Test report · as of 8 September 2026REV. 2026-A
What this website measurably delivers
On 8 September 2026, aureum-tech.com was measured end to end: speed, technology, security, data protection and structure, and, for the first time, what AI assistants say when asked. All 71 pages were covered; the tool measures speed, technology and data protection on the home page. The report runs to 33 pages and can be downloaded below. The most important values are shown here first, so you do not have to open anything to see them.
The measured values
| Largest element visible | 0.4 starget under 2.5 s |
|---|---|
| Server response time | 93 msmedian; 34 ms in the lab |
| Blocked time | 0 mstarget under 200 ms |
| Layout shift | 0,087Target under 0.1, just below it |
| Page weight | 198 kBof which 91 kB fonts |
| Lighthouse SEO | 100 out of 100measures this one page, not the whole website |
|---|---|
| Lighthouse Performance | 98 out of 100100 on 5 September |
| Lighthouse Accessibility | 97 out of 100100 in August, see list of findings |
| Lighthouse Best Practices | 92 out of 10073 on 5 September |
| Cookies set | 0 |
|---|---|
| Third-party servers requested | 0nothing loads from outside |
| Trackers | none found |
| Security headers | 8 out of 8set |
|---|---|
| Mozilla Observatory | A+135 points, 12 out of 12 tests |
| E-mail sender protection | SPF, DKIM, DMARCDMARC set to “reject” |
| HTTPS enforced | one yearwithout subdomains, see below |
| TLS certificate | Google Trust Servicesvalid until 12.10.2026 |
| CAA record and DNSSEC | neither setsee below |
| Blocks of structured data | 72 on 71 pages |
|---|---|
| Machine-readable FAQ questions | 262on 47 pages |
| Text versions for AI systems | llms.txt and llms-full.txtboth present |
| AI bots with access | 23 out of 23none blocked |
| Pages tested | 71that many are listed in the sitemap; the crawl covered 72 |
|---|---|
| Orphan pages | 0every page is linked |
| Internal links | 6.61293 per page on average |
| Pages with only one incoming link | 515 on 5 September |
| Maximum click depth | 2 levels |
| Score | 50 out of 100not part of the overall grade |
|---|---|
| Scope of the measurement | 20 answers5 questions to 4 assistants |
| Question about the company name | 4 out of 4 name usChatGPT, Gemini, Perplexity, Claude |
| Questions about services | 2 out of 16 name usthe weak spot |
SpotLens sums this up as overall grade A, 94 out of 100 points. The report of 5 September 2026 came to 93. Setting the two numbers side by side would still be misleading: this is the first report at the most thorough level. It also checks how four AI assistants answer factual questions, and compares search terms. More ground covered means different numbers, not better ones. Among the sub-scores, discoverability rose from 92 to 94 and content authority from 88 to 92, and security is unchanged at 95. Machine readability fell, from 98 to 95. None of the twenty findings is rated high urgency; ten are medium, ten low. What that means in detail is set out below.
On page weight: the test bar on the home page shows about 150 kB, the report 198 kB. The bar measures what your browser has actually loaded just now; the report measures a first visit without a cache. Both figures are correct, they measure different things. On response time: 34 ms is what the test tool measures in the lab, without a network path; 93 ms is the median over the open internet. The table shows the second value because that is the one you experience.
What is still open
Proof that shows only the good numbers is advertising. That is why all 20 findings of the report are listed here, not a selection: ten medium, ten low, none above that. Below them are the points from earlier reports. Anything a report has found stays listed and is marked “Fixed”; nothing is deleted here. We have collapsed the completed items at the end so that the open points stay at the top; one click and they are back.
Medium
The address with www redirects only temporarily. www.aureum-tech.com answers with status 302 instead of 301. A 302 tells search engines that the redirect is only temporary, so the weight of incoming links is not reliably passed on to the main address. The redirect from http to https is correctly set up as a 301. So exactly one line is missing. This point has been open since 14 August.
