Measure discovery all the way to the enquiry

Define the outcome first

Start from the result the business wants, usually qualified enquiries, and work backwards. Each layer upstream is a step toward it, and each layer has its own measuring instrument. The mistake is to treat a number from one layer as if it came from another: reading impressions as visits, or citations as customers.

Use each report for what it shows

Google’s own guide warns that Search Console clicks and Analytics sessions are calculated differently, so they will not match.^[2] It recommends using them together for a general pattern and not for reconciling exact figures. The same goes for the AI report: its impressions already sit inside the Web search type of the regular Performance report, so do not add the two.^[3]

Make a repeatable review

Choose a cadence you can keep, monthly for most small teams, and review the same things each time. Compare equal periods, such as the same four weeks across months or the same period in the previous year, because search demand has seasons. Write down what you expected before you look at the results.

• Visibility: impressions and clicks for your priority pages, and the Generative AI impressions for the same pages.
• Citations: your audit results and Bing’s citation counts, with their run counts.
• Visits: organic search sessions and AI referral sessions to those pages.
• Engagement: the click-through to the next step, per page.
• Outcome: enquiries, and how each says it found you.

Search Console lets you add custom annotations to charts, to mark events such as launching a page or fixing a bug.^[6] Use them. A chart that shows a change with no note of what you did is nearly impossible to interpret six months later.

Track AI referrals, and know their limits

Some AI visits can be identified. OpenAI says that ChatGPT adds the utm_source=chatgpt.com parameter to referral URLs, so visits from its search results can be tracked in analytics.^[7] For other services, look at the referrer hostnames that actually appear in your acquisition report. Build a segment or custom channel from the hostnames you see, not from a list copied from elsewhere. An example pattern, to adapt:

chatgpt\.com|perplexity\.ai|copilot\.microsoft\.com|gemini\.google\.com|claude\.ai

Treat the result as a floor. Some apps open links without passing a referrer, and those visits are counted as direct traffic. This is why the single most useful addition is a question on your enquiry form: How did you find us? Offer the options that reflect your market, including “an AI assistant (such as ChatGPT or Gemini),” “a search engine,” “a recommendation” and “other.” It costs nothing and catches the visits that analytics cannot attribute.

Do not expect a one-to-one relationship between impressions and clicks. Pew Research found that visits to results pages with an AI summary produced fewer clicks on traditional results than visits without one, 8% against 15%,^[8] and SparkToro’s clickstream analysis found just under 60% of US Google searches ended without a click on any result.^[9] More visibility with flat or lower clicks is a pattern you may well see. It is a reason to measure outcomes directly.

Review a hypothetical change

Suppose a clinic rewrites its sports massage page: a direct answer at the top, a comparison table, a dated price section and an author byline. Before publishing it records the baseline and writes its prediction. This example is invented to show the structure of a review.

What can be concluded? The page improved by more than the control pages did, which is encouraging. The counts of enquiries are small, so the difference between 5 and 8 is weak evidence. The results do not prove the rewrite caused the change: a seasonal peak, a ranking shift or a one-off mention could also explain it. A reasonable decision is to keep the change, keep watching and apply the same pattern to the next page, not to announce a result. The control pages are what turn a before-and-after into something more than a story.

Make the review operational

A review that nobody owns does not happen. Name the owner, put the date in the calendar, keep the template in one place and decide in advance what would change your course: for example, enquiries from organic search falling for two consecutive months, or a priority page losing its Generative AI impressions. Write each review as a short note: what changed, what the data suggests, what we will do and what we will check next time.

Apply it to your website

• You have named the business outcome and the layers that lead to it, with an instrument for each.
• No report is used for something it does not show, and impressions, citations and visits are never added together.
• Search Console and Analytics are used for patterns, not reconciled to the last digit.
• AI referrals are segmented from the hostnames you actually see, and the counts are treated as a floor.
• Your enquiry form asks how people found you, with an option for AI assistants.
• Changes are annotated, compared against equal periods and, where possible, against control pages.
• A named person owns a monthly review, with a note template and decision rules.