NuLogic Independent measurement for the answer layer

Founding notice

An independent register of what AI systems actually recommend

NuLogic records which products and providers large language models name in response to real buying questions, publishes the resulting standings, and documents the method in full. Position on the register is not for sale, at any price, to anyone.

Why a register, and why now

Buying decisions increasingly begin inside an answer rather than a results page. A model is asked what to use, it names a handful of options, and most of the candidate set is never seen. Which names appear is now a commercial fact about a market.

That fact is measurable. It is not currently measured by anyone whose figures a buyer can inspect, and the instruments the software industry inherited from search were built to count a different thing on a page that fewer people reach.

NuLogic exists to record it, on a stated method, on a stated sample, with the date attached.

Finding F-001

Crawl volume does not predict arrivals

The industry counts how often AI systems fetch and cite a site. Measured on one engine at both ends, that count carries almost no information about how many people the engine then sends.

rank correlation
-0.08

Spearman rho between how much a site is crawled and how many visitors it receives. Zero would mean crawl volume tells you nothing. This is fractionally worse than nothing.

efficiency spread
385x

Between the best and worst non-zero property, in visitors delivered per thousand crawler fetches. Same engine, same funnel, same operator.

properties
14

Independent sites, one operator, both ends of the funnel instrumented. Small enough to state as a sample and not as a law.

Figure F-001

Visitors delivered per one thousand crawler fetches, by property. The properties at the foot of this chart are crawled as heavily as the ones at the head.

A · 624/day 115.6
B · 244.3/day 62.9
C · 133.5/day 46
D · 198.9/day 42.6
E† · 96.2/day 34.9
F · 256.3/day 10.2
G · 177.2/day 5.8
H · 139.2/day 1.8
I · 355.9/day 1.7
J · 135.2/day 1.1
K · 137.4/day 1
L · 322.9/day 0.4
M · 268.5/day 0.3
N† · 351/day 0
Instrument. same-engine-efficiency.mjs, run 2026-09-21.
In. Cloudflare AI Crawl Control, OpenAI crawler fetches per day. 45 daily snapshots, 2026-07-30 to 2026-09-20.
Out. GA4 Data API, ChatGPT referral sessions per day, sessionSource. 28d trailing, pulled 2026-09-21.
Corroboration. Bing Webmaster Tools impressions versus GA4 AI-referred session share, Lattice Measurement Atlas 02, pulled 2026-09-21. Reached the same conclusion from a different instrument pair with no shared inputs: search reach and AI arrival are close to independent.
Fewer than five hundred sessions from all channels over the window, so the ratio on that row is noisy and should not carry an argument by itself.
  • The two windows differ. Crawl is a 45 day recent sample and sessions are a 28 day trailing average, so the ratio is directional rather than a same window measurement.
  • The session side is measured by a JavaScript analytics client, so visitors using ad blockers or no-JS clients are not counted. Every session figure here is a floor, not a total.
  • A zero is a measured zero for a referral source string. It is not proof that nobody arrived.
  • Properties differ enormously in size, and a small property's ratio is noisy. The lowest row here draws only 71 sessions from all channels combined over the same window, so its zero is weak evidence. The strongest evidence in this chart is the large, healthy property that draws thousands of sessions and still converts almost no crawler attention into arrivals.
The most dramatic row here is not the strongest evidence, and it is worth saying so before someone else does. The row at the foot draws almost no traffic from any channel, so its zero is close to meaningless. The load-bearing case is the large, healthy property in the lower half: thousands of arrivals a month from other channels, crawled several hundred times a day, and almost none of those arrivals attributable to the crawler. That one cannot be explained away by sample size.

This is a single operator's network and is published as a sample, not as a finding about the whole web. A second instrument pair reached the same conclusion independently on the same day, which is stated in the source rail above.

What NuLogic publishes

01

Recorded standings

For a defined category, which providers are named by how many engines in a recorded sample, stated as a count and never as a superlative. A standing is a description of what was recorded, not a verdict on quality.

02

The method behind every figure

The engine set, the sample depth, the window, the query construction and the refusal conditions. Published before the figures, so the figures can be argued with.

03

Findings, dated and bounded

What the record shows about how these systems select sources, including the parts that argue against our own position. A finding without its caveat is advertising.

What is, and is not, for sale

An index is worth reading only if the reader knows exactly which parts of it money can reach. This is the whole of it.

Earned, never purchasable

  • Inclusion in a recorded standing
  • Position within that standing
  • The engine count attached to a provider
  • Whether a provider qualifies for a badge
  • The wording of any recorded figure

Commercially available

  • A licence to reprint a standing you already hold
  • A subscription to the underlying recorded data
  • Extended historical access to your own record
The rule in one line. A provider may buy the right to quote a position. No provider can buy the position. If those two ever blur, the register is worthless and so is the business built on it.

Status

This is a founding notice rather than a launch. The method and the commitments are published ahead of the register itself, so that both can be held against us afterwards.

Method and honesty floorPublished
Ownership disclosurePublished
FindingsF-001 published
First recorded indexIn preparation
Reprint licensingNot yet open
Every figure on this site carries the instrument that produced it, the windows it was measured over, and what it cannot tell you. Where a figure does not yet exist, nothing stands in for it.