Medium
Our security policy blocks bot detection. The protection layer in front of the website adds a small script to every page, which it uses to tell real browsers from bots. This script carries a different identifier on every request, so it has a different checksum each time. Our policy, however, allows only thirteen fixed checksums and therefore blocks it. The browser console then shows an error message.
Medium
The HTTPS requirement does not apply to subdomains. The header that forbids browsers from ever opening the site unencrypted is set to one year. That is enough for the top grade at SSL Labs. Two additions are missing: one that extends the rule to all subdomains, and one that is a precondition for inclusion in the list built into browsers.
Medium
32 of 34 articles cite no outside source. The articles give hard figures and deadlines: list prices for the Windows 10 extension, cut-off dates for e-invoicing, the end of support for Office 2021. None of it is backed up. The two exceptions point to our own test tool. Anyone who wants to check has to search for themselves, and answer engines prefer texts that trace their figures back to the primary source.
Medium
No page on the specialist software that customers already use. We report this finding only in part, and for a reason: the report’s reasoning relies on statements about other vendors, and we cut those out before publication. The core can still be checked: this website has no page of its own on the industry programs our customers work with every day.
Medium
The Zwickau local page does not rank for its own search term. For “Systemhaus Zwickau” the domain does not appear in the results, although our page on it, at 1,247 words, is the second longest local page. It is also the least crowded of the six search terms tested. So there is ground to be won here.
Medium
The web design page does not name the region. Its title read “Webdesign aus Sachsen” and so named the federal state rather than the area where customers search. It does not appear for “Webdesign Vogtland”. Since 9 September the title names the region, and so does the heading of section 02. Before that, “Vogtland” appeared seven times on the page but not once in a heading. The main heading stays as it is, on purpose. The second part of the finding remains open: at 565 words the page is our shortest service page, and the local pages are one and a half to two times as long.
Low
Nine service pages share the same preview image. 49 of the 71 pages now have their own image for the preview in social networks and chat programs; in August it was the same one on every page. That leaves 21 subpages with the general image, including nine service pages.
Low
The main heading of the home page names neither service nor place. It reads “IT, die läuft. Nachweislich.” (IT that works. Provably.) That is a brand line, and the heading is also the strongest text signal on the page. What we do and where only appears below it. This point has been open since 14 August and is a deliberate decision; it stays in the list all the same.
Low
Sender protection ends with a soft fail, although the rule above it is strict. Our SPF record ends with a character that roughly says: anyone not on the list is suspicious, but accept the message anyway. The DMARC rule, on the other hand, is already at the strictest level. The two contradict each other. In practice DMARC catches most of it.
Low
No CAA record: any certificate authority may issue. A CAA record sets which certificate authorities may issue a certificate for a domain at all. We have none. That means any public certificate authority in the world may do so on request. There is no sign that this has happened, but the safeguard is missing.
Low
DNSSEC is not enabled. DNSSEC signs the answers from the Domain Name System so that they cannot be swapped unnoticed in transit. For our domain both components are missing. Someone positioned along the way could point the domain to a different address.
Low
The DMARC reports go to a personal mailbox. Our DMARC rule is at the strictest level. The reports on who tries to send in our name go to the mailbox of one individual. That has two drawbacks: if that person is on holiday, nobody sees the reports, and the reports belong to the company, not to one person.
Low
Page head and server responses reveal version and platform. The head of every page contained the exact version of our build tool, and three headers name the hosting platform. None of this is a vulnerability. It is a shortcut for someone who searches specifically for websites with a particular version. The version in the page head was removed on 9 September. The three headers come from the protection layer in front of the website and can only be switched off there, which is why the point stays open.
High
Hardly any links from other websites. The most serious finding, and the only one at this level. 133 domains point to us, but of the confirmed sources, six out of seven are automatic company directories that list every newly registered GmbH once. In the reference database for domain authority we are not listed at all yet. For a domain registered in July 2026 that is the expected start. It is still the biggest lever, and bought links are out of the question.
