Six disciplines have answered the question of whom to believe. Search borrowed the youngest answer.

Experience, expertise, authoritativeness and trustworthiness. The four words describe what a human evaluator is asked to look for when judging how much a page and its creator can be relied on. They are written with hyphens because the first two letters are different words that begin with the same letter.

From a single document: the General Guidelines that Google publishes for the people it pays to evaluate search results. The current edition runs to 182 pages and carries the date 11 September 2025. It is public and free, which is unusual for an internal working instruction, and it is the only primary source on the subject. Everything else written about E-E-A-T is commentary on that document.

The three older letters entered the document in 2014 as E-A-T. Experience was added in December 2022, which turned three letters into four. The addition was a response to a specific problem: pages written competently by people who had never done the thing they were describing.

No. There is no E-E-A-T score, no field in an index, nothing a system reads off a page. The four letters describe what the evaluators should look for so that Google can tell whether a change to its systems made results better or worse. The signals that do get computed are meant to correlate with what the evaluators see, which is a much looser relation than a factor.

No. Their ratings do not touch the pages they looked at. The ratings measure the quality of a proposed change across many results at once. A page can be rated poorly by an evaluator and keep its position, because the rating was never addressed to that page.

Because it is the clearest written statement of what the system is trying to reward. A ranking algorithm cannot be read. This document can. It describes the target that the machinery is tuned towards, in ordinary language, with worked examples. That makes it the most reliable available description of the intention behind the ranking.

Whether the person writing has done the thing. Used the product, visited the place, lived through the illness, filed the tax return. It is first-hand contact with the subject, and it is the one quality that cannot be acquired by reading. A well-researched review by somebody who never opened the box fails on experience while passing on everything else.

Expertise is trained knowledge of a field, checkable against what the field holds. Experience is contact with a particular case. A doctor has expertise in an illness; a patient has experience of it. For many questions the reader wants both, and for some questions the patient is the better source.

By what the text shows rather than by what it claims. Depth that a layperson could not fake, correct use of the field's own vocabulary, awareness of where the field disagrees with itself, and restraint at the edges of the subject. The guidelines are explicit that everyday expertise counts: somebody who has kept bees for twenty years is an expert on bees whether or not anybody certified it.

Against everybody else who writes on the same subject. Authoritativeness is comparative and it is external: it asks whether the wider field treats this source as the one to consult. A page cannot make itself authoritative by saying so. The judgement is formed elsewhere, by other people, and the page merely inherits it.

Because the other three are worthless without it. Deep expertise applied dishonestly is more dangerous than ignorance, since it is harder to catch. The guidelines place trust at the middle of the family and treat experience, expertise and authoritativeness as supports for it. A page that fails on trust fails outright, however impressive the rest looks.

Accuracy that survives checking, honesty about who is responsible, transparency about how money is made, a reachable address, corrections when something turns out wrong, and the absence of any attempt to mislead about what the page is for. Most of it is disclosure rather than skill.

Your Money or Your Life: subjects where a bad answer can damage somebody's health, financial stability, safety or well-being. Medical questions, legal questions, financial questions, questions about elections and civic processes. On these subjects the standard rises sharply, and the source's standing matters more than the writing.

The evaluator, by asking what happens to a reader who acts on the page and is misled. The test is the size of the possible harm rather than the category of the site. A cooking site giving advice on food allergies has walked into YMYL territory without changing subject.

A rating of how well a page achieves its purpose. It begins with a question that sounds trivial and is not: what is this page for? A page can only be rated once its purpose is clear, because harmful and useless are judgements about a purpose. E-E-A-T is one input into that rating rather than the rating itself.

A second, separate scale, asking whether this result serves the person who typed the query. A page can be excellent and still fail Needs Met because it answers a different question. The two scales are deliberately independent, and the distinction is more useful than most of what is written about relevance.

Deception, mainly. A page that hides who is behind it, disguises advertising as editorial content, imitates another organisation, or is created to harm somebody. Absence of skill is rated poorly; intent to mislead is rated Lowest. The guidelines devote a whole section to the case where the creator's reputation, rather than the page, is what fails.

The evaluator is instructed to look for what independent sources say about the site or the person, outside anything they control. Not the About page, not the testimonials, but reviews, news coverage, professional bodies, complaints. The standing of a source is established by other people, and the document asks the evaluator to go and find them.

Relevance assessment. Since the Cranfield experiments of the late 1950s the field has employed human assessors to judge documents, and it has studied the judgements themselves as an object. It knows what happens when you ask people to rate quality, because it has been doing exactly that for sixty-five years.

That assessors disagree with each other constantly, and that the disagreement does not ruin the result. Two trained judges asked to rate the same document routinely differ. Yet when the ratings are averaged across enough documents, the comparison between two systems stays stable. The lesson is uncomfortable and freeing: individual judgements of quality are noisy, and the noise cancels at scale. A single evaluator's opinion of your page carries almost no information. The distribution across thousands of pages carries all of it.

