The same website audit, with about half the data sent to the large language model
This is a before/after measurement of the EcoGEO Audit website audit. The same 20 real websites were run two ways — sending the raw page as is, and selecting the key content with our patented method first — and we compared model usage, cost, speed and accuracy.
- Measured 2026-11-05
- 20 real websites (Taiwan 12, Japan 8)
- Model: Claude Sonnet 5
- “Before” was run twice as a noise baseline
Model usage figures are based on the input/output metering reported by the model provider.
One audit, two ways of sending it
Same 20 websites, same model. The only difference is what gets sent. What improved and what got worse are both shown here.
We sent the first ten-thousand-odd characters of raw HTML — mostly <head>, styles and scripts. The actual body text often never made it in.
A feature summary of the page is extracted first and shared by both analyses.
Input sent to the large language model
−49.6%Two analyses combined; average per site 15,343 → 7,730
Model output
+9.4%Output increased (total across 20 websites)
Input + output combined
−43.7%Total across 20 websites
Model cost
−28.6%Total across 20 websites (US dollars)
Average latency
3.6 s slower3.6 seconds slower; the feature summary itself takes about 0.1 seconds
Model usage figures are based on the input/output metering reported by the model provider.
Did accuracy get worse?
Checked site by site against what each page actually says — sites judged correctly out of 20:
We used to send the first ten-thousand-odd characters of raw HTML, which is mostly <head>, styles and scripts — the actual body text often never made it in. Now a feature summary of the page is extracted first and shared by both analyses, so the model reads content instead of code.
The same method applied to text content generation
Measured the same day: 16 input sets × 2 models = 32 pairs, of which 16 were actually reduced.
- The 16 reduced pairs: main-model input 62,435 → 26,127 (−58.2%); counting the usage of the pre-processing step (rough-set extraction) itself, the net change is −39.2%.
- Input + output combined (including pre-processing): Sonnet 5 −18%, GPT-6 Luna −29%.
- Retention of hard facts supplied by the user: 87.3% → 86.4% (flat).
- The other 16 pairs were not reduced: 14 had inputs too short to reduce and 2 did not get smaller, so the original text was sent and main-model input was unchanged. The pre-processing step still used capacity, so input + output for these 16 pairs went up by +48.3%.
- All 32 pairs combined: main-model input 75,304 → 38,996 (−48.2%); net input including pre-processing −11.3%; input + output combined −3.7%. In other words, the method pays off when the input is long; for short inputs it is currently a net increase.
How it works: three steps
We did not switch to a lighter model. We decide what is worth sending before anything reaches the model.
Fetch the raw page of the site being audited: body text and headings, plus a large amount of styles and scripts that have nothing to do with the judgement.
Avg. input per site: 15,343
Rough-set feature reduction trims the features and multi-criteria decision-making (MCDM) picks the passages worth sending, producing one feature summary of the page.
Avg. input per site: 7,730
Both analyses share the same feature summary, so the model reads the key content.
The method in step 2 comes from Taiwan utility model patent M683041, “Low-Carbon E-Commerce Recommendation System Combining Multi-Criteria Decision-Making”.
Converted to carbonEstimate
Every carbon figure below is an estimate: the measured model usage above, multiplied by one estimation factor.
| Basis | Before (estimate) | After (estimate) | Estimated difference |
|---|---|---|---|
| Per audit | 2.05 g | 1.15 g | −0.89 g |
| Per 1,000 audits | 2,045 g | 1,151 g | −894 g |
| Per 1 million audits | 2,045 kg | 1,151 kg | −894 kg |
Average model usage per audit (input + output combined ÷ 20 websites): 17,043 → 9,590
Estimated carbon = measured model usage × 0.00012 gCO₂e per unit of usage. This is the conservative estimation factor our platform’s content-generation feature already uses. The same factor is applied before and after, so the change equals the change in usage (−43.7%).
The large language model runs in a third party’s cloud, and we cannot measure the electricity it actually uses. This figure is “measured usage × an estimation factor in the publicly reported range”. It is meant for comparing relative differences. It is not measured electricity and not a guarantee of carbon reduction.
Model usage figures are based on the input/output metering reported by the model provider.
How the audit report ranks things
The “AI search recommendation potential” score in the report is calculated by the decision engine: rough-set analysis first derives the weight of each criterion from labelled samples, multi-criteria decision-making (MCDM) then applies those weights, and the result is combined with what the large language model actually read.
- SStructured data completenessWhether Schema.org / JSON-LD is complete.
- LSearch and performanceSearch optimisation and loading performance.
- ECarbon gradeEstimated carbon per page view, converted to a grade from A to F.
- RSemantic richnessWhether there is enough content for AI to understand.
- C_llmWhat the large language model actually readHow much information the model could extract when it actually read the site (extraction coefficient).
- Site score (S, L, E, R weighted by rough-set weights + MCDM) × 60%
- What the large language model actually read (C_llm) × 40%
Before that, C_llm was a multiplier and could only pull the score down; since 2026-10-03 it is a weighted sum.
Web page carbon (gCO₂e per page view and its grade) is a separate chapter of the report, and it also enters the decision engine as criterion E. The weights are not fixed; they are derived by rough-set analysis from labelled samples. With the current samples, semantic richness and structured data carry the higher weights and carbon grade carries a low weight. The ranking is the combined result of several criteria — it is not decided by carbon alone.
Our own Software Carbon Intensity (SCI)
A self-disclosure calculated with the ISO/IEC 21031:2024 method: the carbon emitted each time the EcoGEO Audit decision engine completes one diagnostic decision.
- Functional unit
- One diagnostic decision = the decision engine completes one rough-set feature reduction plus one multi-criteria decision evaluation.
- Measurement
- 2026-08-21, AWS EC2 t3.xlarge (ap-east-2), with GreenCalculatorJavaNoUI 1.0.4 from the Taiwan Institute of Economic Research (TIER): 200 warm-up runs, then 20 rounds of a 60-second idle baseline followed by 1,000 loaded decisions.
SCI measures the decision engine’s computation on our own host. The estimate in the previous section covers the part sent to a third-party large language model. The boundaries are different, so the two figures cannot be added together or compared directly.
Limits and boundaries
Everything this measurement did not cover, and everything that got worse, is listed here.
- The sample is 20 websites and a single model (Claude Sonnet 5); the main model used in production was not part of this measurement.
- Output increased by 9.4% and average latency increased by 3.6 seconds.
- Short inputs see no reduction: short inputs for text generation are not reduced, and the pre-processing step has its own usage.
- Carbon is an estimate (see “Converted to carbon”): measured usage × an estimation factor — not measured electricity and not a guarantee of carbon reduction.
- Since the audit switched to the feature summary, scores differ systematically from older reports (most scores go up once the body text can be read). A method-update note has been added to the report.
Big-O was selected to take part in the POC validation of the voluntary Green Software Label promoted by Taiwan’s Ministry of Digital Affairs, as one of the nine participating companies in this phase.