I have run FunInformatique, a French-language tech site, for more than 15 years. It has more than 2,000 articles and reached more than one million monthly visits at its peak. When Google's core updates started hitting our traffic, the question was simple: which of those articles still say things that are true?
That question is why we built EEATClean. Before offering it to anyone, we tested it in two ways. First on pages from other sites, in other sectors and languages, with the base engine. Then on our own articles, with the current version built on it. This post shares both results, with the examples and the limits.
Key figures
- 23 of our articles, 367 claims identified, 70 corrections: between 1 and 7 per article, 3 for a typical one.
- Every article needed at least one correction. None had to be deleted.
- Only one error in four (24%) came from something that changed after the article was written.
- 62 of the 68 corrections that needed a source were proven by the official page of the product maker or the public body concerned.
- On 7 pages from 6 other sites (travel, marketing, software support, fintech, training, events), old pages were wrong because the world changed, recent ones because of claims that were wrong or overstated from the start.
- The free scan, tested by us on 3 articles from 2013 on another tech site, found a tool unsupported since 2017 and a registry step that would delete too much. For the first real user, on the guide to its own software: 3 corrections, all proven by the publisher's own pages.
How we did it
Our own articles. 23 French-language tutorials and guides from FunInformatique: Windows, iPhone, Android, Linux, Facebook accounts, security, phone scams and AI tools. We chose them to test the engine, so they are not a random sample of the site. Each was analyzed between October 6 and October 29, 2026, and we kept the last analysis of each one.
Other sites. On September 27 and 28, 2026, a pilot account ran the base engine twice on the same 7 pages from 6 sites: two Spanish-language travel guides written in 2012, an SEO agency article and an Outlook repair guide both published in 2026, a fintech blog post, a training provider's page and a tech event page. We do not name these sites: the point is the kind of error, not the publisher.
For each article, the engine lists the claims a reader relies on, searches the web for current sources, and only publishes a correction when a source proves it. On our articles, a second check removed 15 proposed corrections before they reached the report, because the proof was too weak or the correction went beyond what the article said. We then read every correction and sorted it by hand into four families.
Finding 1: only one error in four was about time
When people talk about outdated content, they think of old prices and old version numbers. In our 23 articles, those were the minority:
The lesson for anyone auditing a site: sorting articles by date finds only a quarter of the problems. Three errors out of four were in the text from day one, or came from a sentence that promised more than the facts allow. A recent article is not a checked article.
Finding 2: age changes the kind of error, not whether there is one
The pages from other sites made this visible, because their ages were so different.
- The two travel guides from 2012 had 6 corrections, 4 of them about things that had changed: the 9/11 Museum, described as under construction, has been open since May 2014; the visitor pass the guide explained how to book is no longer needed for the memorial plaza; Wikipedia's book tool no longer downloads a whole collection as a PDF. A fifth correction had nothing to do with age: the trees on the plaza are swamp white oaks, not California white oaks, as the memorial's own site says.
- The two articles from 2026, on SEO and on repairing Outlook files, also had 6 corrections, but 4 of them were sentences stated too absolutely: Search Console "tells you the exact queries" people type (Google withholds some of them for privacy), paid advertising "stops producing results the moment the budget's paused", and an FAQ answered "no" to whether data can be lost once a file is repaired. Only one was about change: the Outlook steps applied to classic Outlook, not to the new one.
The other run on the same pages gave the same picture: its 3 corrections on the 2012 guides were all about change, while 4 of its 5 corrections on the 2026 articles were errors present from day one, such as Windows file paths that skipped the user name, or a 2012 close rate compared with a figure from another kind of measure.
So an old page mostly needs its facts brought up to date, and a new page mostly needs its claims and promises checked. Both need reading.
Finding 3: the risky errors were not the old ones
The engine marked 14 corrections in our articles as high severity, meaning a reader following the article could lose money, data or security. Only 4 of them came from a change over time; 5 were sentences stated too absolutely, 4 were wrong from the start and 1 was a step. Some examples:
- Money: a 30-day Acrobat Pro trial that Adobe now lists as 7 days. A reader counting on 30 days would be billed after 7.
- Security: Flash Player in a list of software to keep up to date. Adobe no longer updates it; the right advice is to uninstall it.
- Data: an article on forensic software told readers to install it on the computer they want to examine. The publisher explains how to run it from a USB drive instead, so as not to overwrite the data you are trying to recover.
- Privacy: an article on finding the owner of a phone number still suggested searching Facebook by number, a feature Facebook turned off in 2018.
If you only have time to check part of your site, start with the sentences that give instructions involving money, security, health or the law, whatever their date.
Finding 4: absolute words are a warning sign
Almost a third of the corrections in our articles, and most of those in the recent pages from other sites, were about sentences that were not exactly false, but stated as a rule what the source presents as a possibility. The words were familiar: always, never, any computer, exact, risk-free, permanently, automatically.
- "The USB version runs on any computer": it needs a compatible version of Windows.
- "These tools remove preinstalled apps permanently": without root access, they only uninstall them for the user.
- "You can locate any iPhone from iCloud": only if Find My was turned on before the phone was lost.
- "A password must have at least 12 characters and special characters": the CNIL, the French data protection authority, says there is no universal definition of a good password.
This kind of error is easy to find yourself. Search your articles for those words and check each sentence against the official documentation. You will often only need to add a condition: "with a compatible Windows version", "if Find My was turned on".
Finding 5: the proof is almost always the official page
Of the 68 corrections in our articles that needed a source (the other two were about dates written in the text), 62 were proven by the maker of the product or the public body concerned: Apple, Google and Microsoft support pages, Adobe, Facebook's help center, the French telecom regulator Arcep, the CNIL, the Debian and Linux Mint projects, or the developer's own page.
