Why this page exists

Shelf-level stories move
billion-dollar supply chains.
So we hold ourselves
to the same standard.

A sustainability claim at shelf level is not a marketing decision. It is a market signal. When enough consumers choose the regenerative option, the category tilts. Retailers update range reviews. Buyers add sourcing requirements. Farms change practices. One label claim, credibly told and consistently verified, can reshape a supply chain that feeds hundreds of thousands of people.

The brands we score understand this intuitively. The brands that don’t are losing shelf space to competitors who do.

The compliance reframe
Reporting isn’t a burden. It’s the weapon.

SB 253 made disclosure mandatory for large companies and every significant supplier in their chain. Most brands treat this as a compliance cost. The ones winning buyer relationships treat it as a positioning strategy — their data is ready before anyone asks, their numbers are clean, and they use disclosure to set the benchmark their competitors get measured against.

The credibility test
You can’t hold brands to a standard you won’t apply to yourself.

YKO scores companies on five signals: specificity, transparency, third-party verification, impact coverage, and the ratio of action to marketing. We run investigations. We surface gaps. We flag claims that aren’t supported by evidence. If we won’t do that work on ourselves — publicly, with the same rigor — none of it means anything.

Differentiation

Lead disclosure sets the benchmark. Late filers get audited against your numbers. That’s not accidental — it’s leverage.

Retail access

Buyers ask the same 23 questions. The brands that pass fastest aren’t always most sustainable. They’re most prepared.

Capital access

Disclosure quality signals operational maturity. In a world of ESG due diligence, your report is your pitch.

GreenSpecs — Self assessment

We score brands.
Here’s how we score ourselves.

We ask companies to be specific about what they emit, who verified it, and whether their biggest impact is covered. We’re applying the same rubric to our own operations — with the same honesty we’d expect from anyone else.

Transparency score
7  / 10
GreenSpecs
v2.1  ·  Self-reported, May 2025
Specificity
Transparency
Third-party verified
Biggest impact covered
Action vs. marketing
Operations

What we actually run

Source Estimated footprint Notes
Scope 2
MacBook (founder compute)
~16 gCO ₂e/day
~4 kg/year
~25W active, 3 hrs/day. California grid at 207 gCO ₂e/kWh (CAISO 2024). One person.
Scope 2
Cloudflare Workers + Pages + D1
Disclosed renewable Cloudflare has operated on 100% renewable energy since 2019. Per-request edge compute is microseconds. Not added to the tally — we can’t measure it, but we disclose the basis.
Scope 1
Business travel
Near zero Product testing done online. Local store trips for physical scanning are incidental commuting, not business travel. No flights.

Methodology: Scope 1/2 follows GHG Protocol Corporate Standard. MacBook estimate uses Apple’s published average power draw for M-series chips under active load (~25W) times hours of productive use times CAISO’s 2024 average grid carbon intensity. No adjustment for Apple Silicon efficiency claims — we use a conservative load figure.

Per scan

What one scan costs

Every time someone points their camera at a product, three things happen: their iPhone draws power, data moves over the cell network, and our AI reads the label. We’ve estimated the carbon cost of each step. We assumed iPhone 15 on Verizon or AT&T LTE.

Step gCO ₂e / scan Notes
Scope 3
iPhone 15 — device
~0.009 g Camera + screen active for ~30s at ~3.5W = 0.029 Wh. US average grid at 386 gCO ₂e/kWh.
Scope 3
LTE data transfer
~0.018 g ~500KB image upload + ~3KB JSON response. LTE energy intensity ~0.08 kWh/GB (Verizon/AT&T 2024 estimate).
Scope 3
Gemini 2.0 Flash inference
~0.015 g Industry estimate of ~0.0002 kWh/query for efficient flash-class LLMs on TPUs × Google’s 73 gCO ₂e/kWh effective carbon intensity (64% CFE, 2023). This is the most uncertain figure on this page.
Scope 3
Cloudflare routing + D1 write
Disclosed renewable Served from Cloudflare edge, 100% renewable. Not added to total.
Uncached scan total ~0.042 g Device + LTE + inference.
Cached scan total ~0.027 g Same product scanned again — inference skipped. Device + LTE only.
Put in context — gCO ₂e per action
Google search
~0.20 g
1 min streaming video
~0.036 g
GreenSpecs (uncached)
~0.042 g
GreenSpecs (cached)
~0.027 g

