E-E-A-T: the signals Google actually reads on your site

E-E-A-T is not a ranking factor. Google keeps saying so, and technically it is right: there is no E-E-A-T score inside the algorithm, no dial an engineer could turn. It is an evaluation framework, written for the human quality raters who grade results and act as a check on the ranking systems.

What does exist are signals. Things present in, or missing from, the HTML your pages serve: a byline, a link to a biography, a date, a structured data block, a company identifier in the legal notice. Those signals are not E-E-A-T. They are its machine-readable half, and that half is the only one you control.

So we stopped arguing about it and went counting. 24 French SEO agency domains probed on 27 August 2026, 17 ranked, 170 articles opened and analysed, eleven signals read from the served HTML, one hundred points in total. The scale is published, the ranking too, and the grid at the end of this article scores you on exactly the same scale.

What this score measures, and what it does not

E-E-A-T is four words: Experience, Expertise, Authoritativeness, Trustworthiness. The last one is the centre of gravity, the other three feed it. A human rater judges substance: has this person actually done what they describe, is the company identifiable, does the page deserve trust on this particular subject.

A script judges nothing. It observes. It sees that a name is written, not that it belongs to anyone. It sees that a biography exists, not that it is accurate. It sees that an update date is declared, not that the content ever changed. This ranking scores how detectable the signals are, not the quality of what they claim.

One more limit matters more than all the others, and it applies to every benchmark of this kind: no correlation between this score and SEO performance was computed here. Rankings and traffic for the 17 sites are outside the data available. A score of 98 promises no traffic, a score of 31 announces no penalty. What the score does establish is whether an engine reading your HTML finds enough to tie a page to a person and a company.

If the term itself needs pinning down, the short definition sits in our E-E-A-T glossary entry, and the adjacent question of sensitive content in YMYL.

The scale: eleven signals, one hundred points

The weights are not neutral. They encode an assumption: author identity and company traceability are more expensive to fake than a date tag, so they discriminate better.

• Author name visible on the article: 12 points
• The name links through to an author page: 10 points
• The author page exists and carries a biography: 12 points
• Publication date marked up and visible: 10 points
• Distinct update date: 8 points
• schema.org Article or BlogPosting: 10 points
• author property filled in the schema: 10 points
• Organization on the home page, with at least two sameAs: 10 points
• About page reachable: 6 points
• Legal notice carrying a company identifier: 6 points
• External sources cited in the articles: 6 points

The six criteria that concern articles are averaged over ten articles per site, sampled at regular intervals from the sorted sitemap corpus. That is why the scores are not round numbers: a site where seven articles in ten are signed gets 8.4 points out of 12, not 12 and not 0.

The ranking of the 17

Click a row to open the breakdown of all eleven signals, or sort by criterion to see who wins where.

Three things stand out.

The spread is enormous for an industry that sells this expertise. 67 points separate first from last. Nine sites out of 17 sit at 84 or above, five at 57 or below, and only three occupy the middle. Either the byline, author page and schema chain has been built, or nothing has.

Agency size predicts nothing. Well-established firms land at the bottom while smaller outfits such as agence-slashr.fr, third at 91, or korleon-biz.com, tenth at 73, hold the top. This is not about budget, it is about decisions taken once on the article template.

A 3 or 4 point gap means nothing. The score depends on the ten articles sampled, and on a site whose sitemap mixes blog posts and service pages, it moves noticeably with whichever pages land in the sample. The measurement was then replayed minutes later: all 17 scores came back identical to the point. Reproducible does not mean accurate to the unit.

Where the panel loses its points

Author identity: 34 points at stake, 12.3 lost

This is the sharpest result of the whole measurement. The three criteria forming the identity chain, visible byline, link to an author page, author page with a biography, are worth 34 of the 100 points. The panel averages 21.7 of them.

The drop reads criterion by criterion. 75% of the 170 articles carry a visible author name. Only 56% link that name to an author page. And 59% of sites actually have an author page with a biography on it. Put plainly: one article in four is signed by nobody, and among those that are signed, one in four again signs with a name that leads nowhere.

The common failure is not a missing writer, it is a plain-text name. The template prints "By So-and-so", the reader sees it, and the chain stops there: no link, no page, no bio. Six percent of articles do declare an author URL in their markup, but it points at a LinkedIn profile. That counts as a declared profile, not as an author page on the site: the two are measured separately because they do different work. An external profile ties a name to a social identity. An author page on your own domain ties a name to your company.

External sources: the panel's weakest point

Of the eleven criteria, external sources is where the panel captures the lowest share of the scale: 55% of the available points. 71% of the 170 articles cite at least one external source, but only 29% cite three or more.

Exactly one agency out of 17 scores full marks here. The deficit is economic as much as editorial: citing a source means shipping an outbound link, and the industry spent fifteen years learning not to. The arithmetic has changed. An article that needs nobody to back up what it claims is precisely what a generative engine cannot verify, and therefore will not repeat. We go into it in our piece on optimising content for AI citations.

Dates: marked up for the machine, silent for the reader

93% of articles carry a marked-up publication date. Only 71% display it on screen. The gap is telling: one article in five is dated in its code and anonymous on the page.

It is almost always deliberate. The visible date comes off so the content does not look old, the schema date stays so the SERP display does not break. The result is a contradictory signal, and above all a page asking for trust while refusing to say when it was written. 72% of articles also declare an update date distinct from the publication date, which is the right instinct, provided the update actually happened.

