Start with numbers. Continue with a concrete plan.
We document baseline measurements, investigate available sources and separate confirmed findings from suspected causes. You decide who implements the corrections.
01
Measure the baseline
Assess each agreed question 100 times per system. Save answers and conditions; report confirmed errors and missing information separately.
02
Investigate sources and causes
Examine cited and other relevant pages. Separate observations from suspected causes and prioritise specific actions.
03
Implement agreed corrections
Your team implements the plan or commissions GEO Agentur Wien separately. We edit approved sources within the agreed scope.
04
Measure again and compare
Repeat the agreed questions. Document source changes, changed answers and remaining errors separately.
Which questions we assess.
- Brand & identity: Does AI identify the correct business or person? Does it mix information about similarly named businesses or people?
- Products & professional work: Are services, products, specialisations and experience described accurately?
- Current information: Are locations, contacts, offers and conditions up to date?
- Competition & comparisons: Which verifiable facts appear when AI compares you with alternatives?
- Niche & situation: How does it describe you in answers about a specific task, service or region?
We select questions for your situation. These groups are assessment scenarios, not fixed industry packages.
100 repetitions per question and system.
We assess ChatGPT, Claude, Gemini, Perplexity and Grok. Each agreed question is asked 100 times in each system. This measures how often answers to that specific question contain errors under the defined conditions.
Example: five agreed questions produce up to 2,500 answers per measurement, or up to 5,000 in total for the same five questions before and after the work. The number of questions, languages, models and search modes is agreed separately. We report the actual volume collected.
100 repetitions show how consistently a specific result occurs. Different wording requires additional questions. This sample does not establish an overall error probability for all users.
What we measure and report separately.
- The share of checked answers containing facts about you that include at least one confirmed factual error.
- How often each individual confirmed error repeats.
- Unsupported statements requiring further investigation.
- Answers without information about you, refusals, empty answers and technical request failures, each reported separately.
The error rate is the number of answers containing at least one confirmed factual error divided by the checked answers containing factual information about you. Unsupported statements are not automatically errors. Missing information or mentions do not count as a correction. Results are reported per system and question with the respective sample counts.
Technical foundations and limits
Google explains that search foundations also apply to its generative features. No special AI file or markup is required. This does not promise inclusion, and the documentation describes Google rather than every AI system.
Google Search Central: AI features and your website
Bing provides citation data for supported experiences in its public AI Performance preview. These data must be reported with their actual coverage.
From answers to specific actions.
We examine cited sources and other relevant pages: your website, managed profiles, directories, industry platforms and publications. Different page-capture and extraction tools help us identify outdated information, contradictions, mixed identities and technical access barriers.
Receive specific URLs with the discrepancy, evidence for the correct information, an action, priority, responsibility and verification criteria. Confirmed observations and suspected causes remain separate.