How to Translate 10,000+ Product Pages into English Using AI (and Make Them SEO-Ready)

AI Export Lab · July 30, 2026

The batch AI translation pipeline that turned a Chinese product catalog into 15,000 Google-indexed English pages. Handling technical terms, OE numbers, and the places Google Translate fails.

AI Export Lab | July 30, 2026

Most Chinese factories have a product catalog. It is usually a PDF, a spreadsheet, or an internal ERP export. It is in Chinese. The factory wants to sell overseas.

The obvious move is to translate everything into English and put it on a website. The trap is thinking Google Translate can do it.

Google Translate butchers technical product catalogs. It translates part numbers as if they were words. It turns "散热器芯体尺寸 680×420mm" into something like "radiator core size 680 times 420 mm" — the "×" becomes the multiplication word "times" instead of "by." It translates OE numbers like "BJ-3-12" into whatever it guesses "BJ" might mean. And it has no idea that "铝合金" should consistently be "aluminum alloy" across 15,000 pages, not randomly alternate between "aluminum alloy," "aluminium alloy," and "aluminum metal."

When I built jjradiator.com, I needed 15,000 product pages in English. Here is how I did it with AI, what I learned about preserving technical accuracy, and the SEO setup that got them indexed by Google.

---

Why Google Translate Fails for B2B Product Catalogs

Three specific failures that matter:

1. OE number corruption. Part numbers like "BJ-3-12" or "AA-123/456" contain punctuation that Google Translate interprets as sentence structure. Dashes become spaces. Slashes become "or." You end up with part numbers that do not match the original catalog and cannot be searched.
2. Measurement unit chaos. "680×420mm" is a dimension. MM is a unit. Google Translate sometimes converts units ("680mm × 420mm") and sometimes does not. Inconsistency across 15,000 pages makes your product data unreliable.
3. Terminology drift. A product catalog uses the same terminology hundreds of times. "芯体," "散热器," "铝合金." Without a glossary, each page gets a slightly different translation of the same term. Search engines see 15 different English phrases for the same product concept and treat them as separate content.

AI translation solves all three if you give it rules, not just text.

---

The AI Translation Pipeline

Four steps, each handled by AI with specific instructions:

1. Glossary injection. Before any translation starts, feed the AI a fixed glossary. "散热器" always = "radiator." "铝合金" always = "aluminum alloy." "芯体" always = "core." Lock these down once at the start.
2. OE number protection. Tell the AI: "Any string matching the pattern of a part number — letters + numbers + dashes/slashes — must appear exactly as-is in the output. Do not translate. Do not modify punctuation."
3. Batch processing with consistency check. Process pages in batches of 100. After each batch, run a script that checks the 10 most common terms against the glossary. If 散热器 came out as "heat sink" on page 47, flag it and reprocess.
4. SEO metadata generation. After the body text is translated, generate English title tags and meta descriptions. AI creates these from the translated content, using the glossary to ensure keyword consistency.

The whole pipeline ran across 15,000 pages. Total API cost: under $50. A human translation agency quoted me approximately $2 per page for technical product content. That is $30,000 versus $50.

---

The SEO Layer: More Important Than the Translation

Translation gets you English pages. SEO gets those pages found.

For each product page, the AI also generates:

• A readable title tag. Not "Product AA-123" but "AA-123 Aluminum Radiator for [Vehicle Model] — OEM Replacement." The title includes the search terms a real buyer types.
• A meta description with selling points. Dimensions, material, compatibility notes. These show up in Google's results and directly affect whether someone clicks.
• Structured data. Product schema markup with OE number, manufacturer, dimensions, and weight. This helps Google display rich results.
• Internal links. Each product page links to related OE numbers and the brand page. This builds a crawlable site structure for 15,000 pages.

Without this SEO layer, you have 15,000 English pages that Google treats as thin content. With it, you have a proper product catalog that Google can understand and rank.

---

What to Check Before Publishing

AI translation is fast. It is not perfect. Before publishing a batch:

• Spot-check 5% of pages. Compare original Chinese to English output.
• Check all OE numbers in the spot-check batch. One corrupted part number means you are sending buyers to the wrong product.
• Verify the top 10 most-viewed product pages manually. These are your money pages.
• Run the consistency script I described in step 3. It catches 95% of terminology drift.

This takes an hour per 1,000 pages. It is the difference between "an AI translated this" and "this is a real product catalog."

---

How to Start Your Own Pipeline

You need three things:

• A product database. Spreadsheet is fine. I built mine from PDF catalogs using the workflow described here.
• An AI API. I use DeepSeek V4 through Volcengine Ark. Any API-capable model works. The important part is the glossary and the protection rules, not which model.
• A Python script to batch the calls. Ask the AI to write it for you — the same manager-worker pattern applies.

Start with 100 pages. Verify. Adjust the glossary. Then scale to 1,000. Then 10,000. The pattern holds at any scale.

---

Translating a product catalog? I am curious what languages and industries people are working on. Tell me about yours — I read every message.