KNOWLEDGE LIBRARY

AI Search: Mentions, Citations and Referrals

A screenshot of an AI answer can be useful evidence. It becomes much less useful when its caption claims more than the screenshot shows. “Our page was linked in this answer” and “AI recommends our company to everyone” are not two ways of saying the same thing.

This guide sets out SEOTEC’s practical approach to observing AI Search. It is a method for making a defined sample interpretable, not an attempt to measure every possible answer on every platform.

Quick answers about AI visibility evidence

Does an AI citation mean someone visited the website?

No. A citation shows that a source was linked in an observed answer. A visit requires separate referral evidence, and a business outcome needs evidence of a further action. Do not turn a citation count into a visitor or customer count.

Does one favorable AI answer prove consistent visibility?

No. It records one answer under particular conditions. Preserve the platform, mode, date, question and relevant location or account context, then repeat a defined sample. Do not treat one result as the answer every prospective customer will receive.

Can two AI visibility percentages be compared directly?

Only when the underlying question sets, platforms, periods and counting rules are comparable. Keep the counts behind the percentages and record changed conditions. A higher percentage from a different sample does not by itself demonstrate an improvement.

Define the event before counting it

A brand mention means the name appears. A citation means a source is linked. A recommendation means the answer presents a company or product as an option for the stated need. A referral means an observable visit came through. A business outcome requires evidence of the resulting action.

One answer can contain several of those events, but none should be assumed from another. A page may be cited to explain a category without its owner being recommended as a supplier. A business may be named without a link. A link can appear without anybody clicking it.

Use those categories as separate columns in the record. It is more informative than combining them into one unqualified “AI win.”

Start with a question set that represents real decisions

Choose a small initial set of relevant questions. Include business identification, service or product selection and the comparisons that matter to the intended customer. A useful starting question is one that could change a decision, not simply a prompt designed to elicit the brand’s name.

Keep branded questions and unbranded category questions separate. “What does this company offer?” tests a different thing from “Which type of service fits this situation?” Both can matter, but combining them into a single visibility percentage can hide the difference.

Record the wording exactly. Small changes may be worth testing later; they should not be introduced silently during a comparison.

Preserve the conditions of the observation

For each run, keep the platform, interface or mode, date, question, relevant location or account context, answer and linked sources. Record unavailable conditions as unknown rather than pretending they were controlled.

Google distinguishes AI Overviews from AI Mode and notes that their responses and links can differ. They should therefore be labeled separately in a research log. Google’s AI features guidance

The IRFĒ archive illustrates why this matters. A report describes a screenshot as AI Overview, while the captured interface visibly says AI Mode. Our case preserves the evidence and labels the surface shown, with the discrepancy disclosed in the caption.

View the IRFĒ source-panel example

Repeat the observation without pretending it is a census

Repeated checks help distinguish one appearance from a pattern within the questions being tested. Use the same questions and document changes in the checking conditions.

A rate such as “the brand appeared in 12 of 20 observed answers” describes that sample. It does not establish the brand’s share of all AI searches, and it should not be compared with another program using a different question set as though the denominators were identical.

Our reporting recommendation is to preserve the counts behind any percentage. Keep meaningful answer changes, new citations, removed sources and incorrect statements available for inspection rather than reporting only a summary score.

Google’s dedicated AI visibility report

Google Search Console now has a Generative AI performance report for Search. Its scope is different from a manually checked set of questions: it reports impressions for links to the site in the covered Google experiences, rather than only the answers selected for a monitoring sample. Google’s documentation says these insights became available worldwide on August 31, 2026. Google’s report documentation · Official report announcement

The Search report covers AI Overviews and AI Mode. An impression means a link to the site was shown according to Google’s counting rules. It is evidence of reported visibility, not a count of unique people, brand recommendations, website sessions or inquiries. The report does not inspect whether a generated description of the business was correct.

These impressions are already part of the overall Search performance data. The dedicated report gives a view of that activity; it does not create an additional set of impressions to add to the ordinary Web total. A report that sums the two would count overlapping activity twice.

What the report’s groupings mean

The available groupings let an owner understand where the reported visibility is concentrated. They also set the limits of what can be concluded:

Dimensions in Google’s generative AI impressions report
GroupingWhat it helps explainInterpretation limit
PagesWhich linked URLs received impressions.Google generally assigns performance to its chosen canonical URL, the representative version of a page. Redirects and duplicate URLs can affect where the record appears.
CountriesWhere the searches originated.A gain in one market does not establish the same gain in another.
DevicesWhether visibility came from desktop, mobile or tablet searches.A change in the device mix can change the overall pattern.
DatesHow impressions changed over the selected period.Report dates use Pacific Time; a differently configured analytics calendar can place activity on a different day.
Web search typeText-based searches versus multimodal searches using images.Multimodal covers experiences such as Lens and image-based searches; combining the types can hide which experience changed.

The current documented report does not provide a list of the customer questions, a transcript of every answer, an AI-source position or a separate AI click-and-conversion metric. Its feature coverage includes both AI Overviews and AI Mode; the current documentation does not describe a separate mode-by-mode breakdown. A percentage labeled “AI CTR” cannot be calculated by dividing all organic clicks by these AI impressions: the numerator and denominator would describe different activity.

Why a chart total can differ from the page rows

Google distinguishes aggregation by property, meaning the site being reported on, from aggregation by page. Its documentation says that two results from the same site in a generative AI feature can count as one impression in the property chart. Page rows can count impressions for the individual pages. Applying a URL filter changes the chart’s aggregation to the URL.

