Are meta descriptions a ranking factor?

No. Google has said meta descriptions are not used for ranking. They can influence click-through rate when Google shows them as the snippet, which is why they are still worth writing for important pages.

How long should a meta description be?

There is no fixed limit, but descriptions of roughly 140 to 160 characters usually display without being cut off on desktop. Put the most important information first, since mobile snippets can be shorter.

Does Google rewrite meta descriptions?

Often. Google may show different text from the page when it better matches the searcher's query. Well-written, page-specific descriptions are more likely to be used, but you cannot force Google to show them.

Is it OK to use AI-written meta descriptions?

Yes, as long as they are accurate, specific to each page and reviewed. Problems come from generic, duplicated or invented copy, not from the use of AI itself.

Should every page have a unique meta description?

Every important indexable page should. For very large sites, it is acceptable to leave low-value pages without one and let Google generate snippets, rather than filling them with duplicated boilerplate.

Writing Meta Descriptions With AI at Scale (Without Duplicate or Thin Copy)

Meta descriptions are the perfect job for AI on paper. They are short, repetitive and needed on thousands of pages. In practice, AI-generated descriptions go wrong in familiar ways: generic copy repeated across hundreds of products, invented claims, prices that do not match the page, and descriptions that read like they were written for nobody.

This guide shows how to generate descriptions at scale that are accurate, specific and worth showing, with a workflow that keeps a human in control.

First, decide where descriptions matter

Meta descriptions are not a ranking factor; Google has said so directly. Their value is in the snippet: a good description can make a result more appealing and lift click-through rate. But Google often rewrites descriptions to match the query, so effort should go where a description is most likely to be shown and to matter.

Prioritise:

Pages with high impressions and below-average CTR for their position, from Search Console.
Commercial pages: product, category, service and location pages, where clicks lead to revenue.
Pages with missing, duplicated or truncated descriptions.

Deprioritise low-traffic archive and utility pages. On very large sites, it is better to leave those without a description than to fill them with duplicated boilerplate.

Remember that on queries where AI answers appear, fewer people click any result. A better description helps, but it will not recover all the lost clicks. See how much traffic AI Overviews really take.

The rule that prevents most problems: ground every description in the page

The biggest source of bad AI descriptions is the model filling gaps from its general knowledge. The fix is simple to state: the model may only use facts that appear in the page's visible content.

In practice:

Extract the page's title, main heading, key paragraphs and structured facts such as price, availability, location or duration.
Give only that extracted content to the model.
Instruct it not to add any fact that is not in the content.
Validate afterwards that numbers and names in the description appear on the page.

If the page content is too thin to describe accurately, the right output is a flag, not a description.

Templates by page type

Different page types need different descriptions. Give the model a short pattern per type.

Page type	What the description should include	Example pattern
Product	Product name, key attribute, a differentiator, a practical detail	"[Product]: [key attribute]. [Differentiator]. [Delivery or stock detail]."
Category	What the range covers, variety, a reason to browse	"Shop [category]: [range summary]. [Selection or price range]. [Delivery detail]."
Service	Service, who it is for, location, the next step	"[Service] for [audience] in [location]. [Proof or differentiator]. [How to book]."
Location	Business, place, what is offered there, practical detail	"[Business] in [area]: [services]. [Hours or parking]. [Booking detail]."
Article	The question answered and what the reader learns	"[Answer in brief]. Learn [what the article covers]."

Patterns guide structure without forcing identical wording. Ask the model to vary phrasing so descriptions do not read as templated.

A prompt that works

A good prompt combines the rules, the page type pattern and the extracted content:

Write a meta description for this [page type] page. Use only the facts in the content provided; do not add prices, claims or details that are not in it. Aim for 140–160 characters and put the most important information in the first 120. Include [the product or service name]. Write for a person deciding whether to click, in plain language. Avoid the phrases on the banned list. If the content is too thin to write an accurate description, reply "FLAG: thin content" instead.

Add the client's tone guidelines and banned-phrase list to every prompt.

Validate automatically before review

Automated checks remove obvious failures before a person looks at anything.

