Why Measure AI Visibility Separately
A search result can now be seen without a link being clicked, and it can be seen inside the answer rather than under it. Two independent measurements describe the cost of that shift for websites. Pew found that people clicked a traditional result in 8% of visits to a results page with an AI summary, against 15% without one, and clicked a link inside the summary in 1% of visits[6]. Ahrefs found that an AI Overview correlated with a 34.5% lower click-through rate for the top-ranking page[7].
Until 2026 there was no first-party count of how often a site appeared inside those answers. Search Console’s Generative AI report now provides one: impressions in AI Overviews and AI Mode, by page, date, country and device[1]. Google states that no special optimization or structured data is needed to appear in these features[3]. The practical question is therefore not which technique unlocks them but which pages they use, and that is a question the report can answer.
Instruments
| Instrument | What it counts | What it cannot say |
|---|---|---|
| Search Console, Generative AI report | Impressions of a site’s links inside AI Overviews and AI Mode, combined; by page, date, country, device | Clicks, position or queries; which of the two features showed the link |
| Search Console, Performance report | All Web impressions and clicks, which include the AI-feature impressions | Which impressions were AI; the denominator for AI share only |
| Bing Webmaster Tools, AI Performance | Citations in Copilot, Bing AI summaries and partner experiences; cited pages; a sample of grounding queries with intent labels and citation share | Complete query activity (the queries are a sample); clicks; history before a site is verified |
| GA4, traffic acquisition | Sessions whose source names an AI assistant, with engagement | Visits from assistants that strip the referrer; these are floors |
Table 1. Instruments. AI share, used throughout, is AI-feature impressions divided by all impressions for the same property or page and dates[1][2].
Google aggregates the report’s chart by property and its page table by page, and says the two totals can differ. We quote chart totals for growth and page-table totals for every share, and we name the basis wherever a figure appears. Bing’s definitions are Microsoft’s[4][5].
Properties and Windows
We read the Generative AI report for all 81 properties in our Search Console account[8]. The report’s history begins May 18, 2026, and 40 distinct properties have data for the whole window through October 16; a URL-prefix property inside a domain property is counted once. Of the 40, 17 are client sites, most of them local service businesses, and 23 are other properties, most of them publishing and tool sites the studio operates. Client properties are described here by vertical and state only; studio-operated properties appear only in aggregate. Client data is used for measurement only.
Growth compares the first 28 days of the window (May 18 to June 14) with the last 28 (September 19 to October 16). Weekly series use full Monday to Sunday weeks, May 18 to October 11; the partial week of October 12 is dropped rather than extrapolated.
Finding 1: The Expansion Is the Baseline
AI-answer impressions grew 2.58× across the 40 properties while all search impressions grew 1.06×. The client sites grew 3.01× and the other properties 2.45×. Every group gained, in a period when Google brought AI answers to more searches. The median property grew 5.1× (n = 27), higher than the aggregate because small properties started near zero; the aggregate is the better baseline.
The bar for a single-site claim
Any claim that one site gained AI visibility in this period has to be read against roughly 2.6×, the growth the account as a whole saw over the same windows. A site that grew 2.3× gained less than its peers did; a site that grew 7.6× gained about three times as much. The baseline cannot separate Google’s expansion from everyone else’s optimization, but it removes the most common overstatement.
Finding 3: Answer-Shaped Pages Carry the Visibility
For the five client sites we classified every page by URL and compared each class’s share of search impressions with its share of AI-answer impressions. On four sites the disproportionate class was guides and articles: pages that drew 3% to 21% of search impressions carried 49% to 72% of AI impressions. On the fifth, an entertainment booking agency, it was performer profiles marked up as Person and MusicGroup entities. We call both answer-shaped: pages whose purpose is to answer one question or describe one entity.