That matches what we see when we correct by hand: the answer to "is this still true?" is rarely in another blog post, which may be just as outdated. It is in the documentation of whoever controls the thing you describe.
What a careful check does not do
Some results from the pilot mattered as much as the corrections.
- It does not rewrite history. The 2012 travel guides tell the story of a trip made in 2012. The reports kept that story as it was, and only corrected the practical instructions a traveler would follow today, such as booking a pass that is no longer needed. They also refused to say that EPUB export had disappeared, because no source proved it.
- It does not invent problems. The fintech post, published in June 2026, was judged still accurate: good to keep, no correction. In the other run, when the sources found could not settle its claims about payment habits in several countries, the report listed what remained unproven instead of guessing.
- It does not analyze what it cannot read. The tech event page returned an error: nothing was analyzed and no verdict was invented.
A check that can say "this is still right" or "I could not prove it" is more useful than one that always finds something. On the training provider's page, the only correction was not a fact at all: a paragraph announcing another company's IT services had been pasted into a page about Scrum.
The free scan: our own test, then a first real user
Before sending the free scan to anyone, we ran it ourselves from an ordinary account, on four articles: one of ours, counted above, and three tutorials published in 2013 on another French-language tech site, which we do not name either. On those three, the reports made 8 corrections, including:
- a cleaning tool still presented as the one to use, when its download page says it has not been supported since October 2017;
- update and restore steps for a driver tool that use buttons and tabs its publisher no longer describes;
- a registry tip that has the reader create a value, then says to delete "the key" to undo it, which would remove the other settings stored in that key;
- the same tip presented as protection against attacks, when the security guide describing this setting says it only hides the computer from the network list, and attackers have other ways to find it.
On these 2013 pages, as on the 2012 travel guides, the first thing to fix was what had changed: a tool that is no longer maintained, an interface that no longer looks the same. We have not yet reviewed these 8 corrections one by one, so they are not counted among the confirmed errors below.
Then, on November 2, 2026, a software publisher who had signed up after one of our emails ran a free analysis on the getting-started guide for its own product, published on its own site. We do not name it: the result belongs to the publisher.
The report found three errors, and all three were proven by the publisher's own pages: a feature presented as reserved for paying customers, when the free version includes it; features attributed to the paid version in general, when they require two specific higher plans; and a misspelled product name. Nothing else was flagged.
This case sums up what we saw everywhere. Nobody knows a product better than the company that makes it, and yet its own guide did not match its own pricing and documentation pages. A reader following the guide could have paid for a plan that does not include what they need, or paid for something that is free. The proof was one click away, on the same site.
What we did with the results
The first article we fixed was our guide to ping call phone scams. The report found three errors, which we confirmed: the number that calls you is not always premium-rate, French special numbers are not all premium-rate (Arcep separates free, standard-rate and premium-rate numbers), and +269 is the Comoros, not Mayotte. We corrected the article the same night, and on October 7, 2026 the analysis of the new text came back good to keep. Three weeks later, a newer version of the engine gave the same verdict, as it did for our Gmail backup guide after its corrections.
That is the loop we now use on FunInformatique: analyze, check the proof, correct, analyze again.
The limits of this test
- These are small samples, chosen by us: 23 articles from one French-language tech site, 7 pages from other sites in the pilot, and 4 free scans. Another site, topic or language may show a different mix.
- The engine does not check everything. In our articles, it checked 205 of the 313 important claims (65%). An article without a correction on a point is not proof that the point is right.
- It can miss errors. We confirmed 7 errors by hand in three of our articles; the reports found 5. The two it missed were definitions in our malware article. On two of the pages from other sites, the base engine found 4 of the 7 errors we validated later, across its two runs. This is why the engine kept changing during October (versions 7.6 to 7.18), and why we keep every case to test each new version.
- The pilot ran on the base engine, through several successive settings; our articles were analyzed with the current version built on it. The free scans used the current version too. The series are reported separately and not added together.
This is also why EEATClean never edits a site. It gives you the claim, the source and the proposed correction. You decide.
How to apply this to your own site
- Don't stop at the date. Use dates and Search Console to choose which pages to check first, then read what the page says. Our guide on how to find outdated content explains the full method.
- Adapt the check to the page's age: on old pages, look for what has changed; on recent pages, look for promises the sources do not make.
- Start with what can cost the reader: prices, trials, security settings, legal steps, anything that deletes or installs something.
- Search for absolute words and add the conditions the source gives.
- Replay every procedure against the vendor's own help page, step by step.
- Correct rather than delete. None of our 23 articles needed deleting. See what to do with an obsolete article for the cases where deletion is the right choice.
Sources
- Our data: the last EEATClean analysis of each of our 23 articles (October 6 to 29, 2026), our own free scan tests on 3 articles from another site (October 28, 2026), the first free scan by a real user (November 2, 2026) and the two pilot runs on the 7 pages from other sites (September 27 and 28, 2026), each correction sorted by hand into the families above.
- Adobe: Adobe Flash Player end of life.
- System76: Download Pop!_OS (COSMIC in 24.04 LTS).
- Facebook: An update on our plans to restrict data access on Facebook (April 2018).
- OSForensics: Running OSForensics from a USB drive.
- Google: How to download your Google data and Search Console performance report (anonymized queries).
- 9/11 Memorial & Museum: About the museum.
- BleepingComputer: Junkware Removal Tool (end of support).
- Arcep: Numéros surtaxés : comprendre les numéros spéciaux et les numéros courts (in French).
- CNIL: Mots de passe : recommandations pour maîtriser sa sécurité (in French).
Run the same check on your articles
EEATClean lists the claims in each article you choose, checks them against current sources and shows you each correction with its proof. The 7-day trial covers your first 3 articles.
Start my free trial
We'd love your feedback