Generate Magento Product Meta Descriptions With AI in Bulk — The Anti-Duplicate-Content Prompt

Bulk AI meta-description generation on Magento 2.4.4 — 2.4.9 is the right move when your catalog has more than 1,000 SKUs and the alternative is the rule-based {{name}} - Buy {{name}} at {{store_name}} template that ships duplicate content to Google across every sibling. The workflow described here, shipped on kishansavaliya.com client engagements, rewrote 10,000 product meta descriptions on a live Hyvä store in 60 minutes of compute for $30 of OpenAI API spend, and the affected URLs show a measurable SERP CTR lift in Search Console four weeks on. The pipeline ships as a Magento console command, a 155-character anti-duplicate prompt template, and a rate-limited batch runner that respects OpenAI's Tier-4 RPM ceiling.

Rule-based meta-description templates are the duplicate-content trap

Every Magento store ships with a Page Title and Meta Description template under Stores → Configuration → Catalog → Catalog → Product Fields Auto-Generation. The default reads {{name}} - Buy {{name}} at {{store_name}}. On a 10,000-SKU catalog, that renders 10,000 nearly-identical strings whose only delta is the product name. Google's duplicate-content classifier treats this as boilerplate and collapses the indexed surface — one canonical, 9,999 alternates that never receive impressions.

• Same opening phrase — every meta starts with the product name. 5-gram fingerprint across siblings is >0.9.
• Same closing phrase — every meta ends with the store name.
• No keyword diversity — long-tail buyer-intent keywords never appear because the template has no slot for them.

The AI alternative addresses all three. Each generated meta has a unique opening phrase, a varied middle that paraphrases product attributes, and a closing that rotates across four CTA shapes per category. 5-gram overlap stays under 0.12 across every category we measured.

1. The anti-duplicate-content prompt template

The system prompt is short — 380 tokens of pinned content that ships identically across every API call. Three rules are load-bearing, and the rest is tone and output schema.

You are a Magento meta-description writer for an Adobe-Certified storefront.

RULES — non-negotiable, evaluated by an automated grader:

1. Output must be 155 characters or fewer. Google truncates SERP
   snippets at ~155–160 characters on desktop. Anything longer is
   cut off mid-sentence and loses CTR. Count characters, not words.

2. The primary keyword (supplied per product) must appear within
   the first 60 characters of the output. CTR studies show ~23%
   higher click-through when the primary keyword is front-loaded.

3. You will be given a list of 3 adjacent product names from the
   same category as "DO NOT mention these brands". Do not use
   any of those product names, model numbers, or brand tokens
   in your output. This forces lexical diversity across siblings.

VOICE
- US English, active voice, sentence length under 22 words.
- No exclamation marks, no hype words (best, top, leading, premium).
- One concrete buyer benefit, one specification, one CTA verb.

OUTPUT
- Return only the meta description text, no quotes, no markdown,
  no leading/trailing whitespace. The string is written directly
  to catalog_product_entity_varchar.meta_description.

The three RULES at the top are the entire reason the workflow survives Google's duplicate-content scrutiny. They map to three concrete failure modes the rule-based template inherits by design.

Rule 1 — the 155-character cap, enforced twice

The cap lives at two layers because models drift on length. Layer one is the request body — max_tokens: 50 on both providers, which roughly equals 155 characters at one-token-per-4-characters English. Layer two is a one-line post-process that truncates anything above 155 to the nearest word boundary. About 2% of generations need the post-process.

Rule 2 — front-load the primary keyword

Backlinko's 2024 SERP CTR study^[1] measured 4 million keyword positions and found a ~23% CTR uplift when the primary search query appears in the first 60 characters of the meta. Google bolds matching query tokens in the snippet, and bolded tokens at the start of the line carry disproportionate visual weight on mobile SERPs.

Rule 3 — inject adjacent product names as a do-not-use list

Before each API call, the workflow pulls 3 random adjacent products in the same category and adds their names, model numbers, and brand tokens to the user message as a "DO NOT mention these brands" list. The model treats this as a negative constraint and rewrites with different anchor terms. Two siblings never share the same dominant noun phrase.

The single-line difference between a meta description workflow that gets indexed and one that gets duplicate-content flagged is whether the system prompt tells the model what other siblings have already said.

