rovoki.Private beta

The mechanism

How an AI answer decides which brand to name

Search returned ten links and let the reader choose. An AI answer names one or two brands and moves on. That choice is not random, and it is not a ranking. Here is how it actually gets made.

Every figure on this page opens into its source. Where the evidence is thin or contested, the drawer says so. Nothing here is measured by us.


01

The answer is written, not looked up

When a buyer asks which tool to pick, the engine does not consult a league table. It runs its own searches, reads a handful of pages, and writes a sentence. Your brand is either in the pages it read, or it is not in the answer.

This is why the old instinct only half works now. Ranking higher still helps you get read. It no longer decides whether you get named.

What the measurements show

Ahrefs measured how often AI Overview citations come from pages already ranking in the top ten organic results: 37.9%, across 863,000 search results in March 2026. A year earlier the same measurement was 76%. The fall is attributed to engines running more of their own follow-up searches rather than reading the visible results page.

Rank still matters as a proxy. A separate study of roughly 98,000 ChatGPT citations found pages ranking first were cited at about 3.5 times the rate of pages past position twenty. It is an advantage, not a gate.

A widely repeated figure claims 92% of AI Overview citations come from the top ten. We traced it to a vendor blog with no study attached. It is not supported.

Source: Ahrefs, AI Overview citations and the top 10, March 2026

37.9%

of AI Overview citations come from a page already in the top ten. It was 76% a year earlier.


02

Most of what it reads is someone else's list

The single largest source of AI citations is not brand websites, review sites or news. It is ranked lists. The answer naming your competitor is usually reading a post titled best tools for something.

Which means the sentence a buyer reads about your category was, in most cases, shaped by whoever last updated a listicle. In this category those lists are frequently written by the vendors inside them.

How large the effect is

Evertune examined roughly 25,000 of the most-cited URLs across six engines, drawn from about 400 million citations in March and April 2026. Listicles accounted for 63% of all citations, and between 71% and 86% of those were ranked lists rather than unordered ones.

A separate study of high-intent software questions put listicles at about half of citations overall, rising to 70% for questions phrased as alternatives to a named product.

Source: Search Engine Land, on Evertune’s citation study, 2026

63%

of all AI citations point at a listicle. Ranked lists dominate the rest.


03

Being talked about beats being linked to

The strongest published correlation with getting named is not backlinks and not domain authority. It is people mentioning your brand in text, without linking, and mentioning it on YouTube.

SignalCorrelation
Mentions on YouTube0.737
Brand mentions in text, unlinked0.664
Branded anchor text0.511
Domain rating0.266
Backlinks0.19

That is close to an inversion of how most teams spend their effort. Getting written about, in plain text, with no link attached, tracks more closely with being named than link building does.

Method, and the caveat that matters

Ahrefs measured 75,000 brands against millions of AI responses and reported Spearman correlations for ChatGPT, Google AI Mode and AI Overviews. The figures above are the ChatGPT column.

The authors state plainly that correlation is not causation, and it is worth holding that. A brand mentioned constantly on YouTube is probably a brand people already talk about. What the numbers support is a shift in where attention goes, not a mechanical lever.

Source: Ahrefs, AI brand visibility correlations, 2026


04

The most recommended fix does nothing

Almost every guide to getting cited by AI tells you to add structured data to your pages. It was tested against a control group. The effect on AI Overview citations was slightly negative.

We are including this because it is the clearest example of the category’s problem. Advice gets repeated until it sounds settled, and almost none of it has been measured against a control.

The test

Ahrefs tracked 1,885 pages that added JSON-LD structured data against roughly 4,000 control pages between August 2025 and March 2026. AI Overview citations moved −4.6%. Google AI Mode moved +2.4% and ChatGPT +2.2%, both small enough to be indistinguishable from noise.

Google has separately said its AI features do not read llms.txt, the other fix commonly recommended alongside schema.

Structured data still has uses. Rich results in ordinary search are real. Getting cited by an AI answer is not one of them.

Source: Ahrefs, structured data and AI citations, 2026

−4.6%

change in AI Overview citations after adding structured data, against a control group.


05

Ask in Hindi, and the answer comes from English

This is the finding almost nobody is watching, and it decides how brands are described to a very large number of buyers.

When a question is asked in Hindi, the engine tends to retrieve English sources and answer from those. The model writes fluent Hindi. What it read to get there was mostly not Hindi. The most-cited single source for Hindi questions was English Wikipedia, ahead of every Hindi-language publication.

79.3% accurate

Answers to Hindi questions.

89–91% accurate

The same measurement for every other language region tested.

An Indian brand is therefore being described to Hindi-speaking buyers at roughly double the error rate, out of sources it has never thought about, in a language it did not choose.

The study, and what it could not answer

A Stanford-led team put 2,100 questions to six commercial chatbots across six language regions in February 2026, producing 12,600 model-question pairs. For Hindi questions, English Wikipedia was the single most-cited domain at 3.1%. Of the top Indian sources, only one was primarily Hindi-language. The authors call the pattern an Anglophone retrieval pivot and are explicit that it is a retrieval problem, not a fluency one.

The limitation they flag themselves: every query was issued from servers in the United States, which may have made English sources more likely to be returned. Nobody has yet run the same test from Indian infrastructure. Until someone does, the direction is solid and the size of the effect is not settled.

Source: Suzgun et al., evaluating commercial AI chatbots as news intermediaries, February 2026


06

Ask the same question twice and you get two answers

This is the part that makes most AI visibility reporting unreliable, including reporting sold as a product.

These systems do not return a fixed result. Run the identical question on the same afternoon and the set of brands named will partly change. Any number produced from a single run carries that variation inside it, and almost nobody reports how many runs their number came from.

It is also why we will not put a ranking position on anything. A position implies a stability that does not exist.

How much it moves

Engineers at Thinking Machines Lab sent 1,000 identical requests at temperature zero, the setting meant to make output deterministic, and got 80 different completions. The cause is in how requests are batched on the server, so it cannot be fixed from the outside by changing a setting.

Measured on brand mentions specifically, repeats of the same question on the same day overlapped by a Jaccard index of 0.327 to 0.477. In one study ChatGPT returned no citations at all on 57.8% of runs.

Source: Thinking Machines Lab, on non-determinism in LLM inference, 2025

80/ 1,000

different answers from a thousand identical requests, at the setting meant to remove randomness.


What follows from all of this

Three things. Being named is mostly decided by what other people publish about you, not by what you publish about yourself. The answer changes by language, and in India it changes in a direction that works against you. And a single measurement of any of it is close to meaningless.

That is the whole reason Rovoki reports what it reports, in the way it reports it.