How to check a meta description against the page

The sentence in your <meta name="description"> is a suggestion. Google rewrites the snippet when the page itself is a better answer.

Checking meta description relevance means comparing that suggestion to the page and to the query you hope to match. Three failures show up constantly: the sentence claims something the page does not say, it answers a different intent, or it could sit on any URL in the template.

Google may ignore the description you wrote

Google builds the snippet from the page. The meta description is a fallback for when it is the clearer summary.

The snippet controls documentation says Google primarily uses on-page content, and may use the meta description when it describes the page better than other parts of the content. Snippets are meant to preview the part of the page that fits the search, so the same URL can show different text for different queries.

There is no length limit on the tag. The visible snippet is cut to the device width. A 320-character description is not “more optimized.” It is just more likely to be chopped.

You can refuse a snippet with nosnippet, cap it with max-snippet, or mark a region with data-nosnippet. Those are controls, not ranking tricks. If the description is a keyword list, Google says that kind of text is less likely to be shown.

Three checks before you keep the sentence

Read the description next to the rendered page, not next to the CMS field.

1. Every concrete claim is on the page. Price, date, author, location, and product name belong in the description only when the page states them. Google’s own example is a book description that gathers facts already on the page, and a shop description that names what is sold, the hours, and the place.
2. The sentence matches the query you are willing to be judged on. If the results for that query are comparisons, a description that only says “welcome to our blog” is the wrong pitch. Intent labels are covered in how to classify a keyword’s search intent.
3. The sentence is useless on the neighboring URL. Identical descriptions across a site are called out as unhelpful. Site-level copy belongs on the homepage. Everything else should name that page.

If you cannot write one for every URL, Google says to start with the homepage and the popular pages. That is a priority rule, not a license to clone one sentence.

Generated descriptions are fine when they stay specific

Hand-written lines do not scale on a catalogue. Google says programmatic descriptions are appropriate there, and encouraged, when they stay human-readable and diverse. The raw material should be page-specific: the product, the city, the date. A template that only swaps the keyword is the bad case they describe, the same blurb on every news article.

Two practical tests before you ship a generator:

• Sample 20 URLs. If you can move a description onto another URL and it still reads as true, it is not specific.
• Search a query you care about and see whether Google used your sentence or pulled a paragraph. If it pulled a paragraph, fix the page summary first. The tag is not the snippet.

A model score is not a snippet guarantee

You can ask a model to score a description against text you supply. That is a review aid. It is not Search.

TypeSafe’s introduction of Jev describes a System One model that evaluates typed questions against state you send. It returns structured answers, not an article. The three question types are Choice (pick one option), Score (a rubric), and Noul (a yes-or-no probability from 0 to 1). Choice and Score also return a confidence value. TypeSafe’s advice is to keep each question atomic and combine the answers in your own code.

Their 15 September 2026 announcement says Jev gives up string generation. Published end-to-end times are 70ms to 500ms. Input is listed at $0.042 per million tokens, with output tokens marked free. Those are the vendor’s figures, measured from their side.

What the docs do not say: Jev does not fetch your URL. If the title, the description, and a page excerpt are not in the state, the score is about the packet you built, not about the live page. A high score can still lose to on-page text, because Google chooses the snippet. Treat a low confidence value as a queue for a person, not as a rewrite.

The same limit applies to any classifier. The page still has to contain the facts. For how those facts get chosen when two URLs compete, see blog posts that target the same query.