How Rankli measures whether AI recommends you

In short#

A round asks each of your questions of each AI assistant your plan tracks, and records what every answer says about you. Rounds run every week. 7 and 14 days after an article goes live, its question is asked again, 3 times on each assistant, and the article shows before and after.

Updated 29 September 2026

What a weekly round is#

Each question is asked of each assistant your plan tracks. For every answer, Rankli records whether it named you, whether it recommended you, who else it named, which pages it cited, and whether it answered from the web or from the AI’s memory.

The same question can get a different answer from one ask to the next, so Rankli reports how often across all your questions and weeks, never a single answer.

During the free trial, one round runs on every plan: up to 50 questions on ChatGPT, Claude and Gemini, asked once. Weekly rounds at the plan’s size start with the first paid month.

Where the questions come from#

Rankli drafts them from your site, your products and your competitors, phrased the way a customer asks an assistant: the best option for a need, one product against another, alternatives to a name they know, and whether a product is any good.

You check the list during setup: untick, edit or add. The questions you save are asked word for word, every week, and Rankli never rewrites them.

Asking again after an article goes live#

Every article Rankli writes answers one of your questions. 7 and 14 days after the article first goes live, Rankli asks that question again, 3 times on each assistant your plan tracks. That is more answers than a weekly round gives, so a small change has room to show.

The re-asks run beside the weekly round and never replace it.

Before and after, for each article#

The article splits its question’s answers at the moment it first went live. Before is the last round that asked the question before then. After is every answer since: the re-asks at day 7 and 14 and the weekly rounds.

For each side you see how often AI named you, recommended you and cited the article’s own page. A change counts only when the margins of error before and after do not overlap. Anything less reads as no clear change, and too few answers on either side reads as too few to tell.

Comparing with questions left alone (Scale and Agency)#

On Scale and Agency, Rankli shows whether a change came from the work or from the assistants. Before the first round, your questions are split at random by topic: half are worked on and half are left alone.

Before the work starts, Rankli writes down what it predicts, which questions are worked on, and when the result is due. None of that can be changed afterwards. The result is the change in the questions worked on minus the change in the ones left alone, with a margin of error. When the margin includes zero, Rankli prints inconclusive rather than a number.

How the method is tested#

Rankli’s code carries its own check on that comparison: a calibration harness, run with npm run calibrate. It builds worlds where the true effect of a change is known exactly, by generating each follow-up window twice from the same random seeds, once with the change and once without. Then it scores two methods against that truth: the comparison with questions left alone, and a plain before and after on the worked questions.

The harness states its own limit. It shows that if AI answers behave the way its worlds are built (grouped by topic, drifting from week to week, and partly from the AI’s memory), the comparison recovers the true effect and the plain before and after does not. It does not show that real answers behave that way.

Getting started walks through the steps that lead to your first round.