Our own re-check We checked again: we are in fact listed on 11880, wlw, Cylex and LinkedIn; the report simply did not find those entries. For link strength this still changes almost nothing: only the entry on 11880 sets a link that search engines follow. LinkedIn and Cylex redirect via their own tracking address, and wlw only mentions us in text. For the Yellow Pages (Gelbe Seiten) the report is right: we really are not listed there.
Medium
Ten script errors on two pages. Our provider’s address obfuscation stumbles over links inside the drawings embedded in the page source. There it reads an attribute as text when it is not. The e-mail addresses still appear correctly, so the damage is not something you see. These errors were the reason for the Best Practices score of 73 shown in the report of 5 September; in the Pro report it stands at 92. In August there were six errors on one page, now ten on two: this point has got worse.
Medium
Our AI interface answers the first request with an error. When a page loads, a script from our provider asks our MCP interface for its tools and gets an error back, because it does not send the session identifier from the previous response. Our server works correctly, the script does not. Visitors notice nothing. But because we offer MCP integrations as a service, we are not leaving it like that.
Medium
“Managed services” is missing as a term. Of the terms searched for in our market, this is the only one that appears on none of our pages. Cloud and Microsoft 365 are further down this list, and we state explicitly that ISO 27001 is not a service we offer. We will create a page of its own when we offer the service as a package, not before.
Medium
The guide articles are hardly linked to each other. 15 of the 31 articles have at most three incoming links, two of them only a single one from the overview. The service pages get 67 each, because they appear in the navigation and footer. So it is the strongest pages in terms of content that hang by the thinnest thread.
Medium
No customer reviews are visible. The references page names three projects, but nobody who says anything about them. A competitor in the comparison shows 4.4 stars from 14 Google reviews directly on the page. So far we do not collect any.
Medium
For styles our security policy is too lax. For scripts it works with 13 checksums and allows nothing else. For style declarations it allows embedded CSS across the board. That leaves an attack surface open that is closed for scripts.
Medium
Accessibility has fallen from 100 to 97. The cause is tap targets on phones: some operable elements sit too close together or are smaller than 24 by 24 pixels. In August this was still at full marks. This is the second value that has got worse.
Low
The publishing rhythm is uneven. 32 articles between 18 July and 28 August, produced in batches of two to five. Two to four articles a month at a regular rhythm show ongoing upkeep better than the next batch.
Low
Three kinds of structured-data markup are missing. One for how-to guides, so that step-by-step results can appear in search. One for a search box directly in the Google result. And one for reviews, which only makes sense once there are public reviews; see further up.
Low
65 of 68 pages contain no embedded image. Our drawings sit as SVG directly in the page source. They therefore load without an extra request and stay sharp at any resolution. The price: we practically do not appear in image search, and when shared, all articles fall back to the same preview image.
Low
Three small things in the server configuration. The HTTPS requirement applies for one year, but without subdomains and without an entry in the browsers’ preload list. DNS lacks the record that sets which authority may issue certificates for our domain at all. And three response headers reveal our hosting platform along with the region.
Low
Two values are missing, at no real cost. Field data from Google’s usage report does not exist for our domain yet; Google only collects it once a site reaches a certain number of visitors. And the markup for voice assistants is missing, which Google so far evaluates mainly for news content. Both are listed here only for completeness.
12 points that have since been fixed
Fixed
The sitemap reported modification dates that did not exist. All 71 addresses carried a timestamp, but there were only eight different values; 59 addresses alone shared just two timestamps. For 33 of 35 pages with a date of their own, the value in the sitemap was newer than the one on the page. Search engines use it to decide when to come back.
Since 8 September every timestamp comes from the last commit that changed the page in substance; pages with their own date carry that one. Instead of eight different values there are now seventeen.
