How the numbers are built.

The question

When buyers in a category go looking, who do they find, in Google and in AI answers, and where are you? Everything below exists to answer that with numbers instead of impressions.

Scope of a study

A category is defined by a fixed list of buyer queries, usually 10 to 20, chosen for commercial intent in one market. The same list is used for every brand in the set, typically 3 to 7 brands: the subject, the brands it names as competitors, and the brands search data says it actually competes with.

Demand Control Index

A 0 to 100 score per brand. It is deterministic: the same inputs always produce the same score, and every input is logged with the report.

• Archetypes summarise the shape of a score, for example "Dark Horse" (mid-pack, but owns a viable wedge query) or "Specialist" (real audience in an adjacent sub-topic).
• The AI citation weight is the smallest today and the one we expect to raise as AI answers take more of the journey.

AI citation share

We ask the models what a buyer would ask, then count who they name.

• A fixed set of buyer prompts per category, the same across brands and across months.
• Answers come from a consumer-facing model with web search enabled, because that is what a buyer sees. Raw API models without search would measure the wrong thing.
• Brand extraction is a separate second step, so the model answering is never told which brand we are measuring.
• Citation share is the share of prompts whose answer names the brand. Position is the order in which brands appear in the answer.
• Google AI Overviews are recorded per query: whether one appears, which sources it cites, and what share of the category's search volume that covers.

Dead weight

A crawl of the site is joined against ranking data. Indexable pages that rank for nothing are dead weight. The ratio tells you how much of the site does work in search and how much dilutes it. In the categories we have measured it ranges from 80% to 96%.

Data sources

• Search data: Ahrefs (organic keywords, paid, SERP snapshots, backlinks, domain rating).
• Site data: a full crawl of each brand, respecting robots and rate limits.
• AI data: our own prompt runs against current consumer models, captured with the answer text and cited sources.
• Client data, when available: Search Console and analytics, used to check our estimates against reality.

What we do not do

• No synthetic numbers. Where a data pull is incomplete, the report says so, and the conclusion is softened, not filled in.
• No historical trend claims without a real time series. Trends appear only after we have measured a category more than once.
• No client names on this site without written permission. The sample report has every brand replaced.

What the patent record says

We did not invent this method in a vacuum. Much of it is described in Google's own patents, some from long before AI answers existed. A few you can look up by number:

• Answers are extracted from the opening sentences of a passage, so the answer has to come first, not last. US10783156B1.
• Facts are scored on their own: numbers, dates and names raise a page's credibility. US9672251B1 and US9785686B2.
• Detecting repeated boilerplate and weighting parts of a page differently are old patents, not news. US8898296B2, US8538989B1, US9665617B1.
• Topical authority is a history of interaction with a source on a topic, not an article count. An established source gets a head start on new content. US8949228B2, US8458196B1.
• Information gain is its own concept: what a page adds beyond what other pages already say. US20140280011A1, US11354342B2.
• The record now covers generative answers with citations, hallucination detection and factuality scoring. US12346366B2, US20250225337A1, US20240289395A1.