How does ABKE (AB客) run a “de-AI-ification” content test to compare human expert tone vs pure AI copy—using reading time and scroll depth as proof?

发布时间：2026/03/18

类型：Frequently Asked Questions about Products

ABKE’s GEO “de-AI-ification” test uses an A/B experiment: (A) human expert tone with verifiable fact slices (process/standard/data source) vs (B) pure AI copy without an evidence chain. We measure GA4/Matomo Avg. engagement time (seconds) and Scroll depth (%) under controlled variables (same page layout, same traffic channel, same publish window), and report the median engagement time for each group plus the delta in seconds.

问：How does ABKE (AB客) run a “de-AI-ification” content test to compare human expert tone vs pure AI copy—using reading time and scroll depth as proof?答：ABKE’s GEO “de-AI-ification” test uses an A/B experiment: (A) human expert tone with verifiable fact slices (process/standard/data source) vs (B) pure AI copy without an evidence chain. We measure GA4/Matomo Avg. engagement time (seconds) and Scroll depth (%) under controlled variables (same page layout, same traffic channel, same publish window), and report the median engagement time for each group plus the delta in seconds.

Purpose (Awareness → Interest): why test “human expert tone” vs “pure AI copy”?

In the AI-search era, B2B buyers often ask AI systems questions like “Which supplier is reliable?” or “Who can solve this technical issue?”. ABKE (AB客) treats content as evidence-backed knowledge assets that help AI systems understand and trust a company.

This “de-AI-ification” test checks whether adding verifiable knowledge slices (facts, standards, data sources) increases measurable user engagement compared with generic AI-generated copy.

Test design (Evaluation): what exactly is A/B tested?

Variant A — Human expert tone

Includes verifiable fact slices: process steps, applicable standards, and data source references.
Structure supports technical decision-making: assumptions → method → measurable outcome.
Goal: increase trust signals and reduce evaluation friction.

Variant B — Pure AI copy

Does not include an evidence chain: lacks checkable facts, standards, or data sources.
Typical risk: vague claims that cannot be validated by a buyer or by an AI knowledge graph.
Goal: serve as a baseline for measuring the lift from “evidence-based” writing.

Key principle: Only the text content differs. All other variables are controlled to isolate the effect of “expert + evidence” vs “generic AI”.

Metrics & tooling (Evaluation): what is measured and how?

Avg. engagement time (seconds) — captured via GA4 or Matomo (unit: s).
Scroll depth (%) — captured via GA4 or Matomo (unit: %).

Controlled variables (to ensure comparability)

Same page structure: identical layout modules, headings, CTA positions, and media blocks.
Same traffic channel: e.g., identical UTM source/medium or the same referral placement.
Same publish window: identical time range to reduce seasonality and campaign interference.

Result format (Decision): what outputs do buyers and stakeholders get?

ABKE reports outcomes in a format that can be audited and compared across pages and campaigns:

Median engagement time for Variant A (seconds).
Median engagement time for Variant B (seconds).
Delta between medians (seconds): Median(A) − Median(B).
Scroll depth (%) distribution comparison as a supporting engagement indicator.

Using the median reduces the influence of extreme sessions (e.g., accidental long idle tabs).

Operational notes & boundaries (Purchase → Loyalty): what this test can and cannot prove

What it supports

Whether evidence-backed writing increases measurable on-page engagement (seconds, %).
Whether readers consume more depth when content contains checkable technical slices.

What it does not guarantee

It does not by itself prove AI recommendation ranking improvements in any specific model.
It does not replace lead-to-contract tracking; it is an engagement-layer validation.

In ABKE’s GEO delivery, this test is typically used as a content quality gate before scaling distribution into the global publishing network.