E016: The Noise Floor Experiment — What AI Citation Rate Looks Like With No Intervention

Why This Experiment Exists

Here is an embarrassing thing about running GEO experiments without a noise floor: you cannot tell whether your results mean anything.

I ran Experiment 001 — declarative vs narrative structure, 24pp citation rate gap — and it looked like a meaningful result. The FAQ retrieval experiment found a −1.7% delta and called it null. The E014 Month 1 baseline measured a combined citation rate of 6.7% across Perplexity, ChatGPT, and Google AIO.

The problem: none of those results had a variability reference. A +4pp delta from an intervention looks like signal — unless the platform naturally swings ±6pp on the same query set on consecutive days. Then it’s noise. Without knowing the natural variability range, every result floats in interpretive ambiguity.

E016 was designed to resolve that ambiguity. The hypothesis was simple: the GEO Lab’s citation rate has a stable noise floor that can be characterised within five consecutive daily measurement cycles. The method was even simpler — run the same queries, same platforms, same protocol, for five days, and change nothing on the site while doing it.

Methodology

The design is the simplest possible: no intervention, no changes, no new content. Just the 30-check protocol running on repeat.

Total checks: 150 (10 queries × 3 platforms × 5 days). The publishing freeze was the critical control — any content change during the window would have introduced a freshness signal that could contaminate the baseline.

Results

150 checks total (10 queries × 3 platforms × 5 days). Citation = named source link in AI response. Zero mentions (M) recorded on any platform on any day.

The most striking result is not the combined rate. It is the zero variance on Perplexity — 40% every single day, the same four pages cited every time, for five consecutive days. That kind of stability in an AI retrieval system is not what most people would expect, and it changes how Perplexity results should be interpreted going forward.

The second most striking result is the zero mentions across 150 checks. Not a single M result on any platform on any day. The outcome distribution is binary: citation or nothing. There is no “mentioned without link” middle state on this query set and this site.

What Each Platform Did

The Tier 1 vs Tier 2 Split

The most important finding in E016 is not the noise floor range. It is the absolute split between Tier 1 proprietary queries and Tier 2 category queries on Perplexity.

This is not a gradient. It is an absolute separation. The E014 Month 1 baseline suggested this split; E016 confirms it across five days of consistent measurement. Tier 1 at 80%, Tier 2 at 0% — every day, without exception.

The practical implication is significant. Most GEO frameworks advise optimising content structure — answer-first headings, declarative openings, context-independent sections. Those interventions improve passage-level extractability. But extractability is not the binding constraint on Tier 2 citation for a domain at thegeolab.net’s current authority level. Domain authority is the binding constraint. Perplexity retrieves thegeolab.net on Tier 1 queries because no other source owns those entity associations. On Tier 2 queries, it retrieves established sources regardless of how well-structured thegeolab.net’s content is.

The Noise Floor: What It Means for Everything That Follows

The per-platform interpretability thresholds are different from the combined threshold, and the difference matters for experiment design.

Perplexity is the decisive measurement platform going forward. Zero variance across five days means any movement — even a single additional citation — is interpretable signal. An experiment that increases Perplexity citation rate from 40% to 50% is a real finding. An experiment that produces no Perplexity movement but a 15pp ChatGPT swing is noise.

ChatGPT’s 20pp variability range is the most practically limiting finding in E016. It means ChatGPT cannot be used as a primary measurement platform at this sample size. Ten queries is not enough to distinguish a genuine 15pp improvement from a platform fluctuation of the same magnitude. ChatGPT results from individual experiments should be reported but not interpreted in isolation until the variability range is better characterised at larger sample sizes.

Google AIO at zero is a separate research thread entirely. The absence is structural — it requires a crawl parity audit and a domain authority investigation, not a content optimisation response. It cannot be addressed by the same interventions that affect Perplexity and ChatGPT citation rate.

The six-month System Memory hypothesis — that consistent publishing on a focused topic cluster accumulates citation rate over time — now has its interpretability reference. Month 2, Month 3, and Month 6 baselines from E014 will be evaluated against the E016 noise floor. A Month 6 combined rate above 22.0% would be interpretable as System Memory accumulation. A rate within 13.3%–20.0% would not be.

What Comes Next

E016 completes the prerequisites. The experiment queue is now unblocked.

E003 runs next — heading format as the variable, question-style H2s versus declarative H2s, Perplexity as the primary measurement platform. The noise floor makes the design clean: if declarative headings produce a Perplexity citation rate above 40%, that is signal. If they do not, the null result is credible rather than ambiguous.

E006 follows — context management and passage-level clarity. This is the experiment most directly connected to the LLM readability framework: does removing unresolved pronouns and “as discussed above” references from section openings produce measurable citation rate improvement on the same query set?

The Google AIO absence opens a separate investigation that sits outside the main experiment queue. The question is whether the absence is a domain authority threshold problem or a content-type problem. That investigation starts with a crawl parity audit: confirm GoogleBot and AI crawlers are both accessing the same pages, confirm AIO is processing the correct versions, and identify whether any specific pages are closer to AIO citation eligibility than others.

EXPERIMENT_E016:
  experiment: "30-check protocol reliability / noise floor"
  hypothesis: "Citation rate has a stable noise floor characterisable in 5 days"
  result: "Confirmed — Perplexity 40% ±0pp, ChatGPT 10% ±10pp, AIO 0%"
  noise_floor_combined: "13.3%–20.0% (6.7pp range)"
  interpretability_threshold: "22.0% combined"
  tier_1_perplexity: "80% (4/5 queries cited every day)"
  tier_2_perplexity: "0% (0/5 queries cited on any day)"
  mentions: "0 across 150 checks — citation or nothing"
  platforms: "Perplexity sonar-pro, ChatGPT gpt-4o-search-preview, Google AIO DataForSEO"
  sample: "150 checks (10 queries × 3 platforms × 5 days)"
  publishing_freeze: "13–17 April 2026 — no content changes during window"
  next_experiment: "E003 — heading format (question vs declarative H2s)"
  date: "2026-04-18"

What Practitioners Say

Frequently Asked Questions