Seventeen sites, one thermometer, and three results that did not fit
The first run of the published readings dataset produced three results that argue against the instrument that produced them. Hacker News at 47. MDN at 29 for entity strength. The site that publishes the federation specification at 74, on 255 words.
Updated
The first seventeen readings in this dataset produced three results that argue against the instrument that produced them, and they are the most useful thing in the set. Hacker News reads 47 degrees. MDN's HTML reference reads 29 for Content and Entity Strength. The site that publishes the specification the rest of this portfolio implements reads 74, on a homepage of 255 words.
I built the thermometer first and measured afterwards, which is the wrong order if you want comfortable results and the right order if you want to find out whether the thing works.
Here is what came back.
The set
Ten properties from this portfolio, six well-known reference sites, and example.com as a deliberate floor. All measured the same afternoon, all on the same signals, all reproducible: every reading is published with its date and you can retake any of them.
The portfolio sites clustered between 74 and 91. The reference sites spread from 47 to 81. That spread is the first thing worth explaining, because on the face of it a set of small properties nobody has heard of just outscored W3.org, GOV.UK and Wikipedia.
Result one: Hacker News reads 47
Search Visibility 33. Content and Entity Strength 29. Overall 47 degrees, Warming Up, which puts one of the most linked-to pages in technology two degrees below the midpoint of a scale I designed.
This is close to a controlled experiment, and it did not need designing because Hacker News ran it for eighteen years. The site deliberately does almost nothing the thermometer rewards. No meta description. No structured data. No Open Graph. Minimal markup by policy rather than by neglect.
And it is enormously visible anyway, on the strength of exactly one thing this instrument cannot see: being the thing everyone links to.
The wrong conclusion is that the signals do not matter. The right one is narrower and more useful. Observable page signals are a substitute for authority you do not have. Hacker News does not need them. Your site, probably, is not Hacker News.
Result two: MDN scores 29 on entity strength
1831 words. An 86 millisecond response, the second fastest in the set. Search Visibility 83. And Content and Entity Strength at 29, the lowest number anywhere in the dataset apart from example.com's zero.
That combination looks impossible until you ask what a reference page is for. The MDN HTML element reference is not dated, because it is continuously current. It does not name an author in a machine-readable way, because it has hundreds. It does not cite outside sources, because it is the outside source.
Every one of those is a deliberate, correct editorial decision, and the entity reading charges 20 percent of the temperature for them.
I have left the weighting alone. Not because the reading is right about MDN, but because it is right about the four thousand sites that look like MDN and are not MDN. A page with no author, no date and no citations is usually exactly what it appears to be. The instrument is correctly calibrated for the common case and wrong about the exception, and I would rather say that out loud than tune the scale until my favourite pages score well.
Result three: the spec authority reads 74
This is the one that stung.
DigitalKarmaWeb.com publishes the federation specification that most of the sites in this portfolio implement, including this one. It is, within this system, the authority. Search Visibility came back at 92, which is right. Content and Entity Strength came back at 45 and AI Discoverability at 59, because the homepage carries 255 words.
The thermometer cannot tell the difference between a short page that is authoritative and a short page that is empty. It sees 255 words either way.
I could argue that the reading is unfair. It would be more honest to say the homepage is genuinely thin and the authority lives on the spec files behind it, which is a real finding about a real site and is now on the record with a date attached.
What the portfolio results actually mean
Ten properties between 74 and 91, several of them Hot, none of them with the traffic of any reference site in the set. That is not a claim that these sites are winning. It is a demonstration of what the scale measures.
These sites were built recently, deliberately, on a stack where structured data, answer-first paragraphs, server rendering and federation endpoints are defaults rather than projects. Of course they read hot on an instrument that measures exactly those things. A scale that rewards the work you did is not evidence the work was worth doing.
The evidence for that is whether the numbers move anything, over years, and that argument is not available yet. It will be, and the dates on these readings are the reason it will.
One result that did fit
QuickRankAI came back at 91, the highest in the set, with AI Discoverability at 73 as its only sub-reading below 80.
That site's entire product is telling people what to fix about their AI readiness. It has something to fix. I find that more reassuring than a clean sweep would have been, because a tool that scores its own family perfectly is usually a tool that was tuned until it did.
What happens next
These seventeen get retaken. Not on a schedule that makes a good story, but often enough that the second column is real, and the ones that move will be more interesting than the ones that do not.
The prediction I am willing to put a date on, 18 September 2026: the portfolio sites will barely move, because they start near the ceiling of what this instrument can see. The reference sites will barely move either, because institutions change slowly. If anything in this set shifts more than five degrees in a year, it will be because someone made a deliberate decision, and that is precisely the thing worth catching.
Filed under: dataset, method, measurement