Updated
Your meta tags now feed three different surfaces that treat them three different ways, and the gap between those behaviors is where most meta tag advice falls down. On a classic results page Google throws most of your work away. Inside AI Overviews and the generative answers people now get without a results page, the same words get read and often quoted back word for word.
Start with what Google does to a blue-link listing. Portent looked at 30,000 keywords and found Google rewrote the meta description 71% of the time on mobile and 68% on desktop, Ahrefs ran the same question across 20,000 keywords and landed around 63%, and titles get the same treatment, with Zyppy finding Google rewrote 61.6% of more than 80,000 titles at least in part. For something like two pages out of three, the words you wrote are not the words a searcher reads.
That statistic is why the trade concluded meta descriptions do little. Answer engines changed the math, because they do the opposite of what Google’s results page does.
Does Google use the meta description you write?
Roughly a third of the time. The rest of the time Google assembles its own snippet out of the page, because it is answering the query in front of it and your description was written before the query existed, so a page that ranks for a dozen different searches gets a dozen different snippets.
The length advice is where this gets interesting, because Ahrefs found Google rewrote 61.46% of the descriptions that ran too long and 63.69% of the ones that stayed inside the limit, and those two numbers are close enough that trimming to fit is not what earns you the snippet. Write to a sensible length so a kept description reads well, then stop optimizing the character count and put the effort somewhere it pays.
Why do answer engines treat your description differently?
Because they are not building a snippet, they are looking for a clean summary of the page they can hand to a person, and your description is the shortest one available. Meta descriptions get quoted verbatim by tools like ChatGPT, Perplexity, and Bing Chat, and Google’s AI Overview citation cards show the page title alongside a short description, and frequently that text is your own meta description carried straight through.
So the description Google discards on a results page is the description an answer engine repeats to your buyer. That inverts the old conclusion, because the description does more work now and it does that work in the surface where you have the least visibility and the fewest analytics, which is a reason to write it yourself, because a plugin that generates it from your opening paragraph will be quoted in that state.
What makes Google keep your title?
Alignment, more than anything else in the data. Zyppy found that matching the title to the page H1 dropped the odds of a rewrite from 61.6% down to 20.6%, which is the largest effect in any of these studies, and titles at the extremes get rewritten more than 95% of the time whether they run far too long or far too short.
So the title that survives says the same thing as the headline on the page, sits in a normal length range, and describes the page without padding. If your title promises one thing and your H1 announces another, Google treats the mismatch as a reason to write its own, and it will use your H1 to do it. On a WordPress blog the post title becomes both the title tag and the H1 by default, which means you get that alignment free, and you lose it the moment someone hand-edits one and forgets the other.
How does your title decide whether AI cites you?
It works as a classification signal before anything else happens. Generative engines read the title, the description, the headings, and your structured data to sort out what a page covers and how confidently they can use it, so metadata behaves as a machine-readability layer where a page gets sorted before it ever gets chosen. A page whose title states the question it answers is easier to select than one built around a keyword phrase nobody says out loud.
The engines differ in how strict they are about it. Perplexity leans toward pages whose titles, headings, or metadata mirror the wording of the query, while ChatGPT is comfortable with a partial or semantic match, which means consistency across your title, headings, and body beats cleverness when you are writing for both. A title that is witty in isolation and disconnected from the H1 loses twice, once to Google’s rewriter and once to an engine that could not classify the page.
How do you write a description that earns the click?
Match the search intent first, so the description signals that this page covers the thing the person typed, and then promise something specific, because “Learn about SEO” tells a searcher nothing while “a step by step process to fix declining organic traffic” tells them what they get. Look at the descriptions already ranking for the query you want and write something that does not blend into them, since being the fourth result that says the same sentence as the first three is how a good page goes unclicked.
Front-load the part that carries the value. A truncated description still has to make sense, and the way to survive truncation is to say the important thing before the cut. That front-loading does double duty now, because a quoted description gets read as a standalone sentence with no results page around it to supply context. The same discipline applies to the title, and copywriting techniques that lift click-through rate from Google results go further into how that persuasion works, while using numbers and lists to stand out in search results covers the formatting side of the same problem.
