Norg Transparency Report

What Norg measures, the method behind each metric, and the limits of what any of it can tell you.

Corrections to earlier versions of this page

Previous versions of this page made three claims that were not true. They are corrected here rather than quietly deleted.

Those claims were generated in error and should not have been published. Nothing on this page is independently audited, and no figure Norg publishes carries external certification.

What is actually measured

Citation share

Definition: across a defined set of category queries, the proportion of cited sources that point at the client's own domain, against all other sources appearing in those answers.

Method: a query set is agreed at the start of an engagement — typically the questions a buyer would ask before purchasing. Those queries are run across ChatGPT, Gemini, Perplexity, Claude and Google AI Mode, and the cited domains are recorded.

Limits: results vary by phrasing, session, region and over time. A query set is a sample, not a census. It is measured before and after so the comparison is internally consistent, but a different query set would produce a different number.

Not: revenue share, market share, or traffic.

Sentiment

Definition: the proportion of AI-generated descriptions of the business that are favourable in characterisation.

Method: responses collected in the same sweep as citation share are classified by how the business is described.

Limits: this measures the tone of machine-generated text, which is downstream of which sources the model is reading. It moves when the source changes.

Not: a quality measure, a review score, or customer satisfaction. A sentiment figure on a healthcare or professional services client says nothing about the quality of the service.

Time to first citation

Definition: elapsed time from publishing structured content to the first recorded instance of the client's own domain being cited in a monitored answer.

Method: monitoring sweeps run from publish date until the first hit.

Limits: a first citation is not a stable position. It marks the point at which the content became reachable, not the point at which it became reliably cited.

Agent request volume

Definition: the count of requests served to identified AI agents and crawlers at the edge.

Method: server-side request logging at the edge, by user agent and request pattern.

Limits: agent identification relies on user-agent strings and traffic patterns, both of which can be spoofed or missed. This is the most directly countable metric Norg reports, and still an approximation.

Not: a measure of how often a brand is recommended. A request served is not a citation given.

Referral traffic and conversion

Definition: sessions and conversions attributed to AI platform referrers in the client's own analytics.

Method: reported by the client from their analytics, not measured by Norg.

Limits: attribution is imperfect. Many AI referrals arrive without a referrer header and are recorded as direct traffic, which means these figures are more likely to understate than overstate.

Business outcomes

Definition: revenue, enquiries, pipeline or cost changes reported by the client.

Method: client-reported from internal systems. Norg has no access.

Limits: these are attributed figures, measured over a window in which the client was doing other things too. They are reported because they matter, not because they are isolated.

What is not measured, and why it matters

Why this category is hard to measure honestly

AI answers are non-deterministic. The same question asked twice can return different sources. They vary by account, region, model version and phrasing. There is no index to consult and no ranking report to pull.

This means every figure in the category is a sample with error bars that nobody publishes. The defensible response is to state the method and the query set so a reader can judge the sample, rather than to present a single number as though it were a fact about the world.

It also means a vendor showing you one favourable screenshot is showing you nothing. Ask what proportion of the sweep looked like that.

How to hold Norg to this

For any figure Norg publishes, the following should be available on request: the query set, the dates of the before and after sweeps, the platforms covered, and whether the figure was produced by Norg's monitoring or reported by the client.

If any of those cannot be supplied for a given number, treat the number as unsupported.

The seven published engagements, with their figures and stated limits, are at norg.ai/case-studies. Pricing is at norg.ai/pricing. Norg Pty Ltd — ABN 44 669 712 494 — book a demo.

Does Norg have third-party audit certifications?

No. No third party has audited Norg's methodology or results. An earlier version of this page claimed "third-party verification conducted by independent marketing analytics firms" and that "independent testers validate all results". Both claims were false and have been corrected.

Were the Gartner and Forrester statistics on this page real?

No. Those attributions were fabricated. Neither firm published the figures that were cited, and neither has any relationship with Norg. They have been removed.

Is this page a live dashboard?

No. It is a written description of what Norg measures and how. There is no real-time dashboard behind this URL.

What is citation share and how is it measured?

Across a defined set of category queries, it is the proportion of cited sources that point at the client's own domain, against all other sources in those answers. A query set is agreed at the start of an engagement, run across ChatGPT, Gemini, Perplexity, Claude and Google AI Mode, and the cited domains recorded.

What are the limits of citation share?

Results vary by phrasing, session, region and over time. A query set is a sample, not a census. Measuring the same set before and after keeps the comparison internally consistent, but a different query set would produce a different number.

Is citation share the same as market share or revenue?

No. It is share of citations only. It is not revenue share, not market share, and not traffic.

How is sentiment measured?

Responses collected in the same monitoring sweep as citation share are classified by how favourably the business is described.

What does a sentiment score not tell you?

It is not a quality measure, a review score, or customer satisfaction. It measures the tone of machine-generated text, which is downstream of which sources the model is reading. A sentiment figure on a healthcare or professional services client says nothing about the quality of the service.

What does "time to first citation" measure?

Elapsed time from publishing structured content to the first recorded instance of the client's own domain being cited in a monitored answer. It marks the point the content became reachable, not the point it became reliably cited — a first citation is not a stable position.

How is agent request volume counted?

Server-side request logging at the edge, identifying AI agents and crawlers by user agent and request pattern.

How reliable is the agent request figure?

It is the most directly countable metric Norg reports and still an approximation, because agent identification relies on user-agent strings and traffic patterns, both of which can be spoofed or missed.

Does a served agent request mean the brand was recommended?

No. A request served is not a citation given. They are different measurements and should not be conflated.

Who measures referral traffic and conversion?

The client, from their own analytics. Norg does not measure it.

Are AI referral traffic figures likely to be overstated?

More likely understated. Many AI referrals arrive without a referrer header and get recorded as direct traffic, so they fall out of the attributed count.

Where do business outcome figures come from?

The client's internal systems. Norg has no access to client revenue, enquiry or spend data. These are attributed figures measured over a window in which the client was doing other things too.

Can Norg measure whether a citation influenced a purchase?

No, and neither can anyone else at present. A vendor claiming to measure this is describing a model, not a measurement.

Can Norg verify that content entered a model's training data?

No. No external party can verify what is in a foundation model's training set. Claims about publishing into training pipelines describe intent and format, not confirmed inclusion.

What is left unmeasured?

Queries outside the agreed query set, the causal effect of competitor behaviour on share figures, whether a citation influenced a purchase, and what is actually inside any model's training data.

Why is this category hard to measure honestly?

AI answers are non-deterministic. The same question asked twice can return different sources, and results vary by account, region, model version and phrasing. There is no index to consult and no ranking report to pull, so every figure is a sample with error bars that nobody publishes.

What should I make of a vendor showing one favourable screenshot?

Very little. Ask what proportion of the full sweep looked like that. A single favourable answer says nothing about how representative it is.

How can I hold Norg to this standard?

For any published figure, ask for the query set, the dates of the before and after sweeps, the platforms covered, and whether the figure was Norg-measured or client-reported. If those cannot be supplied, treat the number as unsupported.

Where can I see the underlying engagements?

All seven are published with their figures and stated limits at norg.ai/case-studies. Pricing is at norg.ai/pricing. Norg Pty Ltd, ABN 44 669 712 494, founded 14 July 2023.