Most meta descriptions get rewritten by Google because they were never written to describe the page. They were written to hold keywords, a call to action and a bit of brand messaging, which is exactly the material Google’s snippet algorithms are designed to discard.
The fix is not a clever trick. It is a change of job description. A meta description is a directory listing, not an ad slot, and once you write it that way the rewrite rate drops and the snippet starts working for you.
Two sources of ground truth
There are only two documents that actually govern how meta descriptions should be written: Google’s own guidance, which has been broadly consistent since 2007, and the HTML specification maintained by the World Wide Web Consortium. Everything else is convention, and much of the convention is wrong.
The W3C definition is short enough to quote in full:
“The value must be a free-form string that describes the page. The value must be appropriate for use in a directory of pages, e.g. in a search engine. There must not be more than one meta element with its name attribute set to the value description per document.”
Two things follow. First, “free-form string” means there is no formatting or length rule baked into the standard; the only genuine constraint is that it must describe the page. Second, “appropriate for use in a directory of pages” means it has to make sense out of context, in a list of unrelated results, read by someone who has never seen your site. Google’s snippet systems lean on that standards-based reading, so text that diverges from it invites a rewrite.
What Google actually asks for
Three points from Google’s guidance do most of the work.
Snippets should accurately represent the page
Google’s stated aim is for snippets to accurately represent the result. That is not a soft preference, it is the criterion by which your description is kept or replaced. The common SEO habit of loading the description with target keywords fails this test directly.
The usual defence is that matched terms get bolded in results. That argument has been weak for two decades, since pages rank for large numbers of query variations rather than one exact phrase, and weaker still now that people query in natural language. Describe the page. That is the whole instruction.
Differentiate descriptions across pages
No boilerplate, no site-wide template, no “visit our site today” appended to every URL. If two pages share a description, at least one of them is not being described.
Descriptions do not have to be sentences
This is the most underused part of Google’s guidance. Google explicitly says a meta description is a good place to consolidate factual information otherwise scattered across the page, such as price, manufacturer or specification on a product page and author or publication date on an article. It offers a book listing as the model:
“Written by A.N. Author, Illustrated by V. Gogh, Price: $17.99, Length: 784 pages”
The three habits that guarantee a rewrite
These are widely taught as best practice. All three make the description less likely to survive.
- Using it as keyword storage. You can convince yourself the phrases are relevant. Relevance is not the same as telling a user what they will see after the click.
- Writing it to force a click. Heavy-handed persuasion reads as advertising copy, and advertising copy is precisely what a directory listing is not supposed to be. There are subtler ways to make a description inviting.
- Trying to signal search intent. Meta descriptions have nothing to do with intent signalling. They describe content. Intent is handled elsewhere.
The same trap applies to AI-generated descriptions. Using a model to draft them is completely fine, but a prompt that opens with “you are an expert SEO” imports every one of these habits into the output. You have asked for marketing copy and you will get marketing copy.
Borrowing the rules of an actual directory
Since the standard says the description must suit a directory of pages, it helps to look at how a real directory wrote them. The Open Directory Project, better known as DMOZ, was for years the most valuable listing on the web, and its editorial rules translate cleanly into meta description practice. Descriptions there had to:
- Be concise, informative and objective, telling users what they will find when they visit.
- Use logical sentence or phrase structure with proper punctuation and capitalisation, so listings are easy to scan.
- Avoid reading like advertisements, sales pitches, opinions or reviews.
That third rule came with specifics that remain good discipline: avoid advertising superlatives such as best, most, greatest or cheapest; avoid exclamation marks, all caps for emphasis, ampersands and ellipses; and never insert your own review of the page.
Using AI the right way: summarise, do not sell
The hardest part of writing a good description is that you already know too much about the page. Site owners and SEOs describe pages as they wish them to be rather than as they are. An impartial summary is genuinely difficult to produce from the inside.
This is where a language model earns its place, not as a copywriter but as a summariser. Feed it the page text and ask it to determine what the page is about, with no SEO framing at all.
There is a useful side effect. If the model’s summary does not match what you believe the page covers, the page is probably not covering it. That mismatch is a content diagnostic worth acting on, and it connects to the broader discipline of building pages machines can identify and read cleanly.
A five-step prompt for accurate meta descriptions
Upload the page text, then run this sequence:
- Analyse the document and extract the one question that the entire page answers.
- Use that question to create a short, relevant summary of what the document is about, formatted as if it is an answer to a question.
- Format that summary so it functions as a declarative statement when it is out of context to the question.
- Constrain the summary to 120 characters including spaces.
- Lightly edit it so it can serve as a meta description. Use one imperative verb. It must strictly abide by W3C standards for the meta description.
Each step does specific work. Step one surfaces what the model considers the page’s central query, a useful read on topical focus by itself. Step two anchors a summary to that question, step three makes it standalone so it reads correctly where the question is absent, and step four fits it inside the space Google typically renders. Step five enforces the standard. The imperative verb constraint is optional; it adds warmth without tipping into sales language.
What the output looks like
Applied to an article about WordPress developers objecting to a proposed Knowledge Post Type in core, the prompt produced:
“See why WordPress developers oppose merging the new Knowledge Post Type into core.”
Changing step five to request an imperative verb similar to “read” but not that word gives:
“Discover why developers are pushing back against WordPress’s proposed Knowledge Custom Post Type for core.”
Dropping the imperative verb requirement entirely gives:
“WordPress developers are pushing back on a proposed Knowledge Custom Post Type for WordPress core.”
The last version is the most standards-compliant of the three. It states what the page is, nothing more, and it is the one most likely to be shown unaltered.
Controlling the snippet when the description is not used
A page can be relevant to several subtopics, not just its headline topic. When someone searches a narrow subtopic, Google will often build a snippet from the relevant passage instead of your description, and that is the right behaviour. A tailored passage serves that searcher better than a general summary.
You can still influence it. Work through each major section of the page, identify the question that section answers, and open the section with a direct answer to it. That gives Google clean, quotable material to lift and puts snippet control back in the content itself rather than in a single tag. It is the same principle behind writing for click worthiness when the answer is shown without a click, and it applies just as well to structuring content updates.
Frequently asked questions
Do meta descriptions affect rankings?
No. Google does not use them as a ranking factor and does not require them. Their value is in influencing what appears in the snippet, which affects whether a ranked result actually gets clicked.
How long should a meta description be?
The W3C standard sets no limit. In practice, around 120 characters keeps the text inside the space Google typically renders, which raises the odds of the description being used as written rather than truncated or replaced.
Should the target keyword appear in the meta description?
Only if it appears naturally in an accurate description of the page. Inserting keywords for their own sake is the most common reason descriptions get rewritten, and bolded term matching is not a meaningful benefit when pages rank for many query variations.
Is it worth writing descriptions if Google rewrites them anyway?
Yes. The goal is not a zero rewrite rate. It is to maximise the chance that the strongest, most accurate summary of the page reaches the searcher. Descriptions written to standard are rewritten far less often than descriptions written as ad copy.
The short version
Meta descriptions are not dead, they are misused. Treat the tag as a directory entry: concise, factual, objective, specific to the page, free of superlatives and sales language, and short enough to display in full. Use AI to summarise rather than to sell. Then handle the subtopic snippets from the page content itself by answering the section’s question in the section’s opening line.