| Property | Answer-shaped pages | Pages with AI | Share of search | Share of AI | AI rate | AI rate, rest of site |
|---|---|---|---|---|---|---|
| HVAC contractor, Ohio | Guides and articles | 29 | 3.3% | 50.2% | 24.2% | 0.8% |
| Car rental operator, New Hampshire | Question-led guides | 41 | 5.6% | 51.3% | 33.1% | 1.9% |
| Fencing contractor, Kansas | Resource articles | 14 | 9.9% | 49.2% | 18.3% | 2.1% |
| Tree service, Kansas | Blog articles | 33 | 21.1% | 71.8% | 14.8% | 1.6% |
| Entertainment booking agency, Florida | Performer profiles | 118 | 16.0% | 35.3% | 7.1% | 2.5% |
Table 3. Page tables of the Generative AI and Performance reports, May 18 to October 16, 2026. AI rate is AI impressions as a share of the same pages’ search impressions. The answer-shaped pages’ AI rate ran 30×, 17×, 9×, 9×, 3× the rest of each site’s, in the order shown.
The pattern is descriptive. Answer-shaped pages also differ from service and location pages in topic and intent, so the table does not show that the format causes AI use. It shows where AI visibility lives on these sites, which is the information an owner needs to decide what to build next.
Finding 4: Bing and GA4 Agree on Direction
Bing’s AI Performance report covers only what it saw after a site was verified there. For most of the 71 sites in our Bing account its history begins on October 1 or 2, 2026; 33 had citations by October 16 and 4 had more than 30 days of history[9]. One client property, the Ohio HVAC contractor, has six months: 939 citations across 41 pages. Of the citations attributed to a page, 68.1% went to its guides, the same class of page that leads its Google report.
| Month | Apr 18–30 | May | Jun | Jul | Aug | Sep | Oct 1–16 |
|---|---|---|---|---|---|---|---|
| Citations | 46 | 93 | 92 | 86 | 148 | 346 | 128 |
Table 4. Bing AI Performance, six-month view (April 18 – October 16, 2026), HVAC contractor, Ohio. In the grounding-query sample, citations by intent were Commercial 72, Learn and Solve 70, Comparison 37, Local 12, Research 9; one query naming a competitor is excluded.
GA4 measures the other end: visits that assistants send[10]. On all five client sites ChatGPT accounts for at least 89% of assistant referrals, and on all five its visitors engaged at a higher rate than the site’s average visitor. The shares are small, from under 1% to 4.5% of sessions, and they are floors.
| Property | ChatGPT sessions | Share of all sessions | Other assistants | Engagement, ChatGPT | Engagement, site |
|---|---|---|---|---|---|
| HVAC contractor, Ohio | 103 | 0.85% | 2 | 54.4% | 33.2% |
| Car rental operator, New Hampshire | 153 | 3.90% | 3 | 76.5% | 67.7% |
| Fencing contractor, Kansas | 49 | 1.23% | 6 | 57.1% | 49.8% |
| Tree service, Kansas | 15 | 0.72% | 1 | 66.7% | 40.0% |
| Entertainment booking agency, Florida | 189 | 4.51% | 8 | 64.0% | 36.8% |
Table 5. GA4 traffic acquisition, May 18 to October 16, 2026. Other assistants are Perplexity, Gemini, Copilot and Claude combined.
Implications for Practice
Four working rules follow from these observations, and we apply them to client work. First, measure AI share per page, not AI impressions per site, because the site total hides where the visibility lives. Second, compare any growth with the expansion baseline for the same windows. Third, build answer-shaped pages where the questions are: cost, comparison, eligibility and entity pages, written to be quoted, with structured data that restates the visible text rather than adding to it[3]. Fourth, read Google, Bing and GA4 together; they count different things, and agreement among them is the strongest signal available.
Our Results page applies this method to individual client engagements, named with their owners’ figures.
Limitations
The Generative AI report is in beta, holds five months of history, and has no pre-period for any property, so no before-and-after comparison is possible. It has no query dimension and combines AI Overviews with AI Mode[1]. The roster is the account we manage, not a random sample, and every client site in it receives our work, so the baseline is a reference group, not an untreated control. Page classes are assigned by URL and differ in topic and intent as well as format. Bing’s grounding queries are a sample and its history starts at verification[4]. GA4 undercounts assistants that strip referrers. Impressions are not clicks, and nothing here measures leads, bookings or revenue.