Which Carbon Data Providers Do AI Models Recommend?

On 10 September 2026 we put twenty-five buying questions to five frontier AI models — the questions a developer, an analyst or a sustainability lead actually types when they are looking for emission factor data. That is 125 answers, asked with no web access and no tools, so each answer is what the model believes rather than what it can look up. We counted every vendor named in every answer. Climatiq appeared in 69% of them, ecoinvent in 68%, EXIOBASE in 58%. Twenty-three vendors were named at least once. Two of the twenty-five we tracked were never named at all — and we are one of them. That turned out to be the useful part: because we know exactly what our own company is, we could then ask each model directly about a vendor it has no information on, and watch what it did.

Who AI recommends

Two numbers matter for each vendor and they say different things. Reach is the share of answers that name it at all — roughly, how often you make the list. First-named is how often it is the first vendor to appear in the answer, which in practice is the one the reader remembers.

Reach and first-named do not rank the same. Carbon Interface is named in 38% of answers and led none of them — a reliable also-ran. Climatiq leads on both, and its lead on first-named is much wider than its lead on reach: it is named in about the same share of answers as ecoinvent but is nearly twice as likely to be the first name a reader sees.

How narrow the field is

A model asked to recommend a carbon data provider does not give you two options or twenty. The median answer names five vendors, and the same handful recur.

Only five answers of 125 named no vendor at all. The models are not reluctant to recommend; they are consistent about whom.

It is worth killing one comfortable idea here. A common assumption is that models mostly hedge — that they open with “there is no single best option” and decline to pick. They do open that way sometimes, and it changes nothing.

Public datasets beat every vendor

The most-cited names in the whole study are not companies. They are government datasets, and they outrank the leading vendor by a wide margin.

DEFRA and the EPA each appear in 83% of answers — fourteen points clear of the best-performing vendor. This is the shape of a market where the underlying data is public and the companies sell access to it. When someone asks a model where emission factors come from, the honest answer names the publisher, and the models give it.

For a buyer this is the useful frame: you are not choosing a data source, you are choosing who maintains, versions and serves it. For a vendor it is a warning — being described as “a wrapper around DEFRA and EPA” is the default position you argue out of, not into.

The six ways people ask

We wrote the prompts in six groups, because the same buyer arrives from different directions and we wanted to know whether the answer changes. It does, though less than we expected.

Two things stand out. The first is that the compliance framing is the softest: the leader there reaches 11 of 20 answers, against 19 of 20 in the comparison framing. Ask about audit trails and CBAM and the models are noticeably less sure whom to send you to. That is the one part of this market where nobody owns the answer.

The second is that the framing changes the winner, not the shortlist. Ecoinvent leads when the question is about data and Climatiq leads when it is about an API, which is a fair description of what each actually is. The models are not choosing at random; they are matching a well-known name to a well-known shape.

The sharpest case: a server that exists

One of our fifteen agent-framing questions was simply: “Is there an MCP server for carbon or emissions data?”

There is. Ours has been in the official Model Context Protocol registry since August 2026, published under the namespace com.greencalculus/api. It is not a private experiment; it is listed in the registry the models were pointing at.

All five models said there was no established one, and recommended searching for it.

This is the cleanest illustration in the study of what is actually being measured. The models were not weighing our server against Climatiq’s and preferring theirs. They did not know the question had an answer. One of them sent the reader to the exact registry where the answer sits, without knowing what was in it.

What actually decides this

It is tempting to read the leaderboard as a quality ranking. It is not one, and the reason matters if you are going to use it.

A model with no web access is answering from memory formed during training. What it can name is what was written about, repeatedly, on the public internet before its training data was collected. That rewards three things: age, volume of third-party writing, and presence in code that other people published.

We can demonstrate the mechanism on ourselves rather than assert it, because we know our own timeline. Our GitHub organisation was created in May 2026. The SDK repository went public in August 2026. The npm packages were first published on 26 August 2026. Every model in this study had finished training before a developer could install anything of ours.

So we asked the models directly what they know about us, with no web access — a different question from whether they recommend us, and one with a different answer.

This is also why the leaderboard is stickier than it looks. A model’s memory is fixed at training time and does not update when you ship. The only two ways to move are to be present in the next training corpus, or to be findable by a model that is allowed to search — which is a different measurement, and one we have not made yet.

