Writing meta descriptions one at a time for a site with a few hundred posts is one of those tasks that sounds quick and then eats an entire afternoon. I started batching this through GPT-5 a while back after getting tired of the repetitive part of it, and honestly the setup took some trial and error before the output stopped sounding like the same five sentences reworded over and over. Here’s the actual workflow that’s stuck.
Quick Answer
- Feed GPT-5 a structured list (title + URL + target keyword) rather than pasting raw content for each page
- Set a strict character limit instruction (150-155 characters) directly in the prompt, since the model will run long without it
- Request output in a consistent format (CSV or table) so it drops straight into a spreadsheet
- Batch in groups of 20-30 rather than dumping hundreds at once — quality drops and repetition creeps in on huge single batches
- Always review for near-duplicate phrasing across the batch, since GPT-5 tends to reuse sentence structures when generating many similar items back to back
Why Bulk Generation Needs a Different Approach Than One-Off Prompts
Asking for a single meta description usually goes fine with almost any prompt. Asking for two hundred in one go is a different problem entirely, because the model has a tendency to lock into a pattern early and repeat that structure with just the nouns swapped out — you end up with descriptions that all start “Discover how to…” or “Learn everything about…” which defeats the point of writing them individually in the first place.
There’s also a practical formatting issue: without explicit instructions, GPT-5 will happily generate descriptions that run 180+ characters, which get truncated in search results and waste the effort entirely. Character limits need to be stated as a hard constraint, not a suggestion, or the model treats it more like a rough target.
Step-by-Step Workflow
Step 1: Prepare Your Source Data as a Structured List
Build a simple spreadsheet or plain list with three columns: page title, target keyword or keyphrase, and a one-line note on the page’s actual content or angle if the title alone isn’t descriptive enough. Don’t paste full article text for each page into the prompt — it bloats the input unnecessarily and doesn’t meaningfully improve description quality for this specific task.
Step 2: Write a Prompt With Explicit Constraints
Something along these lines works better than a vague ask:
“Generate a unique meta description for each item below. Requirements: 150-155 characters including spaces, must naturally include the target keyword, active voice, no generic openers like ‘Discover’ or ‘Learn,’ write as if explaining the actual value to a reader deciding whether to click. Output as a table with columns: Title, Meta Description, Character Count.”
Being specific about banned generic openers matters more than it seems like it should — without that instruction, a huge share of bulk-generated descriptions default to the same two or three sentence starters.
Step 3: Submit in Batches of 20-30 Items
Larger batches tend to increase repetition across the set, since the model’s pulling from a narrower internal range of phrasing patterns the longer a single generation run gets. Smaller batches cost more back-and-forth but noticeably improve variety.
Step 4: Review for Duplicate Phrasing Across Batches
This is the step people skip, and it’s the one that actually matters most for bulk work specifically. Since each batch is generated somewhat independently, you can end up with very similar-sounding descriptions across different batches even though each batch internally looked fine. Skim the full combined list before publishing anything, not just each individual batch as it comes out.
Step 5: Verify Character Counts Manually for a Sample
GPT-5 is decent at self-reporting character counts but not perfect — spot-check maybe 10% of the output against an actual character counter rather than trusting the reported numbers wholesale, especially for descriptions that look like they’re pushing right up against the limit.
Step 6: Export and Import Into Your CMS Bulk Editor
If you’re on WordPress with Rank Math or Yoast, both support CSV import/export for meta fields, which lets you skip manually copy-pasting each description into individual post edit screens. Match your exported column headers to what the plugin’s import tool expects — this varies slightly between plugins, so check the specific import format before building your final CSV.
What Actually Worked For Me
My first attempt just dumped a list of 150 blog post titles into one prompt and asked for meta descriptions for all of them. The output looked fine at a glance, but when I actually skimmed the whole batch afterward, a huge number of them followed the exact same “If you’re looking to [X], this guide covers [Y]” template with just the specific words changed. Technically 150 unique descriptions, but functionally repetitive enough that it would’ve looked bad if anyone compared several side by side in search results.
Splitting the same list into batches of 25 and adding the explicit “no generic openers” instruction fixed most of it, though not entirely — I still found maybe four or five near-duplicates scattered across different batches when I did a final read-through of the combined output. Rewrote those manually rather than trying to re-prompt for just those few, which honestly ended up being faster than trying to explain the exact duplication issue back to the model for a handful of items.
Tips for Better Bulk Output
- Include the actual searcher intent, not just the topic, in your source list where it’s not obvious from the title alone — this helps the model write toward what someone would actually want to know rather than restating the title in sentence form
- Ask for 2-3 variations on a handful of your most important pages rather than accepting the first output for high-priority content, then pick the best manually
- Specify tone explicitly if your site has a particular voice (casual, technical, urgent) since default bulk output tends toward a neutral, slightly generic tone unless told otherwise
- Keep a running list of banned phrases as you notice repetition patterns emerging, and add them to your prompt for subsequent batches on the same site
Limitations Worth Knowing
Bulk-generated descriptions still benefit from human review before publishing, especially for pages covering something factual or numerical (pricing, statistics, dates) where the model might smooth over specifics that need to stay exact.
Character counting isn’t always perfectly precise in the model’s own self-reported numbers, so treat that column as a helpful guide rather than a guaranteed accurate measurement, particularly for descriptions near the upper limit.
Very niche or technical topics sometimes produce vaguer descriptions than general-audience content, since there’s less for the model to draw on for natural-sounding specific phrasing — these are worth manually reviewing and rewriting more often than general topics.
FAQ
Is there a limit to how many meta descriptions I can generate in one conversation? No hard limit from the model itself, but quality and variety hold up much better in smaller batches, so it’s a practical choice rather than a technical restriction.
Should I include my target keyword exactly, or can GPT-5 work from a general topic? Giving the exact target keyword produces noticeably better results than a general topic description, since the model needs to know precisely what phrase to work into the sentence naturally.
Can this same approach work for title tags too, not just meta descriptions? Yes, the same batching and constraint-setting approach applies well to title tags, just with a different character limit (usually around 55-60 characters).
Does bulk generation work better with a spreadsheet upload instead of a pasted list? Either works, but uploading a spreadsheet directly (if your interface supports file uploads) tends to keep formatting cleaner for very large batches compared to pasting a long list as plain text.
How do I stop the model from including my brand name in every single description? Explicitly instruct it not to include brand name unless specifically relevant, since default behavior in bulk mode sometimes appends a generic brand mention to every item as a kind of filler.
Editor’s Opinion
the repetition thing catches people off gaurd more then they expect, looks fine untill you actually read the whole batch together. smaller batches plus telling it exactly what not to write (the generic openers) made the biggest diffrence for me, way more then any fancy prompt trick.