Sarah had been freelancing as an SEO consultant for five years. She was good at her job — audits, keyword research, on-page optimisation. But when a fashion e-commerce client handed her a project that included writing meta tags for all 2,400 product and category pages, her heart sank a little.
Not because it was hard. Because it was relentlessly repetitive.
The problem with doing it manually
She'd done projects like this before. Open a spreadsheet. Write a title tag. Write a meta description. Move to the next row. Two hours of focused work would get her through maybe 80–100 pages. 2,400 pages would take most of her working month — time she didn't have, and that the client's budget couldn't justify.
A formula-based approach — something like "Buy [Product Name] | [Brand]" — was fast but lazy. Google actively demotes templated meta tags that don't differentiate pages. Her client's competitors were already doing better.
She'd tried pasting rows into ChatGPT before. It worked for ten pages. At twenty it started feeling like a second job. There was no realistic path to 2,400 that way.
Setting up the batch job
Sarah had heard about AI batch processing from a post in an SEO Slack community. She signed up for PromptMax, started with the free credit, and had a test job running in about fifteen minutes.
Her CSV had four columns — one row per page:
| page_url | h1 | page_type | top_keyword |
|---|---|---|---|
| /products/silk-midi-dress | Silk Midi Dress | product | silk midi dress |
| /collections/occasionwear | Occasionwear | category | occasionwear dresses UK |
Her batch instructions — written once and applied to every row — looked like this:
Batch instructions Sarah used:
You are an SEO copywriter. Write a title tag and meta description for this page.
Rules:
— Title tag: max 60 characters, include the top_keyword naturally, end with the brand name "Bellrose"
— Meta description: 140–155 characters, include a benefit or USP, end with a soft call to action
— Do not use generic filler phrases like "Shop now for the best..."
— Do not use the word "perfect"
— Output format: two lines only — "Title: [title tag]" then "Description: [meta description]"
She ran a test on 20 rows first. The output was solid — distinct, keyword-appropriate, within character limits. She tweaked one instruction (adding the note about the word "perfect" after seeing it appear three times) and ran the full 2,400-row file on Gemini 2.5 Flash.
The result
The batch job completed in just under two hours. Sarah downloaded the output CSV, did a spot-check on 50 rows, and found eight that needed minor edits — mostly pages where the H1 was vague or the keyword column was blank. She fixed those manually and was done.
Total time: about three hours — including setup, testing, the run itself, and QA. A project that would have taken three weeks of focused manual work.
The client was happy. Sarah charged her day rate for the work, which was fair given the quality delivered, and moved on to the rest of the audit.
What to take from this
A few things are worth pulling out if you want to replicate this approach for your own SEO work:
- Test on 20–30 rows first. Running a small batch before committing the full file lets you refine your batch instructions without wasting credits or time. Catching issues early is much easier than correcting 2,400 rows after the fact.
- Specify the output format explicitly. Telling the model exactly what to return — "two lines only, Title: then Description:" — makes the results import cleanly into any CMS without manual reformatting.
- Use all the context you have. The more each row contains — page type, top keyword, existing H1 — the better and more varied the output. Generic input produces generic meta tags.
- Accept 95% as done. A small number of rows will need manual review, especially where the source data is thin or ambiguous. Build QA time into your process rather than trying to write instructions that handle every edge case perfectly.
Bulk meta tag generation is one of the clearest uses for AI batch processing in SEO. The task is well-defined, the input data is structured, the output format is consistent, and the quality bar is achievable. If you're regularly doing this kind of work for clients, it's worth having the workflow set up and ready to run.