What RivalEye measures – and what it does not.
Trust comes from traceability. That is why we set out openly here how reports are created.
1. Data sources
- Public websites of your company and your competitors (homepage and offer pages). We respect robots.txt, read only a few pages per run and do not copy texts into reports – only a summary and a link.
- Public directories and search results for business data (name, address, phone, opening hours, categories).
- Publicly visible review signals (number, average, time). We use official interfaces or publicly indexable information and never circumvent platform rules.
- AI systems: fixed question sets are put to supported AI assistants; the answers are logged.
- Data provided by you (lines of business, area, known competitors, goals). Optionally later: Google Search Console via OAuth.
Every observation stores the source, time, raw value, normalized value and confidence.
2. Change detection
Whether a page has changed is decided by an exact text comparison between two points in time – not by the AI. Meaningless changes (timestamps, tracking IDs, cookie notices) are filtered out. The AI then only assesses what a detected change means for you.
3. The AI Council: a team of AI models
Three AI models from different providers work as one team here. The author model researches and writes the draft. The reviewer model receives the same evidence and checks every statement: agreement, objection, what is missing. The draft is revised. If disagreement remains, an arbiter model decides as an independent instance – or the recommendation is issued as “split” with both positions. The providers are interchangeable; which model was involved is stored for each report. With every recommendation you see agreed or split – with both positions. All models involved know the latest reports so that nothing is recommended twice. Which providers are in use in September 2026 is stated in the privacy policy (section 7) and on the Security page; if the lineup changes, it will be stated here and in the report. There is no partnership with the providers.
4. AI visibility
AI answers vary by time, model, location, wording and account. That is why:
- We measure with fixed question sets (10, 20 or 30 questions depending on the subscription) that match your lines of business and your area.
- Every measurement states the date, sample and systems: “named in 4 of 10 questions” – never “63% visibility worldwide”.
- We log the companies named, their position and the sources cited. The trend over several measurements is meaningful, a single measurement is not.
- We do not promise that your company will appear “on every phone”. Nobody can guarantee that.
5. Prioritization
Every recommendation receives estimates for impact, urgency, confidence and effort. These produce the order (Critical · High impact · Opportunity · Watch · Low). These estimates are a decision aid, not an exact science – and they are labeled as estimates.
6. Fact, assumption, recommendation
Reports separate three levels: fact (direct observation with evidence), assumption (probable explanation, marked as such), recommendation (suggested action). If evidence is missing, RivalEye says “I don’t know” or “RivalEye assumes” – never “RivalEye knows”.
7. What RivalEye does not measure
- The actual quality of work, customer satisfaction or reliability of a company – only the publicly visible signals of these.
- Private business figures of competitors (revenue, capacity utilization, margin).
- Content that is not publicly accessible. We do not bypass logins, blocks or technical protection measures.
- Guarantees: RivalEye detects, watches, analyzes and recommends. Implementation and results are up to you.
8. Fairness toward competitors
RivalEye observes, it does not sabotage. No fake reviews, no spam, no attacks, no copying of other people’s content. Recommendations on prices are worded neutrally (“Competitor offers X – check whether you want to respond”) so as not to encourage price fixing. How to recognize our requests and how to block them via robots.txt is described on How to recognize RivalEye requests.
9. AI labeling
Reports are created with AI support and labeled accordingly (EU AI Act, Art. 50). For every AI-generated assessment we store the model, prompt version, evidence and time so that it remains traceable and reproducible.
10. Corrections
You can mark any observation as “correct”, “wrong” or “no longer relevant”. We correct errors; the history is not silently overwritten in the process. This methodology is versioned: material changes are listed on this page with a date.
11. Free Competitor Check
The free Competitor Check is a one-time, automatic snapshot – separate from the subscription. It checks your website and up to three competitor websites with the same 33 checks in five categories (AI readiness, Google basics, technology & speed, trust & contact, customer acquisition) and awards a maximum of 100 points. The scoring is rule-based: fixed rules, no AI – points, measured values, ranking and effort always come from these rules. The load time is a single measurement from our server. If secondary files such as llms.txt or the sitemap do not arrive in time, the affected checks show “not determined” – without any deduction of points.
AI Council in the check: After your confirmation, the AI Council writes the texts of the PDF report: summary, assessments per category, notes on the competitors and your most important actions with instructions. The author model drafts – in the check a smaller, faster variant than in the subscription reports –, the reviewer model checks every statement against the measured data, and if there are objections the arbiter model decides. The AI Council receives only the measured data of the check (no email or IP addresses) and does not call up any websites itself. Our program checks its answers before they go into the PDF: only checks that are open for you, only statements with evidence from the measured data, no legal judgments; anything unusable is discarded and replaced by the rule-based text. Every action states whether the Council agreed, whether an objection was checked or whether the arbiter model decided. If the daily quota has been reached or a service does not respond, you receive the rule-based report – the PDF then says so. Reports with AI texts are labeled in accordance with the EU AI Act (Art. 50).
The rules sort the actions by possible points, effort and how many of your competitors already meet the check. The AI Council may change this order if it gives reasons, but only among the checks that are open for you; possible points and effort remain rule-based. Not visible to the check are rankings, visitor numbers and content that is only loaded by JavaScript. All checks with their points: methodology of the check. For each website, the check reads robots.txt, the homepage, llms.txt and the sitemap at most once (for a sitemap index, up to 2 sub-sitemaps of the same domain) plus a redirect test from http to https, respects robots.txt and reuses results for further checks for up to 6 hours: how to recognize the requests.
12. Change log
- September 2026 – Version 1.0 published.
- 26 September 2026 – Website: a clearer separation of what is implemented today and what is planned; FAQ “Can’t I simply ask an AI assistant?” added; note in the demo chat that it shows prepared answers. The methodology itself is unchanged.
- 26 September 2026 – Version 1.1: section 11 “Free Competitor Check” added (33 checks, 100 points, rule-based without AI); the identifiers of our requests are now listed on bot.html (linked in section 8). The methodology of the subscription reports is unchanged.
- 27 September 2026 – Version 1.2: Competitor Check: the texts in the PDF are now written and checked by the AI Council, points and ranking remain rule-based; at most 2 instead of 5 sub-sitemaps, fixed time limits per website (“not determined” instead of a deduction of points when secondary files respond too slowly), results per website reused for up to 6 hours. The methodology of the subscription reports is unchanged.