Being used by an AI answer, being cited, being visited and creating value are different outcomes. This essay separates them, reads what the 2025 Pew study and the 2026 Google and Bing reports can and cannot tell us, and suggests an honest way to observe AI visibility.
Consider a familiar question: Is our company visible in search? A decade ago, a marketer might have answered by checking rankings, organic impressions, clicks and qualified traffic. These remain useful measures. But a person can now ask a search engine for a comparison, receive a synthesised answer, learn about a product category, and leave without opening any of the pages that informed the response.
A publisher can be present in that process without receiving a visit. A brand can be discussed without being linked. A linked source can receive an impression without being read. A visit can occur without generating any meaningful business outcome.
None of these states is equivalent. Yet they are often presented as a single thing called AI visibility.

Five questions hiding inside one phrase
I find it more useful to separate five questions.
DiscoveredCan the content be discovered? Is the page accessible, indexable when intended, understandable and relevant to the task? This is partly a technical and editorial question.
What can answer it: Crawl and index status, page reports in Search Console and Bing Webmaster Tools.
RepresentedDoes the answer represent the organisation or its ideas? A model may accurately describe a business, reproduce an idea, omit a limitation or confuse two entities. Representation is an issue of meaning, not merely position.
What can answer it: Manual review of real answers, kept with the date, the surface and the exact question.
AttributedIs the source attributed? Being used as background and being visibly cited are not the same. Even a citation may appear in a collapsed source list that few people open.
What can answer it: Google’s generative AI impressions report; Bing’s AI Performance citations. Both in preview or early form.
VisitedDoes the user visit? Referrals can be measured where tracking and reporting permit. But some AI journeys have no site visit at all.
What can answer it: Analytics referrals, where the platform passes them on.
ValuedDoes the exposure create value? An actual business outcome may involve awareness, later branded search, a qualified enquiry or nothing measurable. Attribution becomes especially difficult when journeys cross platforms and devices.
What can answer it: Enquiries, branded search, revenue. Usually indirect, often delayed.
The intellectual mistake is to collapse these distinct outcomes into one number and then optimise for the number.
What current evidence says about clicks
Pew Research Center analysed browsing behaviour from 900 U.S. adults and the Google search pages they visited during March 2025. In the study, users clicked a traditional search-result link after an AI-summary page in 8% of visits, compared with 15% when no AI summary appeared. Links inside the AI summary itself received clicks in 1% of AI-summary visits.1
The pattern is important, but the limitations matter just as much. This was an observational sample, limited to Google and to browsing in March 2025. The data do not establish that the summary alone caused every difference in clicking. Queries that trigger summaries may themselves differ in intent. The study also does not predict how every market, country or future feature will behave.
Fewer clicks when an AI summary appears
Share of visits to a Google search page that included a click on a link. United States, March 2025.
15%
8%
1%
Even with those boundaries, the finding makes a practical point: an answer can satisfy part of the user’s information need before the traditional click occurs.
For publishers, a decline in referral clicks might coexist with increased visibility to users who no longer need to visit. That does not mean the lost visit is automatically replaced by brand value. The appropriate answer is to measure both effects rather than assume either.
The measurement environment changed in 2026
It would now be inaccurate to say that search engines provide no first-party AI visibility data. Google introduced dedicated generative AI performance reports in Search Console on 3 June 2026, with separate views for Search and Discover.2 The reports show impressions inside generative AI features by page, country, device and date. They report impressions, not clicks or the prompts people used.
Microsoft introduced AI Performance in Bing Webmaster Tools in February 2026, covering cited content across relevant Microsoft AI experiences. In June it announced preview capabilities for intents, topics, citation share and comparisons.3 In October, a topics portfolio beta expanded that topic view.4
These announcements represent meaningful progress for observability. They do not turn all AI answers into a fully transparent ranking system. The underlying systems vary, some features are in preview, and a citation-share percentage is not a market share or a measure of answer quality. Microsoft itself describes Citation Share as an observational metric rather than a ranking system or competitive scoreboard, and says it does not represent traffic share.
- Pew publishes its click studyObservational evidence that AI summaries coincide with fewer clicks.
- Bing Webmaster Tools: AI PerformancePublic preview of citations across Microsoft Copilot and AI summaries in Bing.
- Google Search Console: generative AI reportsImpressions inside AI features by page, country, device and date. No clicks, no prompts.
