Trust
Methodology
How Norund measures AI visibility — and the limits of what any tool can honestly claim.
Which engines are queried
The demo core queries two supported engines through their official APIs: OpenAI (search-grounded Responses API with the web_search tool) and Perplexity (Sonar API). Additional engines are added through the same adapter interface. We never scrape consumer AI interfaces and never use unofficial reverse-engineered APIs.
How prompts are selected
You define the prompts your buyers actually ask — for example “Best CRM for real estate agencies”. For free public scans we generate up to five neutral starter prompts from the brand name and domain. Each active prompt is sent to each configured engine on every scan.
How brands and competitors are matched
Each raw answer is classified server-side by a structured analysis step that may only use the supplied answer text and citations. It must never invent mentions, citations, positions, or sentiment, and it records uncertainty explicitly. A cited URL alone is not treated as a recommendation. Competitors are classified with exactly the same rules as the primary brand.
Direct recommendation vs incidental mention
A direct recommendation means the engine explicitly recommends or ranks the brand as an option for the prompt. Compared means the brand is discussed against alternatives without a clear recommendation. Incidental means the brand appears in passing. Negative references and absence are recorded as such, and genuinely ambiguous cases are marked unclear and excluded from scoring.
Sentiment classification
Sentiment (positive, neutral, mixed, negative, unknown) is classified from the answer text only, with a short supporting excerpt stored as evidence.
How the Norund Score is calculated
The score is deterministic code — a language model never calculates it. Per valid result: base position value (1st direct recommendation 100, 2nd 80, 3rd 65, 4th+ 45, compared 35, incidental 20, negative reference 10, not mentioned 0) multiplied by a sentiment multiplier (positive 1.00, neutral 0.85, mixed 0.65, negative 0.35, unknown 0.70). Engine score = average of that engine's valid results. Blended score = equal-weight average of engine scores. Failed provider calls are excluded entirely — they are never scored as zero.
Answer Coverage vs Recommendation Share
We report two distinct competitive metrics. Answer Coverage is the percentage of valid answers in which an entity appeared (including direct recommendations, comparisons, and incidental mentions) — several brands can each reach 100%, so coverage does not sum to 100%. Recommendation Shareis each entity's share of weighted direct-recommendation prominence across all tracked entities, using the same position and sentiment weighting as the Norund Score; only positive, neutral, or mixed direct recommendations qualify, and it sums to 100%. Recommendation Share is Norund's share-of-voice measure. When no qualifying direct recommendations exist, we show “no recommendation share available” rather than a misleading figure.
How citations are extracted
Supported engines return source URLs alongside answers. We normalise them (strip tracking parameters, unify hosts) and aggregate the most frequently cited domains and pages. Frequently cited sources are evidence signals showing which domains appear in AI-generated answers for your tracked prompts — citation does not guarantee recommendation, and acquiring a link does not guarantee an improved score.
Why AI answers vary
AI engines are probabilistic and their underlying indexes change constantly, so the same prompt can produce different answers between runs — that variance is part of what the platform measures over time. API results may also not perfectly reproduce every consumer interface answer, because consumer apps apply their own context and personalisation. We do not know, and do not claim to know, the exact internal ranking algorithm of any provider.
What the score is — and is not
The Norund Score is a visibility signal for tracked prompts on supported engines. It is not an absolute statement of market position, and the platform cannot guarantee rankings, citations, traffic, or revenue.
Two kinds of confidence
We separate two ideas that are easy to confuse. Extraction confidence is how confidently the analysis layer interpreted a single provider answer — it is stored per result. Assessment confidence is how reliable the overall Norund Score assessment is, based on valid sample size, provider coverage, and scan completeness: low below 10 valid results or below 50% completion; medium at 10–29 valid results with at least 50% provider coverage; high at 30+ valid results with at least 80% provider coverage. A one-prompt scan is never shown as high assessment confidence, even when the extraction model was confident about that single answer.
When scans run and how failures are handled
Plans include weekly scheduled scans plus manual scans with a cooldown. Transient provider errors are retried a bounded number of times with hard timeouts; permanent errors are not. If some calls fail, the scan completes partially and the score uses only valid results. If every call fails, no score is produced for that scan.