Fixed
Our AI interface said “es antwortet undefined”. Our own server for AI assistants delivered a placeholder in the company data tool that was never filled. What went out was, word for word, “es antwortet undefined” (“it answers undefined”), in the middle of a sentence about how to reach us. Anyone who asked about the company through an assistant got that sentence, and quite possibly it was quoted that way. At the very point where our presence is otherwise strong.
Fixed on the evening of 8 September. The sentence now gives the hours and nothing else; the addition also happened to break our own writing rules.
Fixed
In 25 of 34 articles the author stays nameless when only a machine reads. In the author field every article points to an identifier on our About us page. A data record with the name is not in the article itself, though. Only the nine articles by Felix Götz resolve, because he is listed as founder in the company record anyway. For people the name is visible; for tools that do not combine data across documents, it is not.
Fixed on 9 September, in one place instead of in 34 articles: the company record now lists the three people by name, with exactly the identifiers the articles point to. Re-measured on all 34 articles: in every one the author now resolves within the document.
Fixed
The Twitter cards were only half filled in. The card format entry was on all 71 pages, but title, description and image were on none. The platforms then fall back on the Open Graph entries. That works, but nothing guarantees it.
Added on 8 September. All four entries are now on every page; you can see them in the source of this page.
Fixed
Three lines of the text version for AI systems wrote umlauts as letter pairs. In llms.txt, three places had words in which ä, ö and ü were replaced by ae, oe and ue. The summaries on used hardware, warranty and scam calls were affected. Language models read this file as a short version of the website.
Fixed on 8 September, and the build now checks it itself: a rule in our pre-release check stops the build if a transliteration turns up again.
Fixed
34 author links land at the top of the page instead of at the person. Every author line points to an anchor on the About us page that does not exist there. Anyone who clicks the name lands at the top of the page and has to look for the person themselves.
Fixed on 9 September. The three team cards now carry the identifiers the articles point to, and a margin so that the card does not disappear under the header bar. You can check it by clicking an author name under a guide.
Fixed
Almost all meta descriptions are too wide. 65 of 68 exceed the width that Google displays. The character count hides this, because it says nothing about the width of the individual letters. On 43 pages our phone number is in the description, on 42 of them at the end, so exactly in the part that gets cut off.
The Pro report of 8 September states it explicitly: not a single one of the 71 pages is over the display width any more. The widest measures 915 pixels.
Fixed
Only five AI bots are named in robots.txt. The other eleven of the 16 tested fall back on the catch-all rule. They are allowed to crawl, but the decision is not written down anywhere. Spelling out the list makes it verifiable.
Now 23 AI bots are named in it, and none is blocked. You can see it in our robots.txt.
Fixed
The contact form relied on your e-mail program. It opened your e-mail program instead of receiving the message itself. If you had none set up, for example on a phone or with web mail, your enquiry did not arrive, and you did not notice. In the August report this was the most serious point of all.
Since 19 August 2026 our own server receives the form and sends the message itself. Without JavaScript the mail program still opens, but as a fallback, not as the only way.
Fixed
All guides carried the same modification date. In the sitemap the date jumped to the current day with every publication, even for articles that had not changed. For search engines the entry lost its value.
The September report finds the modification dates spread over eight different days. A script now reads each page’s last commit date instead of giving all of them the same stamp.
Fixed
Eight page titles were too long. They exceeded 60 characters and were cut off in the search result, plus two descriptions that were too long.
The Pro report measures titles in pixel width instead of characters: none of the 71 exceeds the cut-off limit of about 580 pixels; the widest measures 575. The same applies to the descriptions, see above in the list.
Fixed
The long text version for AI systems was missing. The short overview at llms.txt was well filled and linked in the page head; the complementary full-text version was missing.
llms-full.txt has been available since 29 August 2026 and is regenerated from the finished pages on every build.
Why we publish this
Because we build the same for customers. A website that loads quickly, contacts no third-party servers without consent and has its security headers set is no accident but craftsmanship; you can read more under Web design & development. The free website check shows how your own website measures up in the same areas.