2. The Claude API request shape

Model pinned to claude-3-5-haiku-20241022 — Haiku is sufficient for 155-character output at $0.80 / 1M input + $4 / 1M output.

The raw HTTP request

curl -sS https://api.anthropic.com/v1/messages \
  -H 'x-api-key: sk-ant-api03-...' \
  -H 'anthropic-version: 2023-06-01' \
  -H 'content-type: application/json' \
  -d '{
    "model": "claude-3-5-haiku-20241022",
    "max_tokens": 50,
    "system": "",
    "messages": [
      {
        "role": "user",
        "content": "Product: 18V Cordless Impact Driver CID-18V-001. Primary keyword: cordless impact driver. Attributes: torque 200Nm, weight 1.4kg, chuck 1/4 hex. DO NOT mention these brands or product names: DeWalt DCF887, Makita XDT13Z, Milwaukee 2853-20."
      }
    ]
  }'

The Python batch client

import asyncio, os, anthropic

client = anthropic.AsyncAnthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
SYSTEM_PROMPT = open("meta-description-system-prompt.txt").read()

async def generate_meta(product: dict, do_not_mention: list[str]) -> str:
    user_message = (
        f"Product: {product['name']} {product['sku']}. "
        f"Primary keyword: {product['primary_keyword']}. "
        f"Attributes: {product['attributes']}. "
        f"DO NOT mention these brands or product names: {', '.join(do_not_mention)}."
    )
    response = await client.messages.create(
        model="claude-3-5-haiku-20241022",
        max_tokens=50,
        system=SYSTEM_PROMPT,
        messages=[{"role": "user", "content": user_message}],
    )
    text = response.content[0].text.strip()
    return text[:155] if len(text) > 155 else text

async def batch(products: list[dict], concurrency: int = 10) -> list[str]:
    sem = asyncio.Semaphore(concurrency)
    async def bounded(p):
        async with sem:
            return await generate_meta(p, p["do_not_mention"])
    return await asyncio.gather(*[bounded(p) for p in products])

The asyncio.Semaphore bounds in-flight requests to 10 — that throttle keeps the batch under the per-minute rate limit while still hitting 10,000 SKUs per hour.

3. The OpenAI rate-limit math

OpenAI publishes tier-based rate limits on the API platform.^[2] Tier 4 — which any production e-commerce account hits inside a month of $250+ spend — allows 10,000 RPM on gpt-4o-mini. That ceiling is what makes 10,000 SKUs in 60 minutes feasible.

The "60 min realistic" row accounts for safety — running at 10,000 RPM risks 429 throttling on adjacent tenants. The CLI defaults to --rate-limit=50/min on Anthropic Tier 2 (Haiku) and --rate-limit=5000/min on OpenAI Tier 4.

4. The CLI batch script

Everything sits behind one Magento console command. It accepts a SKU range, a concurrency level, and an explicit rate limit, and writes a CSV of old-vs-new meta descriptions to var/log/ai-meta-descriptions/run-{timestamp}.csv for review.

bin/magento panth:ai:meta-descriptions:generate \
  --from=1 \
  --to=10000 \
  --concurrency=10 \
  --rate-limit=50/min \
  --provider=anthropic \
  --dry-run

The command class

<?php
declare(strict_types=1);

namespace Panth\AiContent\Console\Command;

use Symfony\Component\Console\Command\Command;
use Symfony\Component\Console\Input\InputInterface;
use Symfony\Component\Console\Input\InputOption;
use Symfony\Component\Console\Output\OutputInterface;
use Panth\AiContent\Model\MetaDescriptionGenerator;
use Panth\AiContent\Model\AdjacentProductPicker;

class GenerateMetaDescriptions extends Command
{
    public function __construct(
        private MetaDescriptionGenerator $generator,
        private AdjacentProductPicker $picker
    ) { parent::__construct(); }

    protected function configure(): void
    {
        $this->setName('panth:ai:meta-descriptions:generate')
            ->addOption('from', null, InputOption::VALUE_REQUIRED)
            ->addOption('to', null, InputOption::VALUE_REQUIRED)
            ->addOption('concurrency', null, InputOption::VALUE_REQUIRED, '', '10')
            ->addOption('rate-limit', null, InputOption::VALUE_REQUIRED, '', '50/min')
            ->addOption('provider', null, InputOption::VALUE_REQUIRED, '', 'anthropic')
            ->addOption('dry-run', null, InputOption::VALUE_NONE);
    }

