E-E-A-T is four questions about the person behind a page, and a text answers them one at a time. On 2026-10-03 we ran claude-seo's content command over a post this blog published on 2026-09-28, and the four factors came back far apart: 18/20 for Experience, 16/25 for Expertise, 14/25 for Authoritativeness and 25/30 for Trust, inside a quality score of 73/100. That spread is the lesson, because the half you produce by working costs an afternoon and the half other people hand you takes months. This lesson reads each factor against the sentences that earned it, then opens the evidence bank seodraft keeps so an article can only cite material a human accepted. On the same day the bank held 135 pieces, and the five it rejected are the most useful part of it.
What the four letters ask, one at a time
Experience asks whether you did the thing, Expertise whether you understand the field, Authoritativeness whether anybody else says so, and Trust whether the page is straight about all three. Four separate questions produce four separate scores over one text, which is why a single "E-E-A-T score" tells you nothing you can act on.
The same command also checks Google's three questions about helpful content: who made the page, how it was made, and why it exists. Our post passed all three, and the report named the "How this post was made" block as the reason: it lists the research, the evidence, the brief, the rules and who approved it. A page that answers those three in its own body is answering the rater's question before anybody asks it.
Experience is the factor you can earn this week
Experience scored 18/20, and every point came from material we already had on the desk. The post names the weekend it describes, 26 and 27 September 2026, and reports what each catalog did to us: a verification checklist at 53/100 on one directory, 37 tools read off the server, a 401 returned to a health check because the server requires a login.
Two of the screenshots are ours, taken from the actual listings rather than pulled from a stock library. The mistakes are in there too, including the five directories that stopped us at signup for five different reasons. Nobody else can write those sentences, and that is the whole mechanism: Experience is the only factor that rewards you for having been present.
Expertise and Authoritativeness are paid in different currencies
Expertise came back at 16/25 and Authoritativeness at 14/25, from the same body of text that scored 18/20 on Experience. Expertise lost its points on depth: the post explains that the registry wants a server.json and a DNS TXT record, and shows neither of them. A reader who arrived to copy something left with nothing to copy.
Authoritativeness lost its points on citation. The post links the live listings it talks about and never links the registry's own documentation or the schema it asks for, so every technical claim rests on the author alone, on a domain barely months old. The run's own fix covers both at once: one code block with a minimal configuration file and a link to the official documentation beside it.
Trust is dates, disclosure and a way to reach you
Trust scored 25/30, the highest proportion of the four, for three things that cost nothing to add. The post admits it has no measured results yet and says so in the body rather than in a footnote. It explains how it was made, and it gives an email address a reader can write to.
The points it lost came from a date. The page said 10 October 2026 while it was live and being read on 3 October, which contradicts a sentence in its own text about going live in September. A future date is a trust problem before it is a technical one, and lesson 11 takes apart how it got there.
The bank decides what an article is allowed to claim
seodraft keeps first-hand material as rows in an evidence bank, and an article's brief has to name at least three accepted pieces before the body is worth writing. Each row carries a kind: a measured number, a named case, a mistake made, a credential, a screenshot still to take. The brief of the post this lesson scores named 15 of them, all confirmed by the founder in one conversation on 2026-09-27: four measured data points, three named cases and eight experiences from that weekend.
Two rules read the plan against the finished article. A brief leaning on a piece nobody accepted fails evidence-unaccepted, and a piece the plan named that never reaches the text fails evidence-unused. Collecting evidence as decoration is the failure those two exist to catch.
Who produced the claim decides its state
A piece the human said in conversation arrives accepted, and a piece the agent inferred, measured or fetched waits as pending until a human rules on it. The flag is saidByHuman, and the agent sets it true only when it is transcribing. Everything else is a proposal, however correct it looks, because the point of the state is to record who is standing behind the number.
Our bank on 2026-10-03 held 135 pieces: 130 accepted, 5 rejected, none pending. Of the accepted ones, 48 were said by the human and 82 were proposed by the agent and approved afterwards. The proportion is worth sitting with: two thirds of what the bank holds was drafted by a machine and survived a human reading it, which is a different thing from material a machine produced unsupervised.
The pieces that look like evidence
Eighteen of our 130 accepted pieces carry a weakFirsthand flag, which marks a general statement anybody in the same field could repeat. They are accepted and usable, and an article built on nothing else reads as a page about an industry rather than a page about you. The flag exists so the writer can see that distribution before starting.
The five rejected pieces are the clearest lesson in the bank. One of them reads "a recent study indicates that 67% of companies already use AI", with no study attached. It passes any rule that counts evidence, it reads like authority, and nobody can check it. A figure with no owner is the shape this whole bank exists to keep out of your articles.
Write the three sentences before you open the draft
Settle which three accepted pieces your next article leans on, and write them out as full sentences with their numbers and dates in them. If you cannot reach three, the article is research you have not done yet, and the honest move is to measure something this week and write next week. Keep the ones you reach in the bank rather than in a draft, so the next article starts from a stock instead of from memory.
Then look at the two factors other people pay for. Link the documentation you are paraphrasing, give the author a page and a profile, and add the code block a reader came to copy. The feature page at features/eeat-evidence shows the bank with its states, and the next lesson takes the same post through the rules that read how it is written.