- Bing adds Intents, Topics, Citation Share, CompareIn preview. Described by Microsoft as observational, not a ranking.
- Bing: Topics Portfolio (beta)Groups cited queries into themes, shown only where a site has enough data.
- Measurement has improvedInterpretation remains difficult.
A score is not a theory of discovery
Several commercial products aggregate prompts and answer appearances into proprietary AI visibility scores. Such dashboards may be useful for tracking a consistent set of questions over time. They can also tempt decision-makers into treating a sample as the entire market.
A visibility estimate depends on which prompts were selected, how often tests are run, the geographic and personalised context, which model and product version was used, and how citations or mentions were counted. Changing the prompt set changes the denominator. Changing the model changes the experiment. An unexplained score may conceal more judgment calls than it reveals.
A better reporting habit is to retain the raw question, time, surface, observed answer, source attribution and interpretation. If a dashboard claims improvement, ask whether the improvement is observable in the underlying answers or only in the vendor’s formula.
| Question | Date and time | Surface | What the answer said | Source credited? | Interpretation |
|---|---|---|---|---|---|
| The words a real reader would use | When you looked | Which product and mode | A faithful summary, with caveats it missed | Linked, named, or neither | What it means, and how sure you are |
Discovery begins with work worth referring to
Google’s guidance for AI features emphasises sound search fundamentals and the quality of web content rather than a separate magic technical recipe.5 In practice, that means clear, accessible pages that answer identifiable questions, accurate author and organisation information, sensible internal links, and original material that adds evidence or interpretation beyond what hundreds of similar pages already say.
A generic page about “five advantages of AI” is easy to summarise and easy to replace. A well-documented experiment, a distinctive explanation of a trade-off, or a carefully scoped first-hand field observation provides more specific value. That does not guarantee selection as a source; it gives a publisher a defensible reason to exist independent of selection.
The distinction matters for a personal publication too. If an essay is written only to trigger a machine citation, it has already surrendered the author’s judgment to a measurement proxy. The more useful ambition is to produce writing that is worth finding whether a person encounters it through a search result, an AI answer or another reader.
An observational framework for a real publisher
Start with a small, stable set of genuine reader questions. Group them by task rather than by keyword alone: definitions, comparisons, judgment calls, and complicated problems. For each question, record ordinary search impressions and clicks alongside any first-party generative-AI visibility data available. Where appropriate, manually inspect answers for correct brand representation, cited sources and missing caveats. Record the sample design openly; do not pretend the sample covers all possible queries.
Look for directional evidence over several periods, not a single impressive screenshot. Compare site engagement with the actual commercial or editorial goal. A few genuinely interested readers who stay with a detailed article may matter more than many fleeting appearances in an answer nobody can verify.
When a report says “AI visibility grew 40%,” the next question should be: forty percent of what, observed how, in which surfaces, over what period, and with what effect?
There is no single metric that answers every part of that question.
The question worth carrying forward
The old language of search encouraged us to imagine a page competing for a position in a list. The newer environment forces a broader question: what role does our knowledge play when other systems construct a response on the reader’s behalf?
Sometimes the role will be a credited source. Sometimes a visit. Sometimes a correctly understood idea. Sometimes nothing at all.
The goal should not be to claim that every appearance is a victory. It should be to understand which appearances are real, which are meaningful, and which reveal that the publication has earned a place in someone else’s process of learning.
ABOUT THIS ESSAY
Published 20 November 2026.
Citations & references
- Pew Research Center, "Google users are less likely to click on links when an AI summary appears in the results", 22 July 2025. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
- Google Search Central, "Introducing Search Generative AI performance reports in Search Console", 3 June 2026. https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports
- Microsoft Bing, "New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare", 16 June 2026. https://blogs.bing.com/search/June-2026/New-AI-Visibility-Insights-in-Bing-Webmaster-Tools-Intents-Topics-Citation-Share-Compare
- Microsoft Bing, "Introducing Topics Portfolio (Beta) in AI Performance", 13 October 2026. https://blogs.bing.com/webmaster/2026/10/introducing-topics-portfolio-beta-in-ai-performance/
- Google Search Central, "AI features and your website". https://developers.google.com/search/docs/appearance/ai-features
Connections
THOUGHT ATLAS
Marketing & BrandCATEGORIES
SEO, AEO, GEO & AI Search
