Which Google feed fields should hold an AI-written product title or description?

Separate authoring from factual review

Record which text came from a model and which product facts support it. For a fictional desk lamp, dimensions can come from a supplier sheet while the wording comes from an AI draft. Approve the dimensions, finish and included components before mapping the approved copy to the feed. Remove invented certifications and performance numbers even when the phrasing is fluent.

Google documents a digital_source_type sub-attribute and content for the structured fields. Consult the linked title and description specifications for the exact syntax supported by your feed method. A spreadsheet export, XML source and API connector may expose different mapping interfaces; do not paste a JSON object into an ordinary title cell without checking its schema.

For generated copy, the required origin value is trained_algorithmic_media in the appropriate structured field’s digital_source_type, with the approved text in content. The detailed specifications cap title content at 150 characters and description content at 5,000 characters. Check both the retained wording and the submitted field, rather than treating an origin declaration as a fact-check.

Prepare a mapping handoff

The worksheet is an editorial handoff, not an implemented Google integration. Keep the old source value so a malformed mapping can be corrected without rewriting the product record.

Watch the ordinary-field precedence trap. Google’s title specification says that when both title and structured_title are supplied it uses title; the description specification states the equivalent rule. For AI-authored content, use the required structured route and inspect whether the connector also emits an ordinary value that takes precedence. Do not assume that adding a structured field beside an unchanged ordinary field completes the mapping.

Validate one product first

Choose one representative SKU and market. Submit it through your normal authorized feed workflow, then inspect the processed title and description. Check that customers see the intended content rather than field names, serialized braces or an empty title. Also inspect variant distinctions: “large” and “small” should not become indistinguishable simply because a model shortened the copy.

CartVisual can help draft creative copy, but it does not infer your connector’s field compatibility or submit Merchant Center data. A planning draft still needs factual approval and correct downstream mapping.

Fictional handoff: lamp SKU L-BLUE receives approved wording revision C3. Save the expected title, expected description, origin value and selected market. Compare the actual outgoing fields with that expectation, then inspect the processed values for the same item. If a literal serialized object appears as the title, or an old ordinary title is still used, hold the bulk change and correct the connector mapping.

Keep a dated origin record

Store the final wording, creation method, fact-sheet version, reviewer and connector mapping together. Recheck the current specification when switching feed providers. Do not describe a title as “Google approved” merely because the generator completed or a spreadsheet accepted the cell. Feed parsing, policy review and campaign performance are separate observations.

Keep this as a proposed validation method: no external feed submission was performed for the article. A repaired serialization should preserve variant identity and verified facts. Follow the CSV boundary check if the connector uses text files, while using the current attribute specification for its inner structured value.