Sitera measures AI visibility in a simple, verifiable way: we run a fixed set of buyer questions — the same ones customers actually use — across five AI engines (ChatGPT, Gemini, Claude, Perplexity and Copilot; four before 19 September 2026), and count citations, position and context for each company.
Every result is verified in a browser, without personalization. The automated scanner produces an indicative frequency; the manual check in a browser, without personalization, is the only standard we publish on. The scanner gives frequency; the browser gives truth.
Our own measurement — MotorikkLab
We started by testing ourselves. Four buyer questions × four engines, without personalization, two draws per measurement.
- 12 July 2026: 11 of 16 — identical across both draws. The raw matrix behind this measurement was not archived, so it is not compared cell by cell with later measurements.
- 11–12 August 2026: 8 of 16 — cells present in both draws (10 of 16 in the first draw, 12 of 16 in the second; 6 cells moved).
The two figures cannot be read as a trend. Three instruments changed between them: Claude moved from Sonnet 4.6 to Sonnet 5, Gemini switched model during the measurement, and Perplexity was measured on an account carrying history.
On the fourth question — "My child is struggling with motor skills, where do we get help in Oslo?" — no cell held across both draws, on any engine. The same pattern we measure in the law barometers: a question phrased as a problem returns different answers than a question phrased as a purchase.
These figures relate to our own business — not an average for Norwegian B2B companies.
What we see recurring
In the searches we've run, we see some clear patterns: visibility on one AI platform doesn't mean visibility on the others, fact-based and structured content is cited more often than pure marketing copy, and structured data (Schema.org) makes it easier for the models to understand who you are. These are observations from our work, not an industry-wide study.
What should you do?
Start by measuring your own visibility. We'll do a free GEO analysis for you within 48 hours — so you know exactly where you stand.
What the research says
The Sitering method rests on three choices. All three are supported by research and large-scale analyses published in 2026. None of the studies below has been peer-reviewed yet, and the Wellows analysis was produced by a company that sells GEO tools.
1. We measure several times, not once.
AI answers change from run to run. In a study where the same questions were asked every day for nine days, the sources typically overlapped by only around 30% in Gemini and around 50% in Perplexity. Another study found that ChatGPT did not use web search in 57.8% of repeated runs. That is why we count every answer. We only publish new figures after at least two independent measurement rounds.
Sources: Sielinski, "Quantifying Uncertainty in AI Visibility", arXiv 2603.08924, 2026 · Schulte et al. 2026, as reported in "A Critical Survey of Generative Engine Optimization (2023–2026)", arXiv 2607.14035, July 2026.
2. We measure five engines, not one.
The engines draw on different sources. In an analysis of 22.7 million AI citations from January to June 2026, 79.6% of the sources appeared in only one engine. Perplexity never cited 89% of the websites ChatGPT cited for the same question.
Source: Wellows, "89% of What ChatGPT Cites, Perplexity Never Touches", August 2026.
3. We measure whether you are cited, not just whether you are recognized.
In a study of 112 startups, ChatGPT recognized 99.4% of the products when they were mentioned by name. When the question was open-ended, the same products appeared in only 3.32% of the answers. The study uses two models, but the distinction is fundamental. This is what we call «gjenkjent, ikke anbefalt» (recognized, not recommended).
Source: Sharma 2026, as reported in arXiv 2607.14035, July 2026.