In a hypothetical example, one result contains links to two different pages on the same website. The property total can record one impression while the page table has one for each page. Adding the page rows would then produce two, although the property chart shows one. The difference follows the counting rule rather than proving missing data or twice as many viewers. Report aggregation rules

The page grouping also follows the final linked URL after redirects, with most performance assigned to the canonical URL. An owner may therefore see a representative page in the report rather than every address that could lead to the same content. This matters when a report appears to show no impressions for an alternate URL.

What an absent report or sudden change can mean

Google says the report may be unavailable when a site has not received enough impressions in the covered features. Exclusion through the Search generative AI control is another possible reason. Neither situation can be diagnosed from a missing screenshot alone. The report also excludes Search Labs experiments, so an experimental answer is not a reliable test of whether an appearance should be in this report.

The newest data can be preliminary. Google’s usual reporting limits also apply, including limits on the number of table rows. An export or a selected table is therefore not a complete record of every possible appearance. Google notes that unavailable values shown as “~” or “-” become zeros in downloaded data. An exported zero therefore needs its original report context; it should not silently become evidence of zero visibility when the original value was unavailable.

Changes in logging can also affect a chart. Google maintains a Search Console data-anomalies record; for example, it documented an August 13–17, 2026 logging error in generative AI Search impressions and later restored that data. That history illustrates why a reporting change and a real visibility change are separate explanations to consider. It is not evidence of a current fault in a particular business’s account.

The Google AI Overviews and AI Mode guide explains the feature and inclusion rules. Here, the important boundary is what the report records: covered Google link impressions, with its documented scope and counting limits.

Which evidence answers the question?

First-party reporting and answer observations complement one another. Search Console can reveal a pattern of link visibility across the covered Google features. An observation can preserve the actual wording and sources of a particular answer. Neither replaces evidence of what a visitor did after arrival.

Different evidence for visibility, arrivals and customers
EvidenceQuestion it can answerWhat remains separate
Google’s Generative AI performance reportHow often were links to the site reported in the covered features?The exact answer text, the original question and subsequent business result.
A defined answer-observation sampleWas the business named, accurately described, recommended or linked in these observed answers?The total audience or all the answers customers received.
Recorded website acquisition and eventsWhich identifiable sources brought recorded arrivals, and what activity followed?People who saw an answer without visiting, and influence not preserved in attribution.
Confirmed inquiry or customer recordsWhich inquiries were real, suitable and progressed to business?The full discovery path unless the available evidence actually records it.

For example, a business could see rising Google AI impressions while a small, repeated sample still contains no recommendation of the company. That is not necessarily a contradiction. The report may include other questions and pages; a source link can support a general explanation without recommending the publisher. Recorded visits could meanwhile remain flat. The supported conclusion would be greater reported Google link visibility, with recommendation and traffic results described separately.

Accuracy deserves its own reading. A response can name the right company but describe an unavailable service, the wrong location or an unsupported qualification. A favorable sentiment score would not correct those facts. The original wording, the source page and the current approved business information make that discrepancy inspectable. Likewise, a citation to an article about the business should identify the article’s publisher; it is not automatically a citation of the business’s own website.

A tool’s share-of-voice figure also needs its definition. A percentage of a selected prompt set, a share of counted mentions and Google’s measured link impressions have different denominators. Their trends can be discussed alongside one another, but replacing the prompt set or platform changes the comparison. The existing question-and-condition record is what makes the sample interpretable.

Keep answer observations separate from referral analysis

Analytics can show observable visits from AI platforms when the available attribution identifies them. Google Analytics now has an AI Assistant default channel for recognized assistant sources. Google explicitly excludes AI Overviews and AI Mode from that channel and includes their non-ad traffic in Organic Search. Google’s channel definitions

This means an AI Assistant traffic chart is not a report of all AI-related discovery. Google’s Search features and other assistants occupy different categories, and missing source information can limit attribution. A recorded arrival also does not expose the complete answer or original prompt. Reports should name the channel definition and scope rather than assume the word “AI” identifies one universal traffic source.

Check what your own report measures before quoting a change. Users, sessions and events are not interchangeable. A screenshot with its column headers missing should not become a precise claim about sessions merely because that interpretation sounds plausible.

Likewise, a referral increase does not by itself show which answer caused it. Preserve that uncertainty rather than connecting an attractive screenshot to an unrelated traffic chart.

Check access without confusing it with training permission

A useful technical review distinguishes search discovery from model training. OpenAI documents separate robots.txt controls for OAI-SearchBot and GPTBot. An owner can make different decisions about those uses. OpenAI crawler documentation

Google says a supporting page in its AI features must be indexed and eligible for a search snippet. Its current guidance also requires the site to be included through the Search generative AI control in Search Console; inclusion is the default, subject to an inherited or selected setting. Google does not require a special AI schema or AI text file. Google’s current guidance · Search generative AI control

For a live business site, the task is to inspect the actual controls and visible content. For a public noindex preview, the task is different: preserve the noindex setting until publication is approved. A diagnostic is not permission to change either site’s access policy.

Turn the finding into a specific task

An inaccurate company description may require clearer business information. A missing explanation may belong on an existing service page. A useful source that cannot be reached may need technical attention. A subject with a genuinely separate reader task may justify a guide.

The task should name the page, proposed change, evidence and intended check. “Optimise for AI” is not an implementation instruction.

The guide to service pages and supporting content explains how to choose a home for new information without multiplying pages around every variation of a question.

What a useful report should contain

Keep the questions and conditions available, identify the answer surfaces, distinguish mentions from citations and preserve the original evidence. Report referral activity separately, with the metric labels and dates intact. Connect observed business actions only where the available tracking supports that connection.

The result is a program that can be discussed and improved. You can see what was checked, what changed and why the next task is worth doing.

Explore AI Search services · AI Search Library · SEO measurement guide

Sources

Platform documentation checked October 30, 2026. Reports and channel definitions can change.