Which mistakes cost you clicks and citations?
Auto-generated descriptions that pull the first paragraph of the post, which almost never read as a standalone pitch and now get repeated by an engine in that state. Duplicate descriptions repeated across a set of pages, which give neither Google nor a model any reason to prefer one of them. Keyword stuffing and the other old SEO tricks that backfire, which read like spam to a person and get rewritten anyway. A brand name eating a third of the character limit on every page. A title that promises more than the page delivers, which buys one click and teaches the reader to skip you next time. And a title written in language your buyer never uses, which survives on a results page they can scan but fails the engine trying to match the question they typed.
Where do AIO, AEO, and GEO fit into this?
They describe the three jobs your metadata now does beyond ranking. AIO covers how models learn what your business is from your public content, so a title and description that state what a page is about teach the model the right thing. AEO covers direct answers, which is why question-shaped headings and a description that answers the question outright get picked up. GEO covers how generative systems assemble a description of your business out of several sources, where consistent phrasing across your pages stops the model from stitching together a muddled version of you.
None of that replaces the classic work, it sits on top of it, and what AEO means and how it differs from SEO goes deeper on the distinction. Meta tags happen to be the cheapest place to serve all of it at once, since one clear title and one clear description feed the results page, the AI Overview card, and the model deciding whether to quote you.
Which pages do you fix first?
The ones already collecting impressions and few clicks, because those pages have proven Google will show them and the snippet is the part that is failing. Pull your queries in Google Search Console, sort by impressions, and look for the rows where click-through rate sits well below the pages around them, and that is your list in priority order. That pass is item 9.2 in the 56-Point AEO and SEO Audit Checklist, low-CTR winners, where the method is to compare click-through rate against the published benchmark for each position and flag pages sitting 40 percent or more below it.
Then run a second pass the reports will not give you. Ask an answer engine the questions your buyers ask and read how it describes you when it cites you, because a description quoted back wrong is a metadata problem you cannot see in any dashboard. If click-through rate stays flat after a rewrite, the cause is likely underneath the snippet somewhere in the page itself, and reclaiming lost clicks with a technical SEO audit walks through the technical causes worth ruling out.
Meta tags are the smallest piece of writing on a page and the only piece that three separate systems read before deciding whether to show you at all, which makes them worth the attention and worth leaving alone once they align with the headline and the content underneath. Content Cucumber writes SEO content, titles and descriptions included, for search and for the engines that answer without a results page.
Explore our writing services
Content Services →Keep Reading
Frequently Asked Questions
Yes, and they treat it the opposite way Google does. Answer engines quote meta descriptions close to word for word, and Google's AI Overview citation cards show the page title with a short description that is frequently your own meta description carried straight through. The description Google discards on a results page is the one an answer engine repeats to your buyer.
Write to a length that reads well when it is kept, then stop optimizing the count. Ahrefs found Google rewrote 61.46% of descriptions that ran too long and 63.69% of the ones that stayed inside the limit, so trimming to a character target does little on its own. Front-loading the value does more, because a quoted description gets read with no results page around it for context.
Match the title to the page H1. Zyppy analyzed more than 80,000 titles and found that alignment dropped the odds of a rewrite from 61.6% to 20.6%, and titles that run far too long or far too short get rewritten more than 95% of the time.
It works as a classification signal before selection happens. Generative engines read the title, description, headings, and structured data to sort out what a page covers and how confidently they can use it. Perplexity leans toward pages whose titles and headings mirror the wording of the query, while ChatGPT accepts a partial or semantic match, so consistency across title, headings, and body beats cleverness.
AIO covers how models learn what your business does, so a plain title and description teach them the right thing. AEO covers direct answers, which favors question-shaped headings and a description that answers outright. GEO covers how generative systems build a description of your business from several sources, where consistent phrasing across pages stops them assembling a muddled version of you.
The pages already collecting impressions and few clicks, because those have proven Google will show them and the snippet is the part failing. Sort your Search Console queries by impressions and look for rows where click-through rate sits well below the pages around them. Then ask an answer engine the questions your buyers ask and read how it describes you, because a description quoted back wrong is a problem no dashboard reports.