Scope 3 emissions are attributed as Category 11 (use of sold products) for device and network, and Category 1 (purchased goods and services) for Gemini inference. We follow GHG Protocol Scope 3 Standard framing even though we’re a free service — the attribution logic is the same.

We do not have a measured figure for Gemini inference energy. Google does not publish per-query energy data for Gemini. Our estimate uses ~0.0002 kWh/query drawn from published academic estimates for flash-class models on TPU v4 hardware. This should be treated as an order-of-magnitude estimate, not a precise number.

Why 7 and not 10

Signal by signal

01 — Specificity 1 / 2

We have numbers — that’s better than adjectives. But our biggest number (Gemini inference) is an estimate drawn from academic literature, not a measured figure. Google doesn’t publish per-inference energy for Gemini. Until they do, or until we instrument our own energy telemetry, this stays at 1. A company should say “80% recycled” not “more recycled.” We’re saying “~0.015 g” when we don’t actually know if it’s 0.005 or 0.04.

02 — Transparency 2 / 2

Every assumption on this page is visible. Every number shows its derivation. The methodology block is not buried in a footnote — it’s right under the number it qualifies. We’ve named our sources (CAISO, published LTE energy studies, academic LLM benchmarks, Google’s Environmental Report). You can check every one of them. That’s 2.

03 — Third-party verified 0 / 2

This report is self-authored. No external auditor has reviewed it. No certification body has verified our methodology. We’re not B Corp certified, nor pursuing any certification. Our position is that a transparent, self-reported methodology is more honest than a certified number whose methodology is opaque. But we’re applying our own rubric fairly: without third-party verification, this signal is 0. That’s the right call.

04 — Biggest impact covered 2 / 2

For a software product with one employee, the biggest controllable impact per unit of value delivered is AI inference. We covered it. We didn’t lead with the easy number (MacBook energy) and bury the hard one. Scope 3 inference is the most uncertain and the most interesting figure — it’s the first thing on the per-scan table.

05 — Action vs. marketing 2 / 2

The architecture decisions that reduce our footprint are real, not rhetorical. Scan results are cached — identical products skip inference entirely. We use a waterfall model selection strategy: cheapest and fastest model first, larger models only when needed. We chose Gemini 2.0 Flash specifically because it’s more efficient than larger vision models. We run entirely on Cloudflare’s renewable-powered edge rather than a traditional data center. These are engineering choices, not marketing claims.

Actions

What we’re actually doing

Scan caching
When a product has already been scanned, we return the stored result. No AI inference, no compute cost. Cache hit = Gemini footprint drops to zero.
Model efficiency
Gemini 2.0 Flash is the most efficient vision model in the Gemini family. We chose it explicitly. We don’t use Pro or Ultra for standard scans.
Edge compute
All serving runs on Cloudflare Workers — renewable energy declared since 2019, and edge nodes are more efficient than centralized data centers by design.
Waterfall model selection
Simple analysis paths route to faster, cheaper models. Complex or ambiguous labels escalate. No unnecessary compute for straightforward claims.
What we can’t tell you yet
Gemini’s actual inference energy. Google doesn’t publish per-query figures for Gemini. Our number is an estimate. If they publish this data, we’ll update this page within 30 days.
User device grid mix. We assumed the US average grid (386 gCO ₂e/kWh). A California user on clean power is lower; a coal-heavy grid is higher. We can’t know at scan time.
Gemini training compute. A share of Gemini’s training energy is attributable to every query, amortized across billions of calls. We can’t calculate our portion. We acknowledge it exists.
Third-party verification. This document is self-authored. We’d rather say that plainly than hide it in the footnotes of a certification we don’t have.