Working with us
Methodology and measurement log
How we’ll measure websites and AI visibility. So far there’s one lab measurement of our technology; the log follows.
Updated
30 minutes, free. Reply within one business day (Mon–Fri).
Four principles behind our measurements
We’ll measure without cookies or browser scripts, run every AI measurement via the APIs three times, state the method and date for every value and also publish drops. That way, you can check every number as soon as it appears here.
No cookies, no scripts
We’ll only count requests, using server logs retrieved without IP addresses. This website sets no cookies and loads nothing from third parties. Our article How to track AI traffic beyond GA4 shows how to count this way yourself.
Every value has a method and date
Every number will come with the tool, version, method and measurement date. We don’t publish estimates.
Three runs per AI question
AI answers vary, so each question runs three times via the APIs; the manual spot check runs once and is labeled as such. Only a change outside the range of those runs that lasts two months counts as a trend.
Drops stay in the record
Worse results stay in the history. We explain them once we know the cause, and we don’t delete anything after the fact.
How we’ll measure AI visibility
We’ll put a fixed list of questions, the kind buyers ask AI assistants, to several AI systems, three times each via the APIs. Then we’ll record whether a company is mentioned and linked as a source. The same method applies, at a different scale, to our own website and to client measurements for generative engine optimization (GEO).
| Aspect | Our website (GEO panel) | Client measurement (AI visibility audit, GEO retainer) |
|---|---|---|
| Questions | 40 core questions (29 German, 11 English), plus 8 rotating ones in months without a full set; every 3 months, all 90 base questions plus currently 7 supplementary questions, with metrics based on the 90 base questions | Audit: 15 of your customers’ questions; retainer: 25 questions a month |
| Question groups | brand, local, provider choice, expert questions | based on your services and audiences |
| Systems | via APIs: Claude with web search (location set to Germany) and Perplexity directly; Google AI Mode, plus ChatGPT, Gemini and Google AI Overviews through the DataForSEO service; manual spot checks: ChatGPT and Microsoft Copilot | ChatGPT search, Perplexity, Google AI Mode, Microsoft Copilot |
| Runs | 3 per question via API; the manual spot check covers the 19 brand and provider questions in a single run | 3 per question |
| Metrics | mention rate, citation rate, correct facts on brand questions, competitors named, domains cited | mention rate, citation rate, correct company facts, competitors named, sources cited |
| Spread | range across the 3 runs | range across the 3 runs |
| Frequency | monthly, on the first business day | audit once, retainer monthly |
The GEO panel is our fixed list of questions for our own website. The mention rate is the share of answers that mention the company. The citation rate is the share of answers that link to one of the company’s URLs as a source. The local questions ask about providers in Bottrop, where we’re based, and elsewhere in the Ruhr area.
We’ll report German and English questions separately. Wherever a system lets us set a location, we’ll set it to Germany, including for the English questions. The results will therefore show the answers that users in Germany get. Users in other countries may get different answers.
Why does each question run three times?
A single answer doesn’t prove much. A SparkToro study put the odds below 1 in 100 that ChatGPT or Google’s AI, asked 100 times, would give the same brand list twice. The authors consider the share of mentions across dozens to hundreds of prompts, run multiple times, a reasonable metric. A ranking position in AI answers, they say, is not. One of the authors works for an AI tracking vendor.
Source: SparkToro (external site), · vendor study · 2,961 runs of 12 prompts in 3 AI tools
Three runs are a compromise between effort and reliability. For single-question conclusions, a study using German-language prompts calls for at least seven runs per prompt per day, and eight for cited sources. So with three runs, we’ll draw conclusions about groups of questions and state the range for every number. For the audit with 15 questions, that means the values show a direction rather than a fixed figure.
Source: Schulte, Bleeker, Kaufmann (preprint) (external site), · queried from servers in Switzerland
When does a change count as a trend?
- Only when it falls outside the range of the three runs and lasts two months.
- We’ll report values before and after a known model change separately.
- We’ll assess major Google Search updates (core updates) no earlier than one week after they end.
What we’ll do with the results
- If a core question goes three months without a citation even though a matching page exists, we’ll analyze the cited sources. If a format is missing, we’ll revise the page. If the sources are directories or trade media, we’ll work on getting mentioned there.
- If a new question comes up twice in customer inquiries or search data, we’ll add it to our topic list.
To measure your own company’s AI visibility, see our article How to measure AI search visibility. The questions to ask an agency about its measurement are in our how-to on choosing a GEO agency.
Generative engine optimizationWhat this measurement isn’t
Not a ranking
The values will show how often a company appears in answers. There’s no fixed position: the same question can produce a different answer on every run.
Not exhaustive
The question list doesn’t cover everything buyers ask. We’ll choose questions based on search data and customer inquiries and document every change.
No view inside the systems
We see the answers. We don’t see the AI systems’ internal signals, and they can’t be measured from the outside.
Not personalized
We’ll ask without a personal account and with no history; if an interface requires signing in, we’ll use an empty account with no history and note it. What signed-in users see may differ: Gemini, for example, personalizes responses using past chats (external site).
What to expect from us
What we won’t promise
- A #1 ranking on Google. No one can guarantee rankings, and Google says so explicitly.
- A guaranteed mention in ChatGPT. AI answers can change quickly. Starting with our first measurement, we’ll report regularly, including any drops.
- Quick wins through volume. Pages with no value of their own, such as city pages with swapped place names, tend to do more harm than good.
