To rewrite meta titles and descriptions with AI, give it your target keywords, your writing rules and each page's own content, then check every line with a script: for a UK life sciences company, IvI SEO rewrote 538 of 740 pages, and all 706 live pages checked afterwards matched the approved sheet. By my estimate, that job would take about 35 hours by hand. Plenty of time saved.
How do you rewrite meta titles and descriptions with AI?
Rewriting meta titles and descriptions with AI works best from a spreadsheet: one row per page, every line written from that page's facts and checked by a script. IvI SEO's five steps:
- Sheet: every page's address, current title and description, body copy and target keyword.
- Rules: set the title formula (keyword first, something helpful, then the brand), length limits and "facts from the page only".
- Draft: the AI writes each title and description from that page's copy alone, and flags keywords that don't fit.
- Check: a script (the AI can write it) checks lengths, keywords, duplicates and claims against the page.
- Upload: after sign-off, upload each value exactly as approved and re-read the live pages.
The rules come first. Not a clever prompt, just your house style written down. Here's a cut-down version of the brief from that job:
Rewrite the meta title and meta description for every page in this sheet.
- Titles: 50 to 60 characters as a guide, not a hard rule. Keyword first, then one helpful detail about the page, then " | [Brand]" where it fits. One keyword per title, no stuffing.
- Descriptions: aim for 120 to 145 characters, with the main point in the first 120. Say what's on the page and why it's worth the click. End with a short call to action where there's room.
- Only use facts that appear on the page itself: numbers, services, results, names. If the page doesn't say it, don't either. Write something plainer instead.
- Keep anything that already works. Change what breaks the rules, not everything.
- If the keyword in the sheet doesn't suit the page, don't force it in. Flag it and tell me why.
- Put it all in the sheet: new title, new description, character counts, and a note on anything you flagged.
For a made-up plumber, that title formula gives you something like "Boiler Repair Harrogate | Same-Day Callouts | Smith Plumbing": keyword first, something useful, brand last. It depends on the site, and on how much room the keyword leaves, but that's where I start.
The title formula: keyword, something helpful, brand
A made-up example. The grey gaps are the separators. Google cuts titles by width, roughly 600 pixels on desktop.
How do you stop AI making things up in meta descriptions?
To stop AI making things up in meta descriptions, make it read each page first and only use facts that appear on that page, then check every claim against a saved copy of the page. Google's own guidance on AI content says to focus on accuracy, especially when generating titles and meta descriptions automatically. When IvI SEO ran this check for a UK life sciences company, it caught a live description claiming testing started from 20 mg when the page itself said 10 mg.
AI is brilliant at writing something that sounds right. That's the problem. Ask it for a punchy description of a science company's team page and it'll happily give you "over 60 expert scientists" whether the page says that or not.
On this job, that exact thing happened. A draft said "over 60 protein scientists". The page said 65 or more. Close, but not what the page said. Another draft said two teams "won" a competition, when one took the judges' prize and the other took the audience vote. Small, but neither matched the page, and both were caught.
How it was caught:
- Every page was saved first. The AI pulled the live text of all 740 pages and saved one evidence file per page: its current title and description, its headings, its keyword and its actual body copy.
- Every draft was scored against that file. Any number, name or claim that wasn't in the page's own copy got flagged and rewritten plainer.
- Nothing was borrowed between pages. A fact from one page couldn't turn up in another page's description.
The evidence files had a bonus. When the client queried a title, I could open that page's file and judge it myself, rather than taking the AI's word for it.
Google spells this out, too. Its guidance on using generative AI says to focus on accuracy, quality and relevance, "especially when automatically generating the content", and it names title elements and meta descriptions specifically. Here's Google's page on it.
How do you prove AI followed your meta title rules?
To prove AI followed your meta title rules, check every title and description with a script: length in characters and pixels, keyword present, no duplicates, brand where it fits. Don't let the AI mark its own homework. On a 740-page site IvI SEO worked on for a UK life sciences company, the final check on the AI's drafts found no titles over 60 characters or 600 pixels and no descriptions over 155 characters on any page, and five fresh AI checkers found three real mistakes in the sheet, all fixed.
AI is bad at counting characters. It'll tell you a description is 150 characters when it's 170. So on this job, the counting was done by a formula, not by asking.