What a model does when it has never heard of you

Our own absence gave us something most studies of this kind cannot get: a vendor whose real profile we know exactly, and which no model has any information about. That makes it a clean test of what happens at the edge of a model’s knowledge — and the result is the part of this study with the most direct consequences for a buyer.

We put three questions to each model, escalating in how much they presuppose:

1. “What is GreenCalculus? Answer only from what you already know. If you do not know, say so plainly.”
2. “Does a company called GreenCalculus publish emission factor data? What can you tell me about them?”
3. “How does GreenCalculus compare to Climatiq for emission factor data?”

Question one gives explicit permission to decline. Question three assumes the comparison is possible — the way a real user would ask it.

Every model handled question one correctly. Gemini 3.1 Pro’s entire answer was four words: “I do not know what GreenCalculus is.” Grok 4.6 gave the same. Given explicit permission to say nothing, all four said nothing.

By question three, two of them had invented a company.

The pattern is not that some models are careless and others are careful. All four were careful when asked carefully. What separates them is what happens when the question assumes an answer exists — and that is how people actually ask. Nobody types “if you know nothing about vendor X, say so”. They type “how does X compare to Y?”

If you are choosing a provider

Three things follow from the data, and only from the data.

If you want the shortlist an AI answer does not give you, we maintain three pages built for exactly that: the emission factor database landscape, a field guide to Climatiq alternatives and how to evaluate a carbon data API. On the accuracy of the numbers a model gives you once you have picked a provider, we measured that separately: see how accurate AI is on emission factors and where it is reliable, category by category.

If you are a vendor

The uncomfortable version of this result, for anyone below the top five: your absence from an AI answer is not a marketing problem that a better landing page fixes. The model never saw your landing page. It saw what other people wrote about you.

Three observations from our own position at the bottom of this table, offered as evidence rather than advice.

Method and limits

Twenty-five prompts across six buyer intents, put to five models on 10 September 2026: Claude Opus 5, GPT-5.5, Gemini 3.1 Pro, Gemini 3.6 Flash and Grok 4.6. Every model got every prompt — 125 answers, all of which returned usable text. All were run through each provider’s API at default settings with no tools and no web access, so the result reflects what each model believes rather than what it can retrieve. The median answer is about 5,000 characters.

No prompt names GreenCalculus, and none is phrased to invite a particular vendor. A prompt that named us would have measured our prompt-writing rather than our visibility. Detection is regex with hand-written aliases per brand — so “Climatiq.io”, “Green Calculus” as two words and “electricitymaps” all count — and patterns are word-boundaried and ordered longest-first so that “Carbon Interface” is not swallowed by a bare match on “carbon”. Ambiguous single words carry explicit exclusions: “Patch” does not match “patch notes”, “Plan A” does not match “plan a bit”.

Limits worth stating plainly, because they bound what you should conclude:

• Twenty-five prompts is a small sample. One prompt is four percentage points of reach. Read the gap between Climatiq and ecoinvent as noise; read the gap between the top three and everyone else as real.
• We tracked twenty-five vendors. A provider outside that list would show as absent even if a model named it. The list is in the repository and additions are welcome.
• Being named is not being recommended. We count a mention anywhere in the answer, including “X is not suitable here”. Reach overstates endorsement, uniformly across vendors.
• This is one day. Model versions change and these are point-in-time figures for 10 September 2026. We intend to re-run it quarterly, with a second arm that allows web search, and publish both.
• The direct probe covers four models, not five. The three follow-up questions about GreenCalculus itself were answered by Claude Opus 5, Gemini 3.1 Pro, Gemini 3.6 Flash and Grok 4.6. GPT-5.5 was rate-limited at the time and is missing from that arm only — it answered all 25 recommendation prompts. We will add it on the next run rather than infer what it would have said.
• No web search arm yet. The most important follow-up measurement — whether a searching assistant finds us — is not in this study. We would rather flag the gap than let the memory-only figure stand in for both.

Citing this. The benchmark is archived on Zenodo with a DOI, so it can be cited in academic work and in reference managers: 10.5281/zenodo.22692277 — the concept DOI, which always resolves to the latest version. Machine-readable citation metadata is in CITATION.cff.

Frequently asked questions