Schema: present, but anonymous

75% of the 170 articles declare a schema.org Article or BlogPosting carrying an author property. Only 56% type that author as a Person. The rest attribute the article to the company itself, which amounts to declaring that no human stands behind it.

Four sites out of 17 serve no article schema at all and drop the 20 points of both criteria at once. On the company side, 88% of sites declare an Organization node on their home page, but only 47% attach at least two sameAs links to it. Two sites in the panel have no Organization node whatsoever. An Organization without sameAs ties the brand to nothing: a business card with no address. For the full mechanics, see our guide to structured data for SEO and AEO and the structured data entry.

The company: reachable, not always identifiable

100% of ranked sites have a reachable About page and a reachable legal notice. It is the one criterion where the panel is perfect, which figures: the law requires it.

But 18% of those legal notices carry no traceable company identifier: no company number, no trade register entry, no VAT number. A legal page without an identifier fills the slot and proves nothing. That is six points left on the table for ten minutes of work.

What this measurement does not say

A benchmark that hides its blind spots is not a benchmark, it is a brochure. Here are ours.

• We are both judge and contestant. vydera.com, first at 98 points, is the site of the people who ran this measurement. digidop.com, eleventh at 66, is our agency of origin. Both are scored by the same script with no manual adjustment, and digidop.com drops 22 points on author identity exactly like any other site in the panel. Manual checks were run in both directions: the author page and date display on vydera.com, but equally the missing bylines and author pages on six sites at the bottom of the table, so that diligence would not be asymmetric.
• JavaScript rendering is not measured. An author block or a date injected after load is invisible to this script. It is also invisible to any crawler that does not render, which covers nearly every generative engine crawler, but a site penalised here may well display the signal on screen.
• 17 sites is a small sample. The per-site percentages rest on 17 values: one site more or less moves the figure by 6 points. The per-article percentages rest on 170 values and are steadier.
• One error was corrected along the way. A first pass credited korleon-biz.com with an author page that turned out to be a schema identifier resolving to the home page. Corrected, the site loses both the author link and author page criteria, and lands tenth at 73 points. A declared signal is not a served signal.
• Pages served in ISO-8859-1 are decoded as UTF-8 by the script, which mangles their accented characters and can hide a date written out in words on those sites.
• A team page counts as an author page as long as the article byline leads to it. The criterion measures that a name leads to a biography, not that every writer has their own page.

Seven domains probed, no score

Of the 24 starting domains, seven are absent from the ranking, and the reasons say something about the market.

Three brands no longer exist as standalone sites. 1ere-position.fr and open-linking.com redirect to noiise.com, search-foresight.com redirects to peakace.fr. Three consolidations visible in a plain redirect trace.

Three sites had nothing to measure. woptimo.com exposes 57 sitemap URLs, all service and training pages, no articles. yumens.fr exposes 41 URLs, of which nine content pages were tested and all redirect to the home page. primelis.com returned only two articles of at least 300 words, below the threshold of five required to be ranked, its resource library being assembled in JavaScript. It appears as a partial measurement, without a score.

One site never answered. resoneo.com returns a 403 on its home page every time, including after repeated spaced attempts. It is not scored zero: it is absent. An unreachable site leaves the ranking rather than receiving a score nobody could compute.

Score your own site on the same scale

Eleven questions, the same weights, the same definition of each signal. Answer while reading your page source, not the rendered page.

Read your result against the panel median, 84, and its mean, 71.5. The bottom five of the panel sit at 57 or below: at that level the problem is not the content, it is an article template that was never equipped to carry an identity.

Where to start, in order

1. Open the source of one of your articles. Not the inspector, the served source. Look for the author name, the date, the application/ld+json block. Whatever is not there does not exist for a crawler that does not render.
2. Fix the identity chain before anything else. Visible byline, link to an author page, author page with a biography: 34 points, and the place where the panel bleeds most. One page per writer works, so does a team page if your bylines point at it.
3. Type the author as a Person in the schema. An author typed Organization says nobody stands behind the article. It is one line in the template.
4. Display the date you already mark up. Nine articles in ten carry the date in their code. Seven in ten show it. Restoring the display is a template change.
5. Add two sameAs links to your Organization node. LinkedIn and one company profile of your choosing will do. Without them the node ties your brand to no known entity.
6. Put a company identifier in your legal notice. Six points, ten minutes, and it puts you ahead of one site in five.
7. Cite three external sources per substantial article. It is the panel's weakest criterion, so the cheapest place to stand out.

Once the identity chain exists, it is worthless if the pages carrying it are poorly connected: the author page has to be reachable from the articles and from the home page, which is a question of internal linking as much as of templates.

Running the measurement yourself

The method is deliberately plain, and it needs no paid tool.

1. Pull the sitemap, keep the URLs that look like articles, sort them, and take ten at regular intervals. Regular sampling rather than random sampling is what makes the measurement replayable.
2. Fetch each article over raw HTTP, without executing JavaScript, and search the HTML for the JSON-LD block, time tags, byline patterns and outbound links to external domains.
3. Fetch the home page and look for the Organization node and its sameAs entries.
4. Test the usual paths for the About page and the legal notice, then search the latter for a company identifier pattern.
5. Weight, average the article criteria, and add up. Then replay the whole thing once: if the scores move, it is your sampling that is not deterministic, not the market that changed.

Budget a few minutes per site. The result is not a ranking forecast, it is an inventory of what an engine finds when it opens your pages without executing anything. What Google does with it next is not measured here. What it cannot read, on the other hand, certainly counts for nothing.