1 · Documentation

How we measure

What augenmerk measures, how a run works and what we do not claim.

augenmerk measures whether and how a brand appears in the answers of AI systems: ChatGPT, Claude, Gemini, Grok, Perplexity and Google's answer box. Every number in the application comes from stored answers, and every statement about a brand is backed by the exact passage in the original answer. This documentation explains how.

What we measure#

• Presence: the share of answers that name a brand, across the counting questions and all runs. See Presence and position.
• Position: the place in the ranking of a market's most-named brands, from ranking questions only.
• Share of voice: the brand's share of all brand mentions.
• Range presence: presence per product line, separate from the brand value.
• Channel contrast: model knowledge without web search against answers with live search against Google's answer box. See Channels and systems.
• Sources: which domains the AI retrieves and cites, and whether your own website is among them. See Sources and website.
• Topics and brand image: what the AI praises and criticises about a brand, with evidence and source. See Evidence.
• Purchase check: what the AI answers when someone double-checks the brand before buying — drawbacks, "not for whom", worth the price. The second half of the buying journey; most tools measure only the first. See Questions.
• Google demand: with Search Console connected, your website's real impressions, clicks and queries next to the AI citations. See The application.
• AI access: whether AI bots may read the website according to robots.txt and HTTP status.
• Findings and next steps: rules over the measurements that say what to do and how to see the effect. See Findings catalogue.

How a run works#

A run is one complete measurement of a project. It has six steps, in this order.

1. The sentence before every question. Every question gets the same preceding sentence, which sets the role the AI answers from, with the project's region and language. It influences which brands are named at all and therefore stays the same across all runs. Its wording is shown in the application under "How we measure".
2. The questions. Every active question is asked once in every channel. Which questions count towards presence and which do not is explained under Questions.
3. The channels. Ten systems, in three channel types: model knowledge, live search, answer box. Every answer is stored in full, with model, channel, timestamp, source list and, where the system reveals them, the search queries it ran.
4. The judge. A second model reads from every answer which brands are named, where, with what verdict and with what source. The judge does not form opinions; it extracts passages.
5. The evidence check. Every passage is checked character by character against the stored answer. Whatever is not found there is discarded, not softened.
6. The evaluation. From the verified mentions come presence, position, source picture, topics and findings. Each run also adds the domain check of your own website.

The raw answer is stored before it is judged and never changed. It cannot be repeated: the same question gives a different answer tomorrow. Better rules or a new competitor can therefore be applied retroactively to all stored answers without a new measurement.

Runs and movement#

A single measurement is a sample. The same question gets two different answers on two days, and a brand can swing by ten or twenty percentage points between runs without anything having changed. augenmerk therefore states a noise band for every project and reports a movement only when it exceeds it. How it is calculated is explained under Noise band and runs.

What we do not claim#

• We show that an AI said something, not that it is true. A statement without a source comes from model knowledge; it may be correct, but need not be. We show it as a statement, not as a fact.
• We know what the AI said and, for some channels, what it searched for. Whether a search led to a particular source we do not know.
• We say what to do, not whether it worked. The next run shows that. A finding counts as resolved when its number moves, not when someone ticks it off.
• Older runs were judged by older rules. Every version of the methodology applies from its filing date and is never overwritten. That keeps the history comparable, and the application names the applicable version for every project.

What Google itself says#

In May 2026 Google published a guide to optimising for generative AI features. It states what is not needed for Google's own AI features: no AI files such as llms.txt, no splitting pages into small pieces, no rewriting for AI, no special markup. What counts is the usual: indexable, useful content, technical basics, genuine rather than manufactured mentions. augenmerk recommends none of what Google declares unnecessary, and measures precisely what Google leaves open: what the other systems say.

Since June 2026 Search Console also offers a report on impressions in AI Overviews and AI Mode, per page. It is the only figure on AI visibility that comes from an AI provider itself, and it applies to Google only. Google does not yet make this report available to tools like augenmerk (as of September 2026); the other Search Console data — impressions, clicks, queries per page — we already read once you connect your property. As soon as Google releases AI impressions, they will sit next to our measurement in the gap view.

Where the data lives#

augenmerk is developed and operated in Hamburg. The data lives in Frankfurt. The AI systems themselves are queried through their providers' interfaces; the questions go there, not the customers' data.

Updated: 2026-09-16