AI Citation Index methodology

The AI Citation Index measures how often AI answer engines name a brand, and which websites they cite, when people ask them the questions buyers ask. This page explains what we ask, where we ask it, how we count, and what the numbers can and cannot tell you.

Industries we track

The index covers these industries, each with its own prompt set and brand list:

• Help desk and AI customer service software: Ticketing help desks and the new AI support agents, asked the way buyers ask.
• Personal loans: National lenders and loan marketplaces, by credit score, purpose and speed.
• Cruise lines: Ocean cruise lines, by trip type, budget and who is travelling.
• Email marketing software: Newsletter and automation tools for creators, stores and small teams.

Each industry is measured and published on its own. We do not compare mention rates across industries.

What we ask

We use a fixed set of 200 prompts per vertical, written the way real buyers phrase questions. The prompts come from the "People also ask" questions Google shows for common searches in the category, from forum and Reddit thread titles, and from keyword variants.

The set has four kinds of prompts:

• Recommendations (about 40%): "What is the best ... for ...?"
• Problems (about 25%): "How do I fix ...?" or "How can I ...?"
• Comparisons (about 20%): "X vs Y" and "alternatives to X".
• Prices (about 15%): "How much does ... cost?"

Only comparison and alternatives prompts name a brand. The set is frozen for a quarter, so week-to-week changes come from the engines and not from us changing the questions. Each set has a version number, for example 2026Q4.1.

Where we ask

We ask Gemini, Google AI Overviews and ChatGPT.

Gemini, Google AI Overviews and ChatGPT are collected with Ujeebu's LLM APIs.

Every table and chart says which engines it includes.

Prompts are written and asked in US English.

When we ask

We collect once a week, between Saturday 00:00 and Sunday 12:00 UTC, in random order and at low speed. Weeks follow the ISO calendar (for example, week 41 of 2026 is written 2026-W41). We keep every raw answer, so we can recount a past week when we improve how we read answers.

How we count

For each brand, engine and week:

• Mention rate: the share of answers that name the brand. If an engine answered 180 prompts and 45 of those answers named the brand, the mention rate is 25%.
• Range: every mention rate comes with a 95% range (a Wilson interval). A small number of answers gives a wide range. Treat two brands whose ranges overlap as roughly level.
• Average position: in answers that are ranked lists, the average place where the brand appears (1 is the top). We only show it when the brand appears in at least 5 ranked lists.
• Share of voice: the brand's mentions divided by the mentions of all brands we track in the vertical.
• Citation share: the share of all cited links that point to a website. For a brand, we count links to the websites that belong to it.
• Week-over-week change: the difference in percentage points from the previous published week for the same engine.
• Overall: the plain average across the engines published that week. We always list which engines are included.

For Google AI Overviews we also report how often an overview appeared at all, because Google does not show one for every search. A search without an overview counts as "no overview", not as a failure.

Quality checks

Before a week is published, each engine has to pass two checks:

1. Coverage: at least 90% of the planned questions must come back with a usable result. Below that, the engine shows "no data this week" with the reason, never a zero.
2. Spot check: a person reads 5 random answers per engine and checks the brands and links we extracted. More than 2 mistakes in the week's sample holds that engine back until we fix the problem and recount.

A week appears on the index only when at least 2 engines pass. We do not publish a number based on fewer than 20 answers.

Which brands we track

Each vertical has a list of tracked brands, with their names, common spellings and websites. When an answer names a brand that is not on the list, we review it and add real brands to the list, then recount. Share of voice only counts brands on the list, so a brand we have not added yet does not reduce anyone else's share.

Known limits

• AI answers change from one run to the next. One weekly run per prompt is a sample, not a census, which is why we show ranges and hold back small samples.
• We ask in one language, for one market. People asking elsewhere, or with their own history and settings, can see different answers.
• We count whether a brand is named, not whether it is described well or badly.
• A large move in one week can come from a change on the engine's side rather than from anything a brand did.

Corrections

If you think we got a brand, a website or a number wrong, tell us. We check every request, fix confirmed errors, recount the affected weeks and note the change below.

Version history

• 0.2 (October 1, 2026): names the four launch industries.
• 0.1 (September 30, 2026): first draft, for review before the first published week.