Understand

Can a machine tell what each page says, and lift an answer from it?

A page can be reached and still be misread. This stage asks whether a search engine or an assistant can tell what each page is about, and lift an answer from it.

30 of 100 points17 checks, on every page they apply toUpdated October 23, 2026

How does a machine read a page?

It files the page by its title, main heading and summary, cuts it into passages at each heading, and quotes the passage that best answers the question it was asked, often with a link. Search engines and assistants each work their own way, but broadly alike.

1It files the page
From the title, the main heading and the summary, before it reads much else. When the three agree, the page is filed under the right question.
2It cuts the page into passages
Each heading starts a passage. A passage that answers one thing, under a heading that says what, can be taken on its own.
3It quotes the passage that answers
An assistant asked a question picks the passage that answers it best and quotes it, often with a link. It quotes passages, not pages.

What it looks at

Four things, each with an example. The examples are written for this page, not taken from a real site.

What each page is about

The title, the one main heading and the summary are the first things read and the last things forgotten. A title of "Home" and a heading of "Welcome" say nothing a machine can file.

Give every page a title that leads with its subject, one main heading that names it, and a summary of a sentence or two from its own words.

Example

Before

<title>Home</title>
<h1>Welcome!</h1>

After

<title>Roof repair in Denver | Acme Roofing</title>
<h1>Roof repair in Denver</h1>
<meta name="description" content="Leaks, storm damage and worn shingles fixed within a day, with a written price first.">

Answers a machine can lift

An answer buried in the third paragraph is rarely the one quoted. Steps run together in a paragraph, and figures written into sentences, are hard to take whole.

Answer the question in the first sentence under its heading, and word the heading the way people ask. Put steps in a list and figures in a table.

Example

Before

<h2>Our process</h2>
<p>At Acme we believe every roof tells a story, and over the years we have found that most repairs, depending on the damage, take about a day.</p>

After

<h2>How long does a roof repair take?</h2>
<p>Most repairs take one day. Storm damage over a large area takes two to three.</p>

Labelled for machines

Structured data says in a machine’s own terms what a page is: a business, a service, an article, a question and its answer. A label that claims more than the page shows is ignored, or counted against it.

Add JSON-LD that names what the page is, with the facts the page itself shows, and nothing it does not.

Example

Before

<!-- no structured data -->

After

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "RoofingContractor",
  "name": "Acme Roofing",
  "areaServed": "Denver"
}
</script>

Enough to say

A page of a few lines gives a machine nothing to quote, and an image with no description is a blank to it.

Say enough on each page to answer what it is about, and describe each image in a short sentence.

Example

Before

<p>Call us for roofing.</p>
<img src="roof-1.jpg">

After

<p>We repair and replace asphalt, metal and flat roofs across the Denver metro, and most jobs start within a week.</p>
<img src="roof-1.jpg" alt="Shingles lifted by wind on a two-storey house">

How it is scored

Understand is 30 of the score’s 100 points. Each check runs on every page it applies to, and the stage scores the share of those checks that pass, weighted by how much each matters.

Some checks need reading, such as whether a page answers before it explains. A model gives an opinion; code decides. A reply that is unsure, cut off or malformed never passes or fails a page: it goes to a person.

A score says pages are well formed. It does not say an assistant cites the site.

Across 258,015 ranked sites

Median59of 100 on Understand

Full marks0%of sites

0100

The check most often failed: Fix description lengths, missed by 87% of sites.

Understand beside the other stages

Each stage’s points, and how the 258,015 ranked sites do on it.

Stage	What it asks	Points	Median site	Full marks
Access	Can it be fetched and indexed at all?	35	91	10%
Understand	Can a machine tell what each page says, and lift an answer from it?	30	59	0%
Trust	Is there a reason to believe it, and to cite it?	20	98	27%
Act	Can the visitor, or an agent, take the next step?	15	57	28%

Sources

The published guidance this stage draws on. The rules and their weights are mttr’s own.

For an agent

A site’s brief gives its score on each stage, and the fixes in this one in order of what they are worth.

mttr_brief({ site: "example.com" })

See how your site does on Understand

No sign-up. Its score, each stage, and what to fix first.