AI Search Visibility Tracking: Set Your Starting Point

Write down the decision the baseline should support. “Understand AI” is too broad to guide a review. “Decide whether our commercial maintenance page needs clearer scope information” is specific enough to investigate. It identifies a page, a customer task, and an action that the team can take.

Separate a content decision from a budget decision. A confusing answer can justify an editorial improvement even when referral volume is too small to evaluate a channel. Moving a substantial advertising budget requires stronger evidence about qualified demand and business outcomes. A useful baseline states what level of decision the available observations can reasonably support.

We'd name one person to own the definitions. Without that, one person counts brand mentions, another counts links and a third reports visits as recommendations. The resulting numbers look comparable while describing different events.

Each layer can be useful without proving the next. Fixing crawler access doesn't earn a citation, and a citation doesn't promise a visit, let alone a good-fit customer. Keep those differences in the report instead of asking one metric to stand for the whole path to a sale.

Start with a manageable collection of important pages. Include the main service explanation, a useful comparison or preparation resource, and the contact destination. Record the intended customer, the question each page answers, and the next action it supports. This inventory helps distinguish a visibility problem from a missing or unsuitable destination.

For a hypothetical commercial kitchen service company, tasks might include identifying supported equipment, understanding scheduled maintenance, and asking for an assessment. A broad recommendation question can be included, but it should not replace the questions buyers use to determine whether the company fits their situation.

Review the actual page before measuring it. Outdated availability or unclear service boundaries can make an appearance less valuable. Google’s AI optimization guidance connects its generative search features with established SEO fundamentals. A useful measurement plan should therefore remain connected to the quality of the underlying website.

Create a short source register. For each report, note the property or site, the dates, the provider, the selected filters, the available metrics, and the person who exported it. For a manual observation, record the question, context, date, and saved response. State whether an unavailable value is missing, preliminary, or outside the tool’s scope.

Use the provider’s terminology when describing a report. Google’s Generative AI performance report reports impressions for supported Google features, including AI Overviews and AI Mode. It does not serve as a universal report for every assistant. Our guide to reviewing generative AI impressions covers the practical review process.

Do not infer a complete history from a short export. A file may contain only the selected window or a limited set of rows. Preserve the original export and the settings used to produce it so another reviewer can understand exactly what the baseline contains.

As reviewed in September 2026, Google's dedicated Generative AI performance report documents impressions with page, country, date, and device dimensions. Dates use Pacific Time, page reporting generally follows canonical URLs, and property totals can differ from page totals. Save these definitions with the export rather than assuming that every row uses the same unit.

The documentation also notes preliminary data and that unavailable values can export as zeros. Save the report settings before you treat a downloaded zero as a confirmed absence. Use completed periods for a baseline when practical. The page's rollout and availability notes should be checked against the actual property instead of guessed from a screenshot of another account.

The broader Search results performance report includes clicks, impressions, click-through rate, and average position. Its Web search view is not a separate historical AI-only series. Keep it useful as a broader search baseline, but label it according to the report you actually exported.

In analytics, choose the traffic source scope that matches the question. Google's traffic-source scope documentation distinguishes first-user, session, and event dimensions. A question about how a particular visit began calls for a different view from a question about where a person was originally acquired or how credit was assigned to a key event.

Also preserve the channel definition. Google's current default channel group documentation includes an AI Assistant channel based on recognized assistant referrals and an ai-assistant medium. Inspect the actual channel and source values in your property. Do not assume that every AI influenced visit is detectable or that Google AI search clicks all belong to that assistant channel.

Give each observation an identifier. Record the assistant or search feature, the exact question, supplied location or business context, date, and source links that appeared. Add a short factual note about the response. For example: “The answer cited the maintenance page and described the supported equipment correctly.”