Length: within the agreed range.
Required terms: the product, service or location name.
Banned phrases: such as "best in the world", "guaranteed", "cheapest", medical or financial promises, and the client's own list.
Fact check: every number, price and proper noun in the description appears in the page content.
Duplicates: no description repeated across pages on the same site, and none too similar to another.
Flags: thin-content flags routed to a separate list for the client or content team.

Descriptions that fail are sent back for redrafting with the reason, not shown to reviewers.

Review by sampling, then approve in batches

Reviewing thousands of descriptions one by one defeats the purpose. Use sampling:

Group drafts by page type and client.
Review a random sample of 20 to 50 from each group.
If the sample is good, approve the group. If not, reject the group, fix the prompt or rules, and redraft.
Always review individually for the highest-value pages, such as top-traffic product and service pages.

Every approved change should be logged and reversible. See human-in-the-loop SEO automation for designing the approval step.

Measure the result

After publishing, measure whether the new descriptions helped.

Compare CTR for updated pages before and after, for queries where average position stayed similar.
Check how often Google shows your description rather than rewriting it, by sampling SERPs for key queries.
Track the share of drafts approved without edits, which shows whether the prompt is improving.

Allow a few weeks for Google to recrawl and for data to accumulate.

Multilingual and regional sites

Sites in several languages or markets need extra care, because translation alone rarely produces good descriptions.

Draft in each language from that language's page content, rather than translating English descriptions. The page itself may differ by market, with different prices, products or delivery terms.
Use local currency, units and spelling as they appear on the page.
Give the model market-specific banned phrases. Advertising rules on claims differ between countries, especially for health, finance and food.
Review samples with a native speaker for each language before batch approval. Fluent-looking text can still be unnatural or wrong in tone.
Check hreflang groups for duplicates across markets that share a language, such as UK and US English, and make descriptions market-specific where prices or delivery terms differ.

Common mistakes

Letting the model use general knowledge, which produces invented details.
One template for every page type, producing generic descriptions.
Descriptions that only repeat the title.
Keyword stuffing. Descriptions are for people deciding whether to click.
Filling every low-value page with near-duplicate copy.
Publishing without review or a rollback.

Beyond meta descriptions

The same workflow applies to other repetitive on-page work: title tags, image alt text, internal link anchor suggestions, and FAQ answers drawn from existing content. In each case, ground the output in page content, validate automatically, sample-review and approve with logging. For more on what agents can safely take on, see what agentic SEO can and can't safely do, and for practical prompts, see Claude for SEO.

A worked example

A fictional outdoor equipment retailer has 4,800 product pages, 2,100 of them with missing or duplicate descriptions. The team extracts each page's name, key specifications, price and delivery note, and uses product-type patterns for tents, footwear and clothing. Drafting produces 2,100 descriptions; validation returns 96 where a price in the draft did not match the page and flags 42 pages as thin. A specialist reviews 40 samples per product type. Footwear samples are too generic, so the prompt is changed to require the activity and terrain from the specifications, and the footwear batch is redrafted. All batches are approved within three days, and the 42 thin pages go to the merchandising team. Six weeks later, the updated pages with stable positions show a modest CTR improvement on commercial queries. The numbers are illustrative; the pattern of grounding, validation, sampling and measurement is what matters.

How Rankwyn helps

Rankwyn drafts meta descriptions from each page's visible content, runs length, duplicate, banned-phrase and fact checks automatically, and puts drafts into the approval queue grouped by page type, so you can sample and approve in batches with a full log and rollback across every client site. See the Rankwyn MCP server.

Key takeaways

Meta descriptions do not affect rankings and Google often rewrites them, so focus on high-impression, low-CTR and commercial pages.
Ground every description in the page's own visible content; flag thin pages instead of inventing copy.
Use short patterns per page type and vary phrasing to avoid templated text.
Validate length, required terms, banned phrases, facts and duplicates automatically before review.
Review samples, approve in batches, review top pages individually, and measure CTR afterwards.

The bottom line

AI can write thousands of good meta descriptions, as long as it writes only what the page supports and a person stays in control of what goes live. Prioritise the pages where snippets matter, ground every draft in page content, validate automatically, sample before approving, and measure the effect. Done that way, a week of tedious work becomes an afternoon, without the generic or invented copy that gives AI descriptions a bad name.