Blog · 2026-10-15 · baryreview editors

E-E-A-T for AI-assisted content: passing the who, how and why test

Most questions about E-E-A-T and AI content come down to one worry: “Will Google penalize this page because a model wrote the first draft?” The short answer, from Google’s own documentation, is no. Google’s guidance since February 2023 has been about “rewarding high-quality content, however it is produced”. The longer answer is that “however it is produced” carries a lot of weight. The same documentation sets out a self-assessment built around three questions, who, how and why, and an AI-assisted page can fail all three without a single grammatical error.

This post walks through those questions, what experience, expertise, authoritativeness and trust actually look like on a page a model helped write, and what a human editor adds that the model cannot.

What Google actually asks

Google’s page on creating helpful, reliable, people-first content describes E-E-A-T as “a mix of factors that can help determine which content demonstrates aspects of experience, expertise, authoritativeness, and trustworthiness.” It then adds a line worth memorizing: “Of these aspects, trust is most important.”

The same page lays out a “Who, How, and Why” self-assessment. The questions are plain. On who:

“Is it self-evident to your visitors who authored your content?” “Do pages carry a byline, where one might be expected?”

On how, the document is explicit that it covers “automated, AI-generated, and AI-assisted content,” and asks:

“Is the use of automation, including AI-generation, self-evident to visitors through disclosures or in other ways?” “Are you providing background about how automation or AI-generation was used to create content?” “Are you explaining why automation or AI was seen as useful to produce content?”

On why, the test is whether the content exists to help people or to rank. Using automation “to produce content for the primary purpose of manipulating search rankings” is called out as a violation of the spam policies, and the scaled content abuse policy names “using generative AI tools or other similar tools to generate many pages without adding value for users” as an example.

One more line matters. Google says that “using generative AI to produce large amounts of text without manual oversight or curation represents little to no effort.” Oversight and curation are human verbs. The documentation does not say a model cannot draft. It says somebody has to be responsible for what the model drafted.

E-E-A-T AI content: what each letter means on a page a model drafted

The four words get repeated so often they stop meaning anything. Here is what each looks like when you open an AI-assisted page.

Experience

Experience is evidence that the author has done the thing the page is about. A model has not done anything. It has read about people doing things. So every sentence that sounds like experience in a raw draft is borrowed, and a careful reader can tell.

On the page, experience shows up as details a model could not have produced from training data: what the product felt like after a month, the step that always goes wrong, the price you were actually quoted, a photo you took. If a plumbing company publishes “How to choose a tankless water heater,” experience is the paragraph that says which models the company has stopped installing and why. That paragraph has to come from a person, and it is usually the only paragraph a competitor cannot copy.

A practical test: read each paragraph and ask, “Could someone who has never done this have written it?” If the answer is yes everywhere, the page has no experience in it, whoever typed it.

Expertise

Expertise is knowing the subject well enough to be right, and to know where the draft is wrong. A model’s draft is fluent in every subject and expert in none. It will produce a confident, specific, plausibly sourced statistic about your industry that does not exist. The expertise on an AI-assisted page is, in large part, the act of removing those.

On the page, expertise looks like correct terminology used consistently, the right level of detail for the stated audience, numbers that resolve to a real source, and the absence of claims the author could not defend if challenged. It also looks like scope. An expert knows what the page should not try to answer, and says so.

Authoritativeness

Authoritativeness is whether other people and other pages treat you as a source. It is mostly earned off the page, through being cited and linked, and you cannot prompt your way to it. What you can do on the page is make the author identifiable and the organization accountable: a real name, a real role, a page that explains who the company is and how to reach it.

Google’s document is blunt about the shortcut: fabricating creator profiles, “such as by using AI-generated headshots, made-up names, or false credentials,” to make content look as if human experts wrote it “is a form of deception.” The Sports Illustrated case in 2023, where product reviews ran under invented author identities, is the cautionary example.

Trust

Trust is the sum of the other three plus one thing: the page is accurate and says what it is. On an AI-assisted page it is also the most fragile, because the draft looks trustworthy before anyone has checked it.

On the page, trust looks like sources linked inline, dates on the content, a correction policy you follow, and a disclosure of how AI was used that a reader can find without hunting. It also looks like an absence: no claims your product cannot keep, no legal, medical or financial statements beyond what you are qualified to make.

The who, how and why test applied to a real workflow

Here is how the three questions play out for a small team that drafts with a model.