    protected function execute(InputInterface $input, OutputInterface $output): int
    {
        $from = (int)$input->getOption('from');
        $to = (int)$input->getOption('to');
        $dryRun = (bool)$input->getOption('dry-run');

        foreach ($this->generator->rangeChunked($from, $to,
            (int)$input->getOption('concurrency'),
            (string)$input->getOption('rate-limit')) as $product) {
            $doNotMention = $this->picker->adjacentNames($product, 3);
            $meta = $this->generator->generate($product, $doNotMention,
                (string)$input->getOption('provider'));
            $output->writeln(sprintf('[%s] %s chars: %s', $product->getSku(), strlen($meta), $meta));
            if (!$dryRun) { $product->setMetaDescription($meta)->save(); }
        }
        return Command::SUCCESS;
    }
}

The command is provider-agnostic. --provider=anthropic hits Claude Haiku; --provider=openai hits GPT-4o-mini. Both reach the same prompt and the same 155-character output schema.

5. Comparison — rule-based vs AI bulk generation

Decision matrix on a 10,000-SKU catalog. Uniqueness is 5-gram overlap inverted (1.0 = no overlap); CTR is the Search Console 28-day average.

The AI workflow lands at 93% of the hand-written CTR for under 1% of the cost. On a 10,000-SKU catalog, that gap pays back in two days of incremental organic clicks.

6. Before-and-after SERP CTR

Search Console data for the affected URL set, 28 days before vs 28 days after the rewrite shipped on 2026-04-15.

• Impressions — 184,200 → 217,400 (+18%).
• Clicks — 2,580 → 6,310 (+145%).
• Average CTR — 1.4% → 2.9% (+107%).
• Average position — 19.8 → 18.4.
• Manual actions — 0 before, 0 after. The dup-guard held.

Going from 0.26 clicks per SKU per month to 0.63 on a 10,000-SKU catalog is 3,700 extra organic sessions per month — at a 1.5% conversion rate and $80 AOV, ~$4,400 incremental monthly revenue from a $30 one-time rewrite.

7. Saving to Magento — the EAV detail that breaks bulk runs

The meta_description attribute lives in catalog_product_entity_varchar, scoped per store. Two gotchas every bulk run hits on Magento 2.4.4 — 2.4.9. First, the default $product->setMetaDescription() writes to the admin scope (store 0); on multi-store-view setups call setStoreId($storeId) once per view or customer-facing views fall back to the template. Second, after a bulk run the meta description is rendered into FPC-cached HTML and indexed by catalogsearch_fulltext — both must clear or the SERP shows the old meta description for hours while Googlebot recrawls.

bin/magento indexer:reindex catalogsearch_fulltext
bin/magento cache:clean full_page block_html
bin/magento cache:flush

On a 10,000-SKU rewrite, catalogsearch_fulltext reindex takes 4–8 minutes on a Hyvä store with the default OpenSearch backend. Run it once at the end of the batch, never per-product inside the loop.

8. Haiku vs GPT-4o-mini — pick by your API tier

Both models work. Claude Haiku at $0.80 / 1M input + $4 / 1M output with the 90% prompt-cache discount lands ~$3.00 all-in on 10,000 SKUs but caps at 50 RPM on Tier 2. GPT-4o-mini at $0.15 / 1M input + $0.60 / 1M output with the 50% prompt-cache discount lands ~$2.40 all-in and runs at 10,000 RPM on Tier 4. Pick OpenAI when wall-clock matters; pick Haiku when you are not yet at OpenAI Tier 4 and the 7-day Anthropic Tier-2 onboarding clears faster.^[3]

What this workflow does not do

It does not write Page Title — that lives in a separate Panth_AdvancedSeo template with stricter brand rules. It does not write long-form product descriptions — see the AI product descriptions workflow. It does not regenerate canonical tags, hreflang, or schema. It also skips configurable parent SKUs — the CLI command writes only to simple SKUs because parent products inherit a child meta-description aggregation under Magento's default catalog setup.

Related reading

• AI-generated product descriptions in Magento — the bulk workflow that doesn't get penalized
• How to integrate ChatGPT with Magento 2 — 4 real patterns
• Magento 2 development service