Meta Description Relevance With TypeSafe Jev
Check meta description relevance by passing the query, the description, and the page content to TypeSafe AI's Jev. It does not open the URL, and it does not write a replacement sentence.
I ask a Score for meta description relevance on ordered levels, and a Noul for whether the description is about this page. Search intent rides in the state as the query and a one-line note of what the searcher wants. Weak or uncertain cases go to a person. Confidence feeds the routing verdict. Confidence is not proof the snippet is right.
Take a /pricing description that says "our story" while the query is "analytics pricing" and the page content is a price table. Meta description relevance is poor because the sentence misses the page and the search intent. You still decide whether to rewrite it. A model score is a sort key for review, not a publish button.
Page content and search intent go in the state
Meta description relevance is a comparison. TypeSafe AI's Jev needs both sides. I send three fields:
query: the search you are judging, in the words a searcher typed.description: the meta description exactly as it is shipped, not a cleaned-up paraphrase.page: the title, the H1, and the first useful paragraphs. Not the whole HTML document if it is a novel.
Search intent belongs in a short note next to the query when the words alone are ambiguous. "analytics pricing" can mean a vendor's price page or a guide to how pricing works. Say which one you mean. If you hide that, the Score will average over both readings and confidence will sag.
Fetch the page yourself. If the description in state is not the description in the HTML, you are grading a draft that is not live.
Use a Score for meta description relevance
One Score, one quality. I do not ask a single question to grade grammar, length, and search intent at once. Length is a character count. Code can do that before the call and skip the model when the description is empty.
Levels for meta description relevance, each understandable on its own:
- The description is about a different page, product, or offer than the page content.
- The description names the topic and misses the search intent of the query.
- The description matches the page content and the search intent, with one concrete detail.
- The description matches the page content and the search intent, and it states an outcome the page actually offers.
The answer includes a score on that scale, a probability per level, and confidence, as Score specifies. A score of 1.4 with probability split between level 1 and level 2 is a fuzzy miss, not a precise grade. Read the distribution when you care why.
A second Score is fair when you truly have a second quality, such as whether the description repeats the title with zero new information. Ask it in the same request. Do not average the two Scores in your head and call that meta description relevance.
A Noul checks the page, and a Choice is the routing verdict
The Noul is a yes-or-no probability: does this description describe the same offer as the page content? The answer field is noul, from 0 to 1. Near 0.5 means the two outcomes are both plausible. It does not mean "medium relevance." Noul answers do not include a confidence field. Use the probability itself. Confidence applies to the Score and the Choice, not to the Noul.
The Choice is the routing verdict, and only the routing verdict:
leave: the description can stay.review: a person should read it.rewrite: hand the row to whoever writes the snippet.
Put the policy in code, not inside the Choice instructions. I do not let the Choice mean "Jev decided to publish." TypeSafe AI's Jev has not seen your click-through rate unless you put that number in state, and even then the Choice should not be your only gate.
relevance = answers["relevance"].score / 3
matches = answers["matches_page"].noul
route = answers["route"].choice
unsure = answers["relevance"].confidence < 0.5 or 0.35 <= matches <= 0.65
if relevance < 0.34 or unsure or route == "review":
send_to_review(url)
Those cutoffs are a starting sketch. Calibrate them on descriptions you have already labeled. A cookbook threshold is not your threshold.
Weak or uncertain cases go to review
Weak means the Score sits on the low levels. Uncertain means confidence is flat, or the Noul probability sits in the middle. Both are review, for different reasons. A weak and confident "this is the wrong product" can skip the debate and go to a writer. An uncertain row should not be rewritten by a script that has to invent a sentence.
I keep a queue with the query, the description, the page excerpt, the Score level that carried the probability, and the Noul probability. A reviewer can disagree. When reviewers disagree with the Score on the same pattern every week, fix the level text. Do not keep prompting until the number matches your hunch.
Automatic rewrites fail in a boring way. They produce a description that scores well against the excerpt and says something the page does not offer. Meta description relevance includes that trap. The Noul is there to catch "pretty sentence, wrong page."
Confidence is not a fact check
Confidence collapses the Score's probability spread into one number from 0 to 1. All of the mass on one level is high confidence. A flat spread is low confidence. High confidence does not prove the description matches search intent in the wild. You might have sent the wrong query.
Do not hide low-confidence rows. They are the ones most likely to be a bad level definition or a thin excerpt. Sample them. If the excerpt was only the H1, send more page content and ask again. That second call is justified because you fetched new state.
Grade the snippets you already ship
Pass real HTML, not a remembered sentence. The sibling guide on how to prioritize SEO issues with TypeSafe AI's Jev is the same pattern for a backlog, not for a single description. If you want the visits on the URL after you change the snippet, start free on EventDash. Pricing is the plan split.
Sources:
Related:
Key takeaways
- Check meta description relevance by scoring the snippet against page content and search intent with TypeSafe AI's Jev.
- A Score rates the fit. A Noul asks whether the description is about this page. A Choice is the routing verdict.
- TypeSafe AI's Jev does not open the URL and does not rewrite the description.
- Weak or uncertain cases go to a person. High confidence does not prove the snippet is correct.
FAQ
- Can TypeSafe AI's Jev open a URL and inspect the page?
- No. To check meta description relevance you pass the query, the meta description, and an excerpt of the page content as state. TypeSafe AI's Jev does not open the URL. If that excerpt is missing, the Score has nothing honest to compare and confidence drops. Fetch the HTML in your own code, then send the text.
- Can TypeSafe AI's Jev rewrite a weak meta description?
- TypeSafe AI's Jev does not rewrite a meta description. A Score rates meta description relevance. A Noul says whether the description matches the page content. A Choice can carry the routing verdict: leave it, review it, or hand it to a writer. The new sentence is generative work, and that work stays outside Jev.
- Does high confidence prove a meta description is correct?
- High confidence means the Score probabilities concentrated on one level. High confidence does not prove the meta description is correct, and it does not prove the search intent you sent is the intent a searcher has. Weak or uncertain cases still go to review. Treat confidence as a routing signal, not a fact check against Google.
- Should every weak meta description be changed automatically?
- No. A low Score is a reason to review, not an automatic rewrite. Some weak or uncertain cases are a fair description for a different query than the one you put in state. Change the copy when a person agrees the page content and the search intent are both missed. Leave the rows that are merely short.