Use separate fields for the different observations:

• Linked citation: Preserve the destination and the claim it appears to support.
• Unlinked mention: Record the business reference without counting it as a linked source.
• Inaccurate description: Identify the disputed fact and the evidence needed to review it.
• No observed appearance: Retain the run so unfavorable observations stay in the denominator.

If a response contains several links to the same page, define whether the log counts one cited page or several link occurrences. Either can be recorded, but the method must be explicit.

Keep unfavorable observations as carefully as favorable ones. Rephrasing a question repeatedly until a preferred answer appears can be useful for exploration, but those exploratory runs should not quietly become the baseline. Follow the fuller citation logging workflow when multiple team members contribute.

Define a qualified inquiry before labeling visits successful. A business might require an offered service, a genuinely served location, and enough information to begin a useful conversation. The definition should reflect the sales process and should not change whenever a channel’s numbers look disappointing.

Keep the website’s measurement and the inquiry record connected through an appropriate internal process. That might include a form record identifier, an available source field, and a later qualification status. Collect only information needed for the task, and preserve unknown source values rather than assigning them to the newest marketing initiative.

Google Analytics defines a key event as an important measured action. A configured event still needs to correspond to the real action you intend to evaluate. Our AI referral and inquiry guide explains the handoff between website reporting and business records.

A baseline becomes more useful when the next review uses the same definitions. Compare equivalent completed periods, keep the page set stable where possible, and note changes in the observation procedure. If the business introduces a new service or replaces a contact form, record that change alongside the data.

Small counts need plain presentation. A hypothetical change from one qualified inquiry to two is one additional inquiry, even though the percentage increase is large. Report the counts and the comparison window together. The percentage alone can make a fragile observation look more substantial than it is.

For manual answer reviews, preserve the size of the tested question set. Three appearances out of twelve recorded questions are different from three appearances out of three. Neither result establishes the share of all customer answers. Read our small dataset review guide before describing early movement as a trend.

Consider an illustrative maintenance provider reviewing a service page, a preparation guide, and a contact page over one completed four week period. The business wants to decide whether unclear eligibility information deserves attention. It is not yet deciding whether to move a large marketing budget. All figures in this example are hypothetical and demonstrate the reporting method.

The team creates a fixed set of twenty-four buyer questions and records two planned runs for each, producing forty-eight observations on one named search or assistant surface. Nine observations contain a linked citation to the provider's site. Three additional observations mention the business without a link. The nine citation observations include two whose descriptions require a factual review.

Separately, the analytics export records eighty sessions under the team's documented assistant source definition. The inquiry system contains seven submission records associated with that defined source group through the agreed record process. One record is a test and one repeats the same request. After those exclusions, five distinct inquiries remain; two match the current qualification definition.

The correct summary does not say that nine citations generated eighty visits or that the two qualified inquiries were caused by those citations. The observation sample and the analytics population are different. Instead, the report identifies two specific descriptions for review and shows how the website activity and qualification records are currently being measured.

Suppose one inaccurate description concerns an old service area, and the service page still contains that obsolete statement. Correcting the page is a supported action. Suppose the other description is inaccurate even though the owned page is current. Record that difference and investigate the sources visible in the answer. Do not assume both descriptions came from the same place.

Define the unit of each important number. For citations, is the unit an answer observation, a unique URL, or a link occurrence? For inquiries, is it a form record, a distinct request, or a person? For outcomes, is it a meeting booked, an opportunity accepted, or revenue recognized under the business's process? Write the definitions before calculating rates.

Add exclusion rules for staff tests, obvious spam, duplicates, and records outside the reporting window. Preserve the excluded counts so a reviewer can follow the arithmetic. If qualification is still pending, use a pending status rather than forcing the inquiry into qualified or unqualified simply to finish the report.

Give changes to the dictionary a date and a reason. If the business expands its service area, a request previously classified as out of scope may now fit. That operational change can improve the apparent qualification rate without any change in search traffic. Keep the old definition available and explain the break in comparability.

