AI visibility
Google finally measures AI-search visibility. Here is what it still cannot tell you
July 17, 2026
Google gave website owners something they have been asking for: a dedicated view of appearances inside AI Overviews, AI Mode, and generative features in Discover. That makes one part of AI-search visibility measurable for the first time.
It does not make the measurement problem disappear.
Google's June 3 announcement says the new Search Console reports show impressions, pages, countries, devices, and dates. The initial rollout covers a subset of websites. Based on the fields Google lists, the report does not identify the triggering query, the wording of the answer, the prominence of your citation, or whether the answer described your business correctly.
That is still a major improvement over screenshots. It is also easy to oversell.
What an AI-search impression proves
An impression proves that a URL from your site appeared somewhere in a Google generative-search experience counted by the report. You can see which page appeared, when it happened, the country, and, for Search, the device class.
That gives a business a baseline. If an answer page is rebuilt and its generative-search impressions rise over the next six weeks, you have evidence that Google is finding a use for it. If only the homepage appears, the site may lack pages that answer the narrower questions Google fans out behind a broad request.
The rollout caveat matters. Google said the dedicated reports were initially available to a subset of sites. A missing report does not prove a site never appeared in generative search, and a quiet week on a small site may say more about sample size than content quality. Record availability and reporting dates before comparing one period with another.
Google explains that AI Mode and AI Overviews can run multiple related searches across subtopics and data sources. One customer question can therefore create several opportunities for a useful page to be included. That favors specific pages with clear facts over a homepage trying to cover everything.
But an impression is not an endorsement. The URL could sit behind a disclosure, support a minor detail, or appear in an answer that sends no visit and no customer.
The report leaves four questions open
The first missing piece is the prompt. A local accounting firm needs to know whether it appeared for "small-business CPA near me," "who handles restaurant payroll in Columbus," or a research question with no buying intent. Aggregate impressions cannot separate those cases.
The second is presentation. A citation can support a positive recommendation, a neutral fact, or a warning. Search Console records the appearance, not the sentence around it.
The third is accuracy. A model may use the correct URL and still report an old price, wrong service area, or discontinued product. That is visible only when someone reads the answer.
The fourth is commercial value. The report announcement does not list clicks, leads, booked calls, store visits, or revenue in its dedicated AI view. A thousand impressions on a research query can be less useful than one accurate recommendation to a buyer ready to call.
Search Console can tell you that Google used your page. It cannot tell you whether the answer helped your business.
A measurement stack that does not pretend
I would track AI visibility in four layers:
- Google exposure. Record generative-search impressions by page, country, device, and week. Keep the classic Search data beside it rather than blending the two into one unexplained score.
- Controlled prompt checks. Run a fixed set of buyer questions in Google, ChatGPT, and Perplexity on a schedule. Save the full answer, cited URLs, date, location context, and account state. Do not quietly change the prompts when the result looks bad.
- Source-page changes. Log what changed on the cited page and when: business facts, product details, author evidence, reviews, structured data, or internal links. This makes later movement interpretable.
- Business outcomes. Ask qualified leads how they found the company. Track calls, forms, branded searches, and sales that mention an AI recommendation. Small numbers are fine. Invented attribution is not.
The prompt checks fill the context gap. Search Console adds platform data that an outside audit cannot manufacture. Business outcomes keep both from turning into vanity metrics.
Prompt tests also need boring controls. Run them from the same account state and location when possible. Keep the wording fixed. Record when an engine declines to answer. Generated results vary, so one appearance is an observation, not a trend. A repeated result across dates and engines is stronger evidence, especially when it points to the same source page.
Google also killed a convenient sales pitch
Google's May guidance on optimizing for generative search says established search fundamentals remain the foundation. It calls for useful, unique content and specifically addresses common AEO and GEO misconceptions. There is no special markup that purchases inclusion.
That is inconvenient for anyone selling a secret AI-ranking switch. It is good news for businesses willing to publish better source material: accurate service pages, useful comparisons, current product facts, visible authorship, and third-party evidence that agrees with the site.
The new report raises the standard for AI-visibility work. A consultant should be able to show both the answers observed from the outside and the exposure Google records on the inside. Screenshots remain useful evidence, but they are no longer the whole measurement system.
Get a baseline before changing the site
My AI visibility audit records buyer prompts across the major engines, checks the source pages behind the answers, applies the agreed fixes, and repeats the same prompts. I also use my own businesses as public tests, including the Bip's Bites local-commerce case study.
The initial check is free. If the engines already recommend you accurately, I will say so.