attached to its source. An impression is not a visit, and a visit is not an enquiry. Combining them into one “AI score” can hide the question you actually need answered.

Record what changed on the website.

Save the affected URL, the original finding and the date the change was published. Describe the change specifically: an indexing instruction was removed from a public page, a broken internal link was corrected, or a service introduction now explains what is included.

Then check it. Re-crawl for the relevant technical result and read the live content against the confirmed business information. When the follow-up audit cannot reach a page or only reviews part of the site, retain that limitation beside the result.

A new report with fewer issues establishes a change in the checks performed. It does not, on its own, establish a change in a provider’s answers or prove that all intended pages were reviewed.

Use platform reports for the experiences they cover.

Google’s Generative AI performance report for Search reports impressions for supported features, including AI Overviews and AI Mode. Check the selected property, date range and grouping. The documentation also describes availability and data limitations, so an absent report is not proof of zero appearances.

Bing’s AI Performance documentation describes citation reporting across supported AI experiences. Record the exact metric and reporting coverage instead of treating it as a count of every mention across the internet.

Neither report should be silently substituted for the other. Keep separate rows for Google impressions and Bing’s reported citation measures, with their original names and units.

A manual prompt check can be a useful observation, but record the prompt, service and date. One answer is not a stable ranking or a complete measure of how often the business appears.

Look for visits and useful next steps.

In your existing analytics, review the relevant source and landing pages. OpenAI’s publisher FAQ documents ChatGPT referral attribution, including its utm_source=chatgpt.com tagging. That can help identify visits; it does not count all mentions or citations.

Then ask what those visitors did. Use the enquiry, booking, purchase or other meaningful event your analytics is already configured to measure. Keep its definition consistent between comparison periods. A new event setup halfway through the comparison changes what the numbers mean.

Do not assume that all traffic from an unfamiliar referrer is an AI recommendation. Investigate the source and preserve attribution limitations. Missing attribution and no activity are not necessarily the same thing.

Keep a small, readable record.

A spreadsheet or an existing reporting note is enough. Use one row per source and measure. The table below is a blank planning example, not collected performance data.

Take care with exported data. Google’s report documentation notes that some unavailable values can become zeros in a download. Preserve the meaning of the original display rather than calling every exported zero a confirmed absence.

Compare like with like.

Choose comparable date windows, keep the same property and filters, and note anything else that changed: campaigns, seasonality, availability, prices or a wider website release. The aim is a useful comparison, not the most flattering one.

When both website quality and platform appearances improve, describe that as an observation. Do not say the page edit caused the increase unless you have evidence that can support that stronger conclusion.

For a small website with little data, the next useful step may be to continue collecting observations while improving an important page. There is no need to invent a score to fill the gap.

Use the result to decide what to do next.

If the issue remains, investigate the implementation. If the page is fixed but there is little outcome data, keep the record and avoid drawing a sweeping conclusion. If visits are arriving but people cannot complete the next step, examine the landing page and the enquiry or purchase journey.

Keep the website work grounded in customer needs. Our practical readiness guide shows how to start with one page; the product overview explains the audit’s scope.

Provider documentation reviewed on 6 September 2026. Reporting availability, definitions and limits can change; use the linked primary documentation when interpreting current data.