The script checked:
- Length in characters and in pixels. Google cuts titles off by width, not character count, so a title full of Ms runs out of room before one full of Is. Google doesn't publish a limit: it says titles are shortened "as needed", typically to fit the screen. Around 600 pixels on desktop is a sensible ceiling.
- The keyword is in the title. On the first 103 priority pages, 63 titles didn't include their keyword at all before. Afterwards, 80 did, up from 40.
- No duplicates. Google's own advice is that identical descriptions on every page "aren't helpful". Check whether repeated metadata belongs to genuinely different pages or to the same page available at two addresses.
- Nothing blank. Of the 527 descriptions in scope, 256 were for pages that had no description at all. If you're wondering whether those gaps are worth filling, I've covered that in do meta descriptions still matter?.
First 103 priority pages, before and after
IvI SEO's metadata job for a UK life sciences company, 2026. One PDF among the 103 was left out of the rewrite.
Then the part I'd recommend to anyone: get the work checked by something that didn't do it. Five fresh AI agents compared the finished sheet with the original. They confirmed three genuine mistakes in the AI's own work, including a column of character counts that had gone blank, and rejected two false alarms. All three were fixed before anything went near the website.
And better yet? Get humans to check it. For this project, that's what we did. It's an unusual case, as the client and project lead aren't usually this hands-on with a task like this, but this client is very keen, so it worked out that way, for the benefit of this little project and the campaign so far.
How do you upload meta titles and descriptions in bulk to WordPress?
To upload meta titles and descriptions in bulk to WordPress, send each approved value exactly as written, either through the WordPress REST API or Yoast's own bulk editor, then re-read (with the human eye or AI) every live page to check it matches. Even with access, the REST API only reaches Yoast's fields on standard posts. IvI SEO had no API access on a UK life sciences client's site, so everything went through Yoast's own bulk editor tools in a logged-in browser, and all 706 checked pages matched the approved sheet, apart from 22 apostrophes WordPress straightened.
This is where AI should step out of the way. The upload script read each cell and sent it character for character. No rewording, no "improving" on the way in.
A few things made it safe:
- A test run first. The first 26 pages went up on their own. Only when all 26 checked out did the rest follow: 1,153 saves in total, with zero failures.
- The old values were kept. Every save recorded what was there before, so anything could be put back.
- Only approved rows went up. The client marked approved rows green in the sheet. The 31 that weren't green were skipped, and the script checked every row's colour rather than guessing.
- The live output was checked. After the upload, 706 pages were checked against the sheet. The only differences were 22 curly apostrophes that WordPress had turned into straight ones.
The checks that proved it
- titles over 60 characters or 600px
- 0
- descriptions over 155 characters
- 0
- real mistakes caught by independent checks
- 3
- live pages matched (22 apostrophes aside)
- 706
IvI SEO's metadata job for a UK life sciences company, 2026. Length checks on the AI's drafts; live check after upload.
Usually on client sites, we use the WordPress API, but this time around, due to permission limits, I had to log into WordPress myself. The AI worked inside my session and never needed the password. If you want the full setup for that, including application passwords, I've covered it in how AI can upload content to WordPress for you.
Should you let AI rewrite every meta title on your site?
No, don't let AI rewrite every meta title on your site: keep the titles that already work and the ones with a hook people like, and only change what breaks your rules. IvI SEO's rule is augment, don't replace: on a UK life sciences site, the first AI drafts changed far too much. In the end, 202 of 740 pages were kept exactly as they were, and the client kept several of its own titles.
One blog post on that site had a title built around a bold comparison. It was 70 characters, well over the "rule". The AI wanted to tidy it into something safer. The client said no, and they were right: it fitted on screen by width, and it was the whole reason anyone would click. So it stayed.
That's the job a person still has to do. The AI does the grunt work: 740 pages drafted and checked in under an hour of running time. My review, the client's feedback and the upload came on top of that. You do the judgement: which keyword belongs where, which titles have personality worth keeping, and whether a claim is one the business is happy to make.
And the client round still matters. On this job, the client went through the sheet and suggested changes to around 45 rows. Their wording went in as written, trimmed only if it broke the length limit, with the count shown so everyone could see why.
Want this done for your site?
Rewriting and checking metadata across a whole site is part of the IvI SEO service: every title and description written from your own pages, checked against your rules, and uploaded from the approved sheet and checked afterwards. If yours haven't been looked at in years, take a look at how I can help. I'll tell you straight whether it's worth doing.