Why your AI-rewritten product feed gets rejected (and how to fix it)

A pattern we see weekly: merchant uses ChatGPT or Claude to bulk-rewrite their product titles and descriptions for “better SEO”, uploads via Shopify Markets or via DataFeedWatch, and within 48 hours the GMC account is suspended for misrepresentation or untrustworthy promotions.

The rejection isn’t because AI rewriting is fundamentally wrong. It’s because the rewriting was done outside the constraints Google’s review engine enforces. Here are the four failure modes we see most often.

Failure 1: invented factual claims

The big one. A merchant asks ChatGPT to “rewrite this product description to be more compelling”. The model - being a probabilistic language model, not a fact lookup - fills in gaps with plausible-sounding details: “ergonomic design tested by physiotherapists”, “made from sustainably-sourced materials”, “winner of the 2023 Red Dot Design Award”. None of these are true. The merchant uploads the rewrite without checking. Google’s review engine - which can verify award claims, certification claims, and supplier claims via its own knowledge graph - flags the listings as misrepresentation.

Fix: any production AI rewrite pipeline must explicitly forbid factual invention. Our AI Feed Optimizer uses prompts that constrain the LLM to reformulation only - never adding claims not present in the source. Restricted-category products (healthcare, supplements, finance) get the strictest constraint and run through Opus-4.7 for legal-grade carefulness.

Failure 2: title stuffing past Google’s quality threshold

The legacy SEO playbook says “put your top 5 keywords in the title”. Google’s 2026 review engine treats title stuffing as a quality signal - the more keywords you cram in, the more likely it flags the listing as low-quality. Specifically, listings that exceed roughly 70% non-content-words (brand, color, size, generic adjectives) trigger an automated quality flag.

A title like “Premium Luxury High-Quality Soft Cotton Comfortable Stylish Designer Men’s T-Shirt Black Large XL Best Gift Sale” gets flagged. A title like “Acme Cotton Crew T-Shirt - Black, Large” doesn’t.

Fix: Optimize for clarity, not keyword density. The order Google’s quality model rewards is: brand → product type → key attribute → secondary attribute. Keywords that don’t fit that pattern hurt rankings as much as helping them.

Failure 3: ALL CAPS or promotional language in titles

A common pattern: merchant runs a sale, asks AI to “make the titles attention-grabbing for the sale”. Output: ”🔥 BEST PRICE EVER! Acme T-Shirt - 50% OFF - FREE SHIPPING”.

Google bans promotional language in feed titles outright. The [title] field is for product description, not promotion. Sale prices and shipping promos belong in their own structured fields (sale_price, shipping).

Fix: Run a deterministic post-process on every AI-generated title that strips emoji, ALL-CAPS sequences > 3 letters, and a fixed list of promotional keywords (FREE, SALE, BEST, OFF, DISCOUNT, NOW, LIMITED). This is what our optimizer does before any push.

Failure 4: missing or invalid GTIN / brand attributes

When AI rewrites a product, it sometimes drops or modifies structured attributes alongside the prose. The [gtin] field, [brand] field, and [google_product_category] need to remain stable - even if the title and description change.

A common bug: the AI removes a brand reference from the title, the merchant assumes the field is also no longer needed, deletes the attribute. Google’s reviewer requires [brand] for branded products and flags missing values.

Fix: Treat structured attributes (gtin, mpn, brand, googleProductCategory, condition, ageGroup, gender, color, size) as separate fields that the AI never touches. Only [title], [description], and [image_link] rewrites should be in scope. Our optimizer makes this separation explicit at the prompt level.

The right architecture for AI feed rewrites

Reverse-engineered from running thousands of corrections through Merchant API:

Push corrections to a supplemental data source, never to the primary feed. Google merges supplemental on top of primary, so reverting is one click - dataSources.delete() makes the original feed take over again within minutes.
Read back via products.get after every push. Verify the merged result actually has the change. Sometimes Google rejects supplemental fields silently - only the readback catches it.
Compare item-level issues before vs after. If new issues appeared (rare but happens), automatically revert. If old issues persisted, flag for human review.
Cap LLM cost per merchant per month. Restricted categories, high-value SKUs only - don’t let the optimizer churn through 10K low-revenue products eating budget.
Force Validate mode on restricted categories regardless of plan. Healthcare, pharma, finance, alcohol, gambling, weapons. Auto-pilot is too risky here - the legal exposure on a wrong rewrite belongs to the merchant, not the optimizer, but only if the merchant approved the rewrite.

This is the architecture we run in production. The product is GMC Guardian’s Auto-Fix, and the design is exactly the constraints above hardened into the API contracts.

Try the Validate-mode optimizer free on your first 5 product rewrites - see exactly what corrections we’d suggest before paying for any push.