- Legal advice. We build privacy and security into the technology. Legal reviews belong with lawyers.
What we do commit to
- Written acceptance criteria with a defined test method, such as Lighthouse scores and accessibility.
- A fixed price once the analysis is done.
- Monthly measurements in the GEO retainer.
- Code, accounts and documentation under your control.
No one can guarantee a #1 ranking on Google.
Is there a measurement log yet?
Not fully yet. Since Nov 9, 2026, there’s been one lab measurement of our technology, updated on Nov 14, 2026: Lighthouse on mobile, local build, 20 pages. Field data from real visits, indexing, rankings, AI crawler visits and AI visibility are still missing.
AI crawlers are the programs AI providers use to fetch web pages. Mohammad Alboush, the owner of Datafalke, is responsible for the method and the measurements.
| Metric | Budget | Method |
|---|---|---|
| JavaScript per page | 0 KB | build check, every page |
| CSS per page, gzipped | at most 25 KB | build check |
| HTML per page, gzipped | at most 30 KB | build check |
| Fonts on first load | at most 84 KB | build check |
| Page weight without images | at most 150 KB | build check |
| Third-party connections | none | check after every deploy |
| Cookies | none | check after every deploy |
Our lab measurement from Nov 14, 2026 (Lighthouse 12.6.1, mobile, local build, simulated mobile connection, median of five runs per page) found a load time (LCP) of 1.5 to 1.8 s across 20 measured pages, a lowest Lighthouse category of 99 or 100 out of 100, and no third-party connections. It doesn’t measure real visits and isn’t a commitment for client projects.
How this website is built| Measurement | Target | Frequency | Method |
|---|---|---|---|
| Lighthouse mobile: performance, accessibility, best practices, SEO | at least 95 out of 100 each | nightly | Lighthouse, mobile, median of 5 runs, the 10 most important URLs |
| Largest Contentful Paint (LCP), lab | at most 2.0 s | nightly | as above |
| LCP, INP, CLS for real visitors (p75) | at most 2.0 s, 150 ms and 0.05 | nightly, once data is available | Chrome UX Report |
| Indexed pages | all submitted | monthly | Google Search Console page indexing report; Bing Webmaster Tools |
| Google and Bing rankings | not specified | monthly | search results retrieved via DataForSEO on the measurement day, not personalized; for Google, a keyword counts at a position only if it reaches that position on both desktop and mobile; for Bing, the desktop ranking counts |
| AI crawler visits per month | not specified | monthly | server logs retrieved without IP addresses, network operator check; implausible requests are not published |
| Referrals from AI assistants per month | not specified | monthly | server logs, referring website; AI Overviews and AI Mode can’t be separated from Google Search |
| AI visibility | not specified | monthly | GEO panel, range across the 3 runs |
We publish targets only for our own technology. Visibility targets are our own assumptions and stay internal. Before we start counting from server logs, we’ll add this to our privacy policy.
Why these methods?
LCP measures how long the largest element takes to load. INP measures how quickly the page responds to input, and CLS how much content shifts during loading. Our targets for real visits are stricter than the thresholds Google gives for good values. Those are 2.5 s for LCP, 200 ms for INP and 0.1 for CLS, each at the 75th percentile of page loads. Our explainer on Core Web Vitals covers each metric in detail.
Source: web.dev (Google) (external site),
Results from Lighthouse, Google’s auditing tool, vary from run to run. According to its documentation, the median of five runs is twice as stable as a single run.
Source: Lighthouse project (Google Chrome) (external site),
Field data, meaning values from real visits, comes from the Chrome UX Report. Pages with too few visits aren’t included, and Google doesn’t disclose the exact threshold. Until a page has enough visits, the log will show “no field data yet.”
What will the history look like?
There’s no history yet. From the first measurement, this section will show one row per month with method, questions, systems, runs and values, plus a short monthly “GEO report.” Drops stay in the record. We’ll also provide the data as a CSV file under the CC BY 4.0 license.
The first monthly report will appear with the first measurement under method v1. Each report will state:
- measurement date, method and scope
- for brand questions, how many of the seven are answered correctly by at least two of the three systems, each in at least two of three runs
- for expert questions, how often datafalke.de appears as a source, with the range
- drops with a possible cause; if there are none, the report says so
- AI model changes and Google core updates that month
- the next action, with a link to the page it affects
Before this website launched, we ran a baseline measurement on October 24, 2026, with one system, mostly in a single run. It’s listed below as version v0.
Method versions
Every change to the questions, systems or runs becomes a new version with a date, the change and the reason. So far there’s only the baseline; the other two versions are planned.
- v1, planned40 core and 8 rotating questions, 3 runs, systems as listed in “Method v1 at a glance.” Plus the full set as the new basis: the 90 base questions and the supplementary questions, with metrics based on the 90 base questions.
- s1, planned Google and Bing rankings for fixed brand and local keywords plus the target keyword of every published page. Google on desktop and mobile, Bing on desktop, Google Maps from Bottrop, Gladbeck and Gelsenkirchen, each retrieved once per measurement day.
- v0, Oct 24, 2026 Baseline before this website launched: 96 answers, Claude with web search, mostly 1 run.
Built on
Editorial note
Researched and drafted with AI assistance. Mohammad Alboush has not yet checked the sources, figures and claims.
Content not yet reviewed
Spotted a mistake? Email us at kontakt@datafalke.de. We’ll fix it and log the change.