Your AI Visibility Score Is a Domain. The AI Cites a Page. I Counted Which Ones.

AI attribution is broken, so we report AI visibility. Then we report it for the whole domain.

The consensus this quarter is that you cannot attribute AI search the way you attribute Google. No query data, no clean referrer, answers that change week to week. So the industry moved the KPI. Report visibility, mentions, and citations instead of clicks. Fine. I agree with that part.

Here is the part I do not agree with. Almost every AI visibility number I see in a client report is for the domain. Your site was cited 40 times. Your share of voice is up. Great. Which page? Because the AI never cites a domain. It cites a URL. And if one page is doing all the work, the domain number is a weather report for a city when you needed the address.

Semrush's October update to affiliates said the Visibility Overview report in the AI Visibility Toolkit now takes a subdomain, a subfolder, or a single URL, not just the root domain. Its knowledge base entry already lists a Cited Pages metric: your domain's pages cited in AI-generated answers. Before I told anyone to use it, I wanted to know whether page level actually changes the story for a local business. So I counted.

786 citations of a business's own site, sorted by the page that earned them

The data is our own. Every month our app asks four engines the same local service prompts for every business we track: ChatGPT, Perplexity, Gemini, and Google AI Mode. It stores every source each answer cites, down to the full URL. Four runs, July 11 to October 1. 67 businesses. 3,848 answers that completed. In 786 of the citations, the source was the business's own website. 48 of the 67 businesses got cited at least once. The other 19 never did.

Those 786 citations landed on 206 different pages. I sorted every page by what it is. Service pages: 269 citations, 34 percent. Service pages with a city in them, the roof-repair-in-somewhere kind: 239, 30 percent. Homepages: 213, 27 percent. City-only pages: 39, 5 percent. About, contact, FAQ, reviews, menus, and blog posts combined: 26, about 3 percent. Blog posts alone: one citation. One, out of 786.

Now the concentration. For the median business, its single most-cited page carried 63 percent of all its citations. Across everyone, each business's top three pages carried 78 percent of the owned citations. 12 of the 48 businesses were cited on exactly one page, ever. Their entire AI presence is one URL.

And the size of the site does not matter much. For 16 businesses we have a full crawl on file. Those sites have 820 pages between them. 85 of those pages were ever cited, by any engine, in any run. The median site has about 48 pages and three of them get cited. About 7 percent.

Last one, and it is the one that stung. I matched each cited page against the service in the prompt that triggered it. Ask about one service, does the AI cite the page for that service? Of the 573 citations that were not homepages, 185 went to a page that was not about the service being asked, a different service or a city page. Add the 213 homepage citations and 398 of 786, about half, were the AI reaching for a page other than the one built to answer the question.

The domain number is an average of one good page and a lot of furniture.

If a business has 40 citations and 25 of them are one page, the domain trend line is mostly that page's trend line. When it drops, the domain drops, and you go hunting through the whole site for a problem that lives at one address. When it rises, you take credit for a sitewide strategy that one service page earned.

The engines also do not agree on what to cite, which is another reason not to blend them. Perplexity cited the business's own site in 35 percent of its answers. Gemini did it in 7 percent, Google AI Mode in 5, ChatGPT in 4. When Perplexity cited the site, it picked the homepage 31 percent of the time. Gemini and AI Mode picked the homepage 18 and 19 percent of the time and went deeper, to service and service-plus-city pages. Same business, same prompt, different door.

Industry shifts it too. Restaurants and shops got homepage citations 36 percent of the time, because for a restaurant the homepage is the menu and the hours. The three legal and financial firms got service-page citations 58 percent of the time. The six health, dental, and wellness practices got city-only pages cited 16 percent of the time, three times the overall rate. Home services, 31 businesses and most of the data, split almost evenly between service pages, service-plus-city pages, and homepages. If that sounds familiar, it is the AI version of the positions 4 to 20 test: a few pages you already have do almost all the work.

And none of this is stable. In the ChatGPT volatility test one in five prompts flipped on whether the business was named between two monthly runs. If the answer can move that much, you want to know which page it moved off of, not just that the domain went down.

Report by page. Fix by page.

One. Find the page that carries you. For most local businesses it is one service page or one service-plus-city page. Know its URL by heart. Watch it like a hero product, because it is one. If it breaks, gets redirected in a redesign, or loses its title, your AI visibility goes with it.

Two. Make every money service its own page, and make that page the obvious answer. A third of the non-homepage citations went to the wrong service. That usually means the right page is thin, missing, or buried, so the engine grabbed the closest thing it could find. One service, one page, the service name in the title and the URL, the city in the body.

Three. Stop expecting the blog to do this job. One citation out of 786. For local service prompts, the engines cite the page that sells the service, not the post that explains it. Blog posts still help people and still build topical depth. They are not where AI citations land for a plumber.

Four. Split your report by engine. A blended number lets Perplexity's 35 percent hide ChatGPT's 4. They behave differently, they pick different pages, and they need different evidence. If your report shows one line, ask for four.

Five. Keep the methodology fixed. Same prompts, same cadence, same page buckets every month. The reason this test works is that the prompts never changed. That is also the whole argument of the AI search KPI check.

Point the report at the page, not the domain

Our app does this because I built it to. You do not need to. The Visibility Overview report in the AI Visibility Toolkit tracks citations and cited pages across ChatGPT, Gemini, Google AI Overviews, and AI Mode, and per Semrush's October update it now takes a subfolder or a single URL as the target. That is the test above, done in a search bar. Put in your top service page and watch it as its own line. Put in your service folder and see whether the rest of your service pages are pulling any weight at all.

The other half is knowing what that page is worth. The Traffic and Market Toolkit shows the traffic picture by source, so when the page that carries your AI visibility also carries your traffic, you know which URL to protect first. Semrush runs a free trial, seven days, cancel anytime, which is enough time to find your one page.

The AI does not know your domain. It knows a handful of your pages.

3,848 answers. 786 citations of the business's own site. 206 pages. Most businesses live on one to three of them, a site of 48 pages gets three cited, and the blog got one citation total. A domain-level AI visibility score averages all of that into a single number and hides the only thing you can act on, which is the address.

Find your page. Protect it. Build the missing ones. Then report AI visibility the way the AI hands it out: one URL at a time.