The Noma Intel Way

Our measurement methodology · July 2026

Digital analytics has a credibility problem. Ad platforms grade their own homework. Third-party tracking is blocked by every major browser. And AI assistants and crawlers now read websites in volumes conventional analytics cannot see at all — or worse, silently counts as people.

Noma Intel is built on a different premise: measure first-hand, count honestly, and separate what we know from what we suspect. This page explains how every number the platform reports is produced — and why some of our figures deliberately read lower than the numbers other tools show.

1. First-party data, measured on your own site

Website numbers — visitors, sessions, conversions, channels, AI visibility — are captured by the Noma Intel tag running on the client's own domain. Browsers treat this as first-party traffic, the same trust level as the website itself, so nothing is lost to the third-party tracking blocks built into Safari, Firefox, and Chrome.

The consequence: our figures are counts, not estimates. Nothing is sampled, nothing is modelled, and no statistical guesswork fills gaps. Where something genuinely cannot be observed, the report says so rather than inventing a number.

Channels are classified once, at the moment of capture

Every visit's acquisition channel — Paid Search, Paid Social, Organic Search, AI Referral, Direct — is decided once, while the evidence still exists: the ad-click identifier in the landing URL, the referring page, the campaign tags. The verdict is stored permanently and reports never re-derive it, so history never re-shuffles when a report is re-run.

Honest "Direct", honest "unknown"

"Direct" is the junk drawer of most analytics tools. In Noma Intel, bots never hide there — non-human traffic is excluded from channel reporting and shown in its own bucket — and unknowns are labelled unknown: a visitor whose journey began before first-touch collection started shows as "Not yet known", never silently counted as Direct. Structural limitations (such as Safari capping attribution cookies at about seven days) are disclosed in the product, not modelled away.

Privacy by design, enforced in code

IP addresses are truncated — not hashed — at the moment of capture, so no stored value can single out one connection. Geography stops at city level. URLs are scrubbed to a whitelist of campaign parameters. Data older than 24 months is automatically de-identified: trends survive, individual visitor threads are cut. These rules run inside the collection pipeline and are guarded by automated tests on every release. See the Privacy Policy for the full statement.

2. AI measurement: tiered confidence, never blended

AI systems interact with websites three ways, and each demands different measurement: humans arriving from AI tools (AI referrals), automated agents driving real browsers, and crawlers that read pages without ever running the site's code — the audience that is invisible to every conventional analytics tool.

Every AI classification in Noma Intel carries one of three confidence levels, and they are never added together:

Behind the scenes, every collected event is classified twice — once in the browser, once at the server — by independent detectors, and their agreement rate is recorded. Detection quality is measured, not assumed. When detection improves, it improves forward from a dated point with a marker on the charts: a step-up means better recognition, not a traffic spike.

Seeing the invisible readers

AI crawlers never run an analytics tag, so Noma Intel observes them at the network edge of the client's own site. Reporting honesty follows the instrument: where a sensor has complete coverage, counts are exact; where a sensor sits behind page caching, its volumes are reported as "at least N" — a floor, never a precise total. Each instrumented site also invites well-behaved agents to identify themselves through a standard AI doorway, and those interactions are recorded in their own layer — never blended into human sessions.

3. Paid media: conservative counting, one yardstick

Ad-platform figures — spend, clicks, impressions, and the platforms' own conversion counts — are synced daily from Google Ads and Meta, and are always labelled as the platform's own measurements. They are shown alongside first-party outcomes, never mixed with them.

One correction is applied at the counting level. Meta's default attribution credits an ad that was merely on screen for one second — no click — if the viewer later converts. Google counts on clicks. Noma Intel therefore counts Meta conversions on click-through only (7-day click window): our Meta figures read lower than Meta's own Ads Manager, deliberately, and Meta and Google become genuinely comparable.

Cross-channel comparison then uses a single yardstick: "cost per outcome" always means ad spend ÷ the client's own on-site outcomes — human visitors only, attributed by the ad-click identifier actually observed on the visit — never a platform's self-reported conversions.

4. Change with a date on it

Measurement methodologies improve, and improvement changes numbers. Our rule: methodology changes are forward-dated, marked on charts, and explained — never retroactively smeared across history. Where a change is applied to past data (as with the Meta counting restatement), it covers the whole reported period under one rule, so no chart ever mixes two counting regimes.

The Noma Intel Way in ten rules

1. Measure first-hand. Website numbers come from first-party capture on the client's own domain — not sampled, not modelled, not bought.
2. Classify once, at capture. Channels are decided while the evidence exists, then frozen. History never re-shuffles.
3. Humans only in human metrics. Bots and crawlers never inflate sessions, conversions, or conversion rates.
4. Never mix confidence tiers. Verified, named, and suspected AI are reported separately. Suspicion is never added to a headline.
5. Audit your own detectors. Two independent AI classifiers run on every event and their disagreement is measured, not assumed away.
6. Separate AI visibility from human traffic. AI discovery, declarations, and crawls live in their own layer.
7. One yardstick across channels. Cost-per-outcome divides spend by first-party human outcomes — never by a platform's self-graded conversions.
8. Count conservatively, disclose the difference. Meta view-through conversions are excluded; the gap versus Ads Manager is explained where the number appears.
9. Say "at least N" when that's what you know. Floors are labelled as floors. Unknowns are labelled unknown, not defaulted to Direct.
10. Privacy is enforced in code. IP truncation, city-level geography, URL scrubbing, and 24-month de-identification run in the pipeline, guarded by tests on every release.

Questions about any of these rules? Contact us — the methodology is the product, and we're happy to explain any number in it.