Cited, Not Clicked: What Visibility Still Doesn’t Tell You

AI visibility reports show where links appear. They leave harder questions open: what the answer takes from a source, whether the source receives credit, and what value follows.

By Stefano Zoia · 13 October 2026 · 6 min

Google’s dedicated generative AI reports in Search Console count how often links to a site appear in AI features. They do not provide a separate click metric. The reports launched in June and completed their worldwide rollout on 31 August 2026. [1]

That makes one part of the picture easier to see. It does not settle whether those appearances are useful.

A link can appear without contributing much to an answer. A source can contribute useful information without its brand being remembered. Even an accurate, clearly attributed answer may leave a publisher with no visit and no revenue.

These are different outcomes. Counting citations alone cannot distinguish them.

What the dashboard actually measures

Google defines an impression in the generative AI report as a link to a site being shown to a user within a generative search feature. It is not a count of how often a model fetched or read a page. [2]

The absence of a click metric in that dedicated report also does not mean clicks have disappeared. Google’s general Search Console documentation explicitly counts clicks on external links in AI Overviews. [3]

The reporting gap is narrower: a dedicated view of appearances does not, by itself, explain the traffic or value associated with them.

That matters when an organisation uses rising AI impressions as evidence that its content strategy is working. The number establishes exposure under the platform’s counting rules. It leaves the next questions unanswered.

Fewer clicks still matter

Pew Research Center examined nearly 69,000 Google searches made by 900 US adults in March 2025. Users clicked a traditional search result on 8% of visits with an AI summary, compared with 15% without one. Links inside the summaries received clicks on 1% of visits. [4]

This was observed behaviour, not a randomised experiment. The comparison does not isolate every difference between searches that produced a summary and those that did not. It nevertheless documents an important pattern within that sample.

A Digital Content Next review published on 5 October brings together further research on declining referrals and the limited replacement traffic arriving from AI platforms. Its central commercial concern is that publishers can supply information without receiving the visit through which they previously earned revenue or developed a direct relationship. [6]

For a publisher supported by advertising or subscriptions, that lost opportunity matters. For a consultancy or manufacturer, useful exposure might contribute to a later enquiry. Neither outcome can be inferred simply from the content appearing in an answer.

The business model determines what value needs to be measured.

Being cited is not the same as being used

An April preprint by Zhang Kai, He Xinyue and Yao Jingang proposes a distinction between citation selection and citation absorption: whether a source is selected, and how substantially it appears to contribute to the generated answer. [5]

Using a dataset covering three platforms and 602 prompts, the authors report differences between citation breadth and apparent contribution. Their influence score combines observable features, including reference placement, coverage and textual similarity.

It is a constructed proxy, not a direct observation of the model’s reasoning or a measure of influence on the reader. The study is descriptive, and its authors explicitly limit causal claims.

The useful takeaway is therefore a question, rather than a new universal score: what did the answer actually take from the source?

A citation list alone does not answer that.

Contribution does not guarantee credit

Consider a company that publishes an original comparison of maintenance costs. An assistant uses its figures to explain the trade-offs between two products.

Several outcomes are possible.

The answer might name the company and link clearly to the comparison. It might include the link among several sources without explaining whose analysis it used. It might reproduce the figures while dropping the conditions that made them comparable.

These are hypothetical examples, but they expose three different concerns: contribution, attribution and accuracy.

Even in the strongest case, the company still does not know whether the user noticed its name, trusted the analysis or changed a purchase decision.

This is where “absorption” becomes insufficient as a business objective. Information can be incorporated into an answer without returning recognisable value to its originator.

Four questions to keep separate

A practical measurement approach should preserve those distinctions.

The first three can be investigated through recorded answers and source comparison. The fourth requires evidence about people and their behaviour.

Combining them into one visibility score risks hiding the very distinctions the organisation needs to understand. A frequent citation attached to an inaccurate claim is a different result from an occasional, accurate recommendation that leads to a qualified enquiry.

Start with a repeatable sample

Choose a small set of questions that matter to your audience. Include different intentions: learning about a problem, comparing options and checking a specific claim.

Record the exact prompt, platform, date, language and relevant location or account context. Save the answer and its links. Repeat the exercise over time.

For each answer, examine whether your organisation appears, what information seems to be used, how it is attributed and whether important qualifications survive. Where several sources contain the same information, record uncertainty about its origin rather than assuming the answer came from yours.

This is an audit of sampled responses. It does not estimate how often the wider public encounters those responses.

Alongside it, continue measuring relevant referral traffic and business outcomes. Branded searches, direct visits and prospects mentioning an assistant can provide additional signals. They remain imperfect: campaigns, seasonality and other channels may explain changes occurring at the same time.

For example, “How did you first hear about us?” can reveal a route that referral analytics missed. It cannot establish the total contribution of AI across all customers.

Improve the source without writing for a score

The editorial response is to make evidence useful and checkable.

An original figure should carry its date, scope and method. A comparison should explain its criteria. A recommendation should retain the conditions under which it applies. A paragraph should remain accurate when read outside the surrounding page.

These are standards for readers as well as for systems that summarise content. They make reuse easier to inspect, but do not guarantee citation, faithful reproduction or commercial return.

The mistake would be to replace the pursuit of clicks with the pursuit of an equally incomplete citation count.

As with measuring completed work rather than AI activity, an intermediate signal becomes useful when its relationship to the intended result is clear.

A visit remains valuable evidence of arrival. A citation is evidence of inclusion. Apparent contribution tells us something about the answer.

What any of those means for the business is the next question—not a conclusion already contained in the number.

Stay close.

— Stefano

Sources

1. Google Search Central, “Introducing Search Generative AI performance reports in Search Console”, 3 June 2026, with an update confirming worldwide rollout on 31 August 2026.
2. Google Search Console Help, “Generative AI performance report (Search)”, accessed 13 October 2026.
3. Google Search Console Help, “What are impressions, position, and clicks?”, accessed 13 October 2026.
4. Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 July 2025. The analysis concerns US browsing behaviour in March 2025.
5. Zhang Kai, He Xinyue and Yao Jingang, “From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms”, arXiv:2604.25707v2, 29 April 2026. A descriptive preprint proposing an observational proxy for source contribution.
6. Michelle Manafy, “The evidence is clear: AI search is costing publishers”, Digital Content Next, 5 October 2026. A synthesis from a publisher trade association, used here for its discussion of publisher economics.