Who. Put a byline on the page, and make it a person who read the final version and will answer for it. If the model drafted and a marketing manager edited, the manager is the author. Link the byline to an author page that says what this person does and why they are a credible voice on this topic. Do not invent credentials, and do not use a stock headshot with a made-up name. If nobody on the team is willing to put their name on a page, that is information about the page.

How. Decide on a disclosure and apply it consistently. It does not need to be long. “Drafted with AI assistance and edited and fact-checked by [name]” at the bottom of a post answers Google’s question and tells readers what they are looking at. Google asks whether you are “providing background about how automation or AI-generation was used.” A site-wide editorial policy page does this once for every post.

Why. Be honest about the reason the page exists. If the only answer is “we needed a page for this keyword,” the page will read like it. If the answer is “customers keep asking us this and we have an opinion,” the model’s draft is just a faster way to get to the editing.

A page-level E-E-A-T checklist for AI-assisted content

Run each page through this before it goes live. It takes ten minutes and catches most of what matters.

1. Byline. A named person who read the final version. Linked to an author page with a real role and contact route.
2. Disclosure. A short, consistent note on how AI was used, on the page or in a linked editorial policy.
3. One first-hand detail per section. A decision, a number you measured, a mistake you made, a photo you took. If a section has none, it is filler or it needs a person.
4. Every statistic resolves. Each number links to a primary source you opened. If the source does not exist or does not say that, the number goes.
5. Every quote is real. A model will generate plausible quotes from real people. Confirm each one or cut it.
6. Claims about your own product are true today. Features, prices, turnaround, guarantees. A model will describe the product it imagines you have.
7. Scope is honest. The page does not give legal, medical or financial advice beyond what the author is qualified to give, and says so where it stops.
8. Dates and a correction route. A publish date, an updated date when it changes, and a way to report an error that someone reads.
9. The title’s promise is kept. If the title says “how to choose,” the body contains a decision, not a survey of considerations.
10. It would survive a skeptical reader. Read it as your most informed customer. Every place they would say “that’s not quite right” is a trust leak.

What a human editor adds that the model cannot

A model can draft, summarize and rewrite. What it cannot do is the part E-E-A-T is actually measuring.

Judgment. The draft will contain sentences that are grammatically perfect and substantively empty, and sentences that are confidently wrong. Telling the two apart requires knowing the subject and the audience. A model does not know which of its own sentences it made up. An editor who knows the field reads the stat and asks where it came from.

First-hand knowledge. The model’s experience is secondhand by construction. The only way to get a first-hand detail onto the page is for someone who has that detail to put it there. This is the “E” that was added to E-A-T, and it is the one a model structurally cannot supply.

Accountability. Google’s “who” question has a simple implication: a byline is a promise that a person stands behind the page. A model cannot be held to anything. When CNET corrected 41 of 77 AI-assisted explainers in 2023, the corrections were made by people, under the publication’s name. The accountability was always human. The error was letting the draft skip the person.

This is the practical argument for a review step. One informed reader going through the page paragraph by paragraph, marking what is wrong and what is missing, is the “manual oversight or curation” Google describes. If nobody on your team has time to be that reader, that is the gap a professional editor’s review fills: you keep the draft, the tool and the voice, and a person checks the facts and claims before the byline goes on.

A short before-and-after

A fictional SaaS company, Lumen Analytics, drafts a post titled “How long does a data warehouse migration take?” The raw draft’s opening:

Data warehouse migrations are a critical undertaking for modern businesses. According to a 2024 Gartner study, 68% of migrations exceed their original timeline. It’s important to note that careful planning, stakeholder alignment and robust testing are essential for success.

Three problems: the Gartner figure cannot be found; the third sentence is a triad of abstractions; nothing here required having done a migration. After an editor’s pass:

We have run eleven warehouse migrations for customers in the last two years. The shortest took five weeks. The longest took seven months, and the reason was not the data; it was that nobody had written down which of the 240 dashboards were still in use. Here is what we now do in the first week to avoid that.

Same topic, same author byline, same tool used for the draft. The second version has experience (eleven migrations, a specific failure), expertise (the diagnosis), an accountable voice (“we”), and no unverifiable numbers. That is what E-E-A-T looks like on a page a model helped write.

Takeaway

Google’s test for AI-assisted content is not “was a model involved” but “who is responsible, how was it made and why does it exist.” Answer those on the page with a real byline, a plain disclosure, first-hand details and sources that resolve. Then have one person who knows the subject read every paragraph before it goes live, because trust is the part the model cannot supply.