Testimony, and the trustworthiness of the person giving it. The field studies how anybody comes to know something they cannot check, which is nearly everything anybody knows. It has a worked vocabulary for the problem: epistemic dependence, epistemic vigilance, peer disagreement, and the conditions under which accepting somebody's word is reasonable rather than lazy.

That credibility is distributed, and that it is distributed unfairly. The same claim receives different amounts of belief depending on who makes it, and the difference tracks accent, gender, institution and class as much as it tracks accuracy. A rating system built on human judgements of authority inherits that pattern whole. Nothing in the document acknowledges it, and no measurement in the field would detect it, because the evaluators and the population share the same prior.

Ethos, the credibility of the speaker, one of the three means of persuasion alongside logos and pathos. Aristotle broke it into three parts: practical judgement, good character and goodwill towards the audience. Written in 350 BCE, and the correspondence to the modern four is close enough to be startling.

Practical judgement covers expertise and experience: knowing the subject and having handled it. Good character covers trustworthiness: the audience believes the speaker is not trying to trick them. Goodwill has no counterpart in the four letters at all, and it is the one that decides whether a reader stays. Rhetoric also holds that ethos is established inside the speech rather than brought to it, which is the older and better statement of what authoritativeness means.

Source credibility, and it measured the thing experimentally in 1951. Identical texts were given to readers under different attributed sources and the attributed source alone changed how much the readers believed. Two dimensions came out of that work and have survived every replication since: perceived expertise and perceived trustworthiness. Two of the four letters were established empirically seventy-five years before they were written into a guideline.

That source credibility does most of its work when the reader is not thinking hard. Under close attention people weigh the argument; under light attention they weigh the source. Most reading is light attention. This means the visible marks of a credible source, the named author, the disclosed interest, the reachable address, are doing more work than the quality of the reasoning for the majority of readers, which is an unflattering finding and a useful one.

Disclosure duty, and unlike the other five it is enforceable. Who operates this site, how to reach them, when advertising must be labelled as advertising, what has to be said before somebody buys. In Germany the imprint requirement makes a large part of trustworthiness a legal obligation with penalties attached, and the transparency rules of the Digital Services Act extend it across the European Union.

Because it converts a preference into a floor. The guidelines ask an evaluator to look for who is responsible for the content; the law already requires the answer to be on the page. Any site that satisfies its disclosure duties properly has satisfied a substantial part of trustworthiness as a side effect, and any site that treats the imprint as a formality has left the most checkable part of its credibility undone.

The search one. It is the only one of the six with no measurement procedure of its own. Information retrieval has assessor protocols and agreement statistics. Media psychology has controlled experiments. Rhetoric has a structural analysis. Law has courts. Social epistemology has two thousand years of argument. The search reading borrows human evaluators, receives a distribution of opinions back, and the industry then discusses the result as though it were a property of the page. That gap between a borrowed judgement and a page attribute is where nearly every piece of bad E-E-A-T advice comes from.

Three of the four can be shown, and one has to be earned. Experience can be shown, by writing what only somebody who was there could write. Expertise can be shown, by depth and by accuracy at the edges. Trustworthiness can be shown, almost entirely through disclosure. Authoritativeness is formed by other people talking about you, which no change to the page reaches.

Naming the author, with enough about them to establish why they may speak, and making that page reachable from the article. It is the cheapest of all the changes, it satisfies the disclosure duty at the same time, and it addresses the dimension that seventy-five years of credibility research identified as the strongest. Most sites still publish unsigned articles.

What the person has done, stated concretely and checkably, rather than adjectives. Years in the field, the institution, the qualification, the cases handled, the things built. A link to somewhere outside the site that confirms it. An unverifiable claim of expertise on a page you control is worth nothing to a reader and nothing to an evaluator.

How the page makes money. Affiliate arrangements, sponsorship, the commercial relationship to anything reviewed. The disclosure costs a sentence and buys the reader's willingness to keep reading a recommendation. Concealment, when discovered, moves a page from mediocre to Lowest in the rating scheme, which is the steepest penalty in the document.

Detail that could not have been researched. The thing that went wrong, the step the instructions leave out, the measurement you took, the photograph you made. A generated summary of the subject can be produced by anybody; what happened when you did it cannot. This is also the part that survives being quoted by an answer system, because an extract of a specific first-hand detail carries the source with it.

Not by watching rankings, because there is no score to move. Watch instead for the things that would follow if the standing had changed: being named as a source rather than paraphrased, being cited by other sites without asking, direct and branded queries rising, and readers arriving at the author page. Those move slowly and they mean something. A ranking that moves in a week means the query changed.

Four letters describe an old problem, and they describe it less precisely than five other fields already do. They are worth taking seriously because they state what the largest search system is aiming at, in public, in plain language. They are worth holding loosely because they are one field's recent vocabulary for a question that has been worked on since antiquity, and the older answers are in several respects better.