How likely is the internet to be right about your company?
That is the question the Digital Reality Score answers. This page publishes how: the definition, the rules, the thresholds and the limits. Every Score page links back to the evidence behind its number.
What the Score means
The Digital Reality Score is the expected accuracy of what people and machines encounter about a company, weighted by how much each fact matters and how often each surface is seen.
Take every material claim about the company whose Business Reality could be established. Take every place that claim is represented: the company’s own website, registries, reference works, databases, press, search results and AI answers. For each representation, estimate the probability that it matches what is actually true. Weight it by the claim’s importance and by how much that surface is encountered. The Score is the weighted average, on a scale of 0 to 100.
Because the weights work as sampling weights, the number keeps a plain reading: a representation of the company, drawn in proportion to importance and exposure, is accurate with probability DRS / 100, across the material claims and surfaces assessed.
Earning the percentage sign
A probability is only meaningful if it is calibrated: of all the representations the model rates 80% likely to be accurate, about 80% should turn out accurate when checked independently. Until that has been shown, the Score is published as an index, 80 / 100, with this reliability interpretation. It will be published as a percentage only when calibration evidence supports it.
Calibration is measured against a growing set of claims whose truth has been verified independently. That set deliberately does not rely on claimed companies alone. Companies that claim their profile are self-selected and interested parties, so a random sample of unclaimed companies is verified by analysts every quarter. Calibration results, and the size of the validation set, will be published here.
Bands and confidence
Higher means more reliable. The Score always travels with a band, a confidence level, Evidence Coverage, the date and the methodology version.
| Band | Range |
|---|---|
| Aligned | 85–100 |
| Minor gap | 70–84 |
| Moderate gap | 50–69 |
| Significant gap | 0–49 |
Claims are the unit
A claim is an important fact about the company, written as subject, predicate and value.
Acme → primary_category → Enterprise operations platform
Acme Flow → product_available → Generally available
Claims are weighted by importance (high 3, medium 2, low 1) and come in four kinds, because not every claim can be known the same way:
- Fact
- Independently verifiable: leadership, founders, headquarters, ownership, size, certifications. Needs independent evidence.
- Self-defined
- The company is the main authority: category, positioning, products, pricing. The company’s own statement counts for more.
- Subjective
- Opinion or marketing, such as “leading provider”. Recorded, never scored.
- Diagnostic
- About being found rather than about the company, such as whether a search for the name leads to it. Shown as diagnostics, never part of the Score.
Two questions, kept apart
The Score answers two different questions, and never uses the answer to one as evidence for the other.
1. What is true?
For each claim the methodology estimates the probability that the value the company states is Business Reality.
- Start from a prior. A publicly stated fact starts at 60%; a self-defined claim at 78%. When the company states nothing, a value proposed by sources starts at 50%.
- Only independent, citable evidence moves it: registries, reference works, press, databases, directories. Each one shifts the log-odds by its authority × reading confidence × freshness, times 2 when it supports the stated value and -2 when it contradicts it.
- Copies collapse to their origin. A directory that repeats Crunchbase is Crunchbase again. One hundred copies are one vote, cast by the strongest of them.
- Search results and AI answers never count as proof. They are how a company is represented, not evidence of what is true. Five AI answers agreeing leave a claim exactly as unresolved as none.
- The company’s own properties never prove themselves. The website sets the stated value; it cannot corroborate it.
- Evidence ages. It loses half its weight every 365 days.
Business Reality counts as established when the probability rises above 80% or falls below 20%. Below 20%, the company’s own statement appears to be wrong, which is itself a finding: Stated Reality is often where an error starts. Between the two, the claim is unresolved.
2. What is shown?
Given that probability, each representation of the claim is scored on how likely it is to be accurate: a representation that matches the stated value is accurate with the same probability; one that is outdated or contradicts it is accurate only if the stated value is wrong. Each representation is weighted by exposure, how much that surface is encountered. The same wrong fact costs more in a knowledge panel than in an obscure directory.
Exposure in drs-0.2 uses published proxies by kind of surface. They will be replaced by measured exposure: search demand × rank position for search surfaces, answer frequency for AI.
| Surface | Exposure proxy |
|---|---|
| Search results, knowledge panels | 1.00 |
| Company website | 0.80 |
| AI answers | 0.70 |
| Company profiles (LinkedIn, maps) | 0.50 |
| Wikipedia, Wikidata | 0.50 |
| Press | 0.35 |
| Company databases | 0.30 |
| Directories and review sites | 0.20 |
| Registries and filings | 0.05 |
Three claim states
- Aligned
- Important sources represent the claim as the company states it.
- Gap
- Sources represent the claim differently from the company, with enough evidence to say so.
- Unresolved
- Not enough independent evidence to decide. Unknown is not treated as wrong.
A claim is aligned when Business Reality is established and its representations are at least 75% likely to be accurate. Below that it is a gap, and a gap always names its kind, using the Digital Reality Gap framework’s claim states:
- Incorrect
- Wrong, at the source or somewhere downstream.
- Stale
- True once, and still circulating after it changed.
- Conflicting
- Different sources give different answers.
- Unsupported
- Confirmed by the company, with no independent evidence behind it.
- Missing
- True, and absent where people look for it.
- Misunderstood
- Present, and interpreted as something else by search or AI.
- Divergent
- Right in one market or language, wrong in another.