A factual correction can proceed from one confirmed contradiction. A usability repair can proceed from a reproducible failure. A content experiment may be justified by several recurring questions and a sensible proposed answer. A major budget change needs evidence about qualified demand and commercial outcomes, not merely a favorable movement in an early citation sample.

Set the review date around the business's buying cycle and the amount of usable data expected. A monthly operational review can still be useful when sales take longer, provided it does not pretend that every open opportunity has already reached a final outcome. Report the stage and age of pending opportunities instead of silently treating them as failures.

If counts remain small, continue the observation period or narrow the decision. The team may be able to conclude that a service explanation is accurate and inquiries are being handled correctly while remaining unable to estimate a channel's return. Stating that limit preserves the usefulness of the work without turning incomplete evidence into a forecast.

Record the observation, your interpretation, the proposed action, and what would confirm completion. If a page misstates availability, the action is to correct the information and verify the published version. If several realistic questions reveal missing preparation guidance, the action may be a new section on the existing page. A lack of sufficient outcome data may justify continuing the observation period.

Keep the interpretation separate from the observation. “The business did not appear in this saved answer” is directly supported. “The business did not appear because its schema is incomplete” needs additional evidence. This distinction helps avoid spending money on a remedy that has not been connected to the problem.

Assign a review date based on the decision. Technical completion can often be checked soon after a change. Business outcomes may take longer, particularly when a proposal or purchasing process has several stages. A review date organizes responsibility; it does not promise that an AI system will respond within that period.

1. Decision: State the business question and the action it might inform.
2. Scope: Identify the pages, services, platforms, and dates included.
3. Definitions: Explain citations, mentions, referrals, and qualified inquiries.
4. Evidence: Link the source exports and saved observations.
5. Findings: Describe what was observed without assigning unsupported causes.
6. Actions: Give each task an owner and a completion check.
7. Next review: State what will be compared and what remains unavailable.

A short document with stable definitions is easier to maintain than a complex dashboard whose numbers nobody can explain. Add visual reporting after the underlying records are trustworthy. The report should allow another person to trace an important conclusion back to its evidence.

Ask a second person to reproduce one finding from the saved material. Give them the source register, the selected page, and the definition used in the summary. They should be able to locate the underlying observation and understand how it became the reported statement without relying on the original analyst’s memory.

For a hypothetical citation finding, that person should identify the exact question, the cited destination, and the accuracy note. For an inquiry finding, they should understand the qualification rule and the reporting period. If either route breaks, improve the record before expanding the dashboard.

Keep the review proportionate. The purpose is to catch ambiguous definitions, missing context, and accidental double counting. A small team can perform this check on a few important conclusions instead of building a complicated approval process around every observation. Record the corrections so the next reporting cycle starts with a clearer method.

Do we need a score for every platform?

Use a score only when its method, inputs, and limitations are clear enough to support a decision. An unexplained score may be convenient for presentation while concealing missing data. It is acceptable to report a verified observation without converting it into a universal rating.

Can one cited article prove that our GEO strategy works?

It proves that the article appeared in the recorded context. To evaluate a broader strategy, review repeated observations, relevant website activity, and business outcomes using a consistent method. A single appearance can be encouraging without carrying the entire conclusion.

What if we have no AI referral data yet?

Begin with the evidence available: page quality, relevant access checks, platform reports you can actually inspect, and a carefully maintained observation log. Mark referral measurement as a gap. Avoid estimating visits from citations or treating unknown traffic as AI traffic.

How should a team handle a change in the measurement tool?

Keep the earlier baseline, record the change date, and document which definitions or coverage changed. Run a short comparison where both methods are available if that is practical. Otherwise start a clearly labeled series instead of joining unlike numbers into an apparently continuous trend.

Use the AI search reporting playbook to organize recurring reviews and the search to sales attribution playbook to connect visibility work with the commercial process. For implementation help, request a measurement review with your important pages, current tools, and the decision you need to make.