AI SEO Case Study: 297,901 AI Citations Across Four Properties I Own

Most AI SEO case studies are a traffic chart with a new label

The genre has a shape now. A percentage growth figure, a line going up, the words “AI search visibility”, and underneath it a Google Analytics screenshot measuring the same organic traffic everyone was measuring in 2019.

Nothing in it distinguishes an AI result from an ordinary one.

The opponent is that substitution. AI visibility is a separate measurement with its own reports, and if a case study can’t name which report a figure came from and over what window, it hasn’t measured AI anything.

What follows is four properties I own, all pulled on 12 August 2026, the reports named per figure, and the surfaces I couldn’t measure listed rather than implied.

Finding one: citation and traffic move independently

ZPlatform.ai took 1,420,120 Google impressions over 16 months and returned 8,666 clicks. A 0.61% click rate. Over six months the same pages were cited 241,500 times by Copilot, which works out at a minimum of 97 citations for every organic Bing click the site received.

CoimbatoreJunction.in makes the point harder, because there the two lines move in opposite directions. Clicks fell across the window while Copilot citations quadrupled to 40,101 in 180 days, and the single most-cited page on that property converts at 0.09%.

Two pages on that site have near-identical organic impression volume and opposite outcomes. A short ward-list lookup people click through to, and a 47-question reference guide with real pricing tables that the SERP and the models can answer from without ever sending anybody anywhere.

Same site, same author, same month. The difference is page shape.

If you’re reading a traffic report to judge your AI visibility, that’s the effect you’ll misread. Falling clicks on a page being cited more is what success looks like in this channel, and it’s indistinguishable from failure on the dashboard you’re probably checking.

Finding two: the rate runs against scale

MakeMoneyOnline.lk is the smallest property in the portfolio at 71,572 impressions. 8.1% of those impressions come from Google AI features, across 124 pages, and that’s the highest rate here by a distance. Roughly five times the rate of the largest property.

I expected the opposite.

The assumption going in was that AI visibility scales with site size, and that a site this small would show up in AI features rarely if at all. One property can’t isolate a variable, so this is a lead rather than a conclusion, but the plausible mechanism is specificity: the site answers a narrow set of questions about Sri Lankan online income properly, including which methods don’t pay, and a narrow well-answered subject looks easier for a model to attribute than a broad partially-answered one.

It’s now the first thing I test on anything new, ahead of scale.

Finding three: citation share follows query shape

VelankanniShrine.in holds 100% of Copilot’s citations for the flag hoisting date query, and 84% for the feast itself, out of 16,300 citations in total. It does that from an average position of 9.1.

The pattern across all four properties is that citation concentration is set by what the question is, not by how well the site is optimised. A factual date query has one right answer and one source gets it. A “recommend me a site” query resolves to a single destination instead, which is why another property here puts 87% of its AI appearances on one URL and gets an average of two pages cited against 35 on the city guide.

That distinction decides where your work goes. Chasing citation breadth on a query shape that will only ever reward one page is wasted effort, and building one hero page for a subject that rewards a long tail leaves most of it on the table.

Look at what the query is asking for before you commission anything.

What produced these numbers

The mechanism is the same across all four, and none of it is specific to AI.

Answer the question in the first line, in the shape it was asked. Real HTML tables for anything numeric, never an image of a table, because an image is invisible to a crawler and useless to a model. One comprehensive page per question rather than several thin ones. Plain sentences, because a model rendering your fact into another language will drop it if the sentence is clever. First-hand detail an aggregator can’t copy, whether that’s three hours inside a piece of software or which counter at the booking office actually opens.

The local SEO cut of this same evidence reaches those four decisions from a different direction, on two properties with nothing in common except the structure, which is the closest thing to replication I have.

The report you haven’t opened

Then measure the right thing.

The AI Performance report in Bing Webmaster Tools and the Generative AI features report in Search Console are both free, they’re probably already sitting in an account you own, and neither one appears in the traffic report you’re looking at.

If you haven’t opened either of them, you don’t currently know whether your site is being cited.