“Incorrect” is reserved for a stated value that the evidence says is wrong, a company correction, or a registry-grade source (authority 0.9+) contradicting it. Disagreement alone is “Conflicting”.
Evidence Coverage
The Score measures how reliably the company is represented. Evidence Coverage says how much of the company’s material reality could be assessed confidently.
Coverage is the importance-weighted share of fact and self-defined claims whose Business Reality could be established. There is one Score, not a public one and a verified one. Unresolved claims are left out of the Score and lower Coverage instead. Unknown is never treated as wrong, and a Score of 95 on 30% coverage says something very different from 95 on 97%.
Diagnostics
Six views explain where a Score comes from. From drs-0.2 they are not averaged into it.
| Diagnostic | Question |
|---|---|
| Accuracy Probability that a source, any source, represents each claim correctly, with every source counted equally. The Score weights the same thing by exposure. | Are important company facts represented correctly? |
| Consistency Whether sources tell the same story as each other. Diagnostic only: sources can agree on something wrong, so agreement never raises the Score. | Do important sources agree? |
| Evidence How many independent origins corroborate each claim. Copies of the same source count once. Two independent origins count as sufficient. | Are important company claims sufficiently supported? |
| Authority The standing of the strongest independent source supporting each claim. Registries and reference works rank above directories. | Are credible and authoritative sources reinforcing those claims? |
| Discoverability Whether searching for the company leads to the company rather than to someone else, weighted by exposure. | Can the right company, products and information actually be found? |
| Representation Probability that search results, knowledge panels and AI answers represent each claim correctly, weighted by exposure. | Do search and AI systems describe the company correctly? |
Sources and provenance
Every observation keeps its provenance, using the framework’s source states:
- Declared
- Supplied and confirmed by a verified person at the company.
- Observed
- Found on properties the company controls.
- Evidenced
- Found in an independent, citable source.
- Perceived
- How search results and AI systems represent it.
- Unknown
- Not enough evidence either way.
Authority and exposure are different properties. Authority is how much a source counts as evidence of truth; exposure is how often it is encountered. A registry has high authority and little exposure; an AI answer has exposure and, for truth, none.
| Source | Counts as evidence | Authority | Exposure |
|---|---|---|---|
| Companies register, regulatory filing | Yes | 0.95 | 0.05 |
| Wikipedia | Yes | 0.85 | 0.55 |
| Business press, analyst reports | Yes | 0.80 | 0.40 |
| Industry publications | Yes | 0.75 | 0.30 |
| Company databases | Yes | 0.70 | 0.30 |
| Review sites | Yes | 0.55 | 0.35 |
| Directories | Yes | 0.50 | 0.15 |
| Company website | No | — | 0.80 |
| Search results | No | — | 1.00 |
| Search knowledge panel | No | — | 0.90 |
| AI assistant answers | No | — | 0.70 |
When the company answers
A verified person at a claimed company can confirm, correct or ask to clarify a claim. Every answer is evidence, labelled Declared. None of them sets the Score. An answer shifts the log-odds by 2 while pending and 3.5 once an analyst has reviewed it, and it can never make a claim certain.
Answers mostly raise Evidence Coverage. They can lower the Score: confirming a fact the wider web gets wrong turns an unresolved claim into a gap. A company cannot choose or buy a better Score. It can only help establish what is true. Company-supplied answers are private unless the company publishes them.
Badges
A claimed company can display its Score when Evidence Coverage is at least 80%, the Score is no older than 120 days, confidence is not low and no high-importance claim is unresolved. A badge means the company’s assessed digital representation is highly likely to be accurate and current. It does not say the company is trustworthy, ethical or good at what it does. The Score measures information, not character.
Limitations
- An automated Score cannot know every fact about a business. It knows what can be observed and established. That is why Coverage is published and unknown stays unresolved.
- Priors, evidence strengths, authority and exposure are published starting estimates. Calibration will adjust them, and each change is a new methodology version.
- Dependency between sources is recorded where it is known. Undetected copying still overstates corroboration.
- AI answers vary by model, date, language and phrasing. Each run records which assistants were tested and when.
- drs-0.2 scores in English. A company can be represented accurately in one language and badly in another.
- A Score is a snapshot at the date shown on it.
Versions
| Version | Change |
|---|---|
| drs-0.2 | The Score becomes an exposure-weighted expected accuracy of representations. Truth is estimated from independent evidence only, with copies collapsed and freshness decay. Company answers become weighted evidence. Consistency and the other five dimensions become diagnostics, outside the Score. Findability claims move to diagnostics. Badges gain eligibility rules. |
| drs-0.1 | Weighted mean of six dimensions, including source agreement. |
Published Scores keep the version that produced them. Scores from different versions are not directly comparable.
How often Scores are updated
Published Scores are re-run periodically, and every run is kept in the company’s history. Continuous monitoring of search, mentions and AI representation is not part of the Score; that is what PageRadar SIGNAL does.
Independence, and the framework
The Score is maintained by PageRadar, which also sells services that help companies improve how they are represented. How that conflict is handled is set out in the independence policy. The Score is built on the Digital Reality Gap framework: the framework explains why the realities diverge, the Score measures it.
Preview status
This is methodology drs-0.2. The companies published today are fictional demo profiles, run through this exact methodology from invented evidence. Scores of real companies will be published once each has been reviewed by a person.