Rule-Based vs AI Product Titles: Which One Your Feed Needs

Google says users "will usually notice only the first 70 or fewer characters" of a product title, even though the attribute accepts up to 150. Most feeds are built as if the opposite were true: the distinguishing detail sits at the end, and it is the first thing cut. This guide is for feed managers and agencies who have to decide how titles get written at catalog scale: by template rules, by a language model, or by a mix of both.

By the end you will be able to score your own catalog against six criteria, pick an approach, and apply the guardrails that keep each approach inside the title specifications of Google, Meta and TikTok Shop. We include one worked feed row, the exact Google attribute you need when a model writes the title, and a Turkish casing trap that breaks naive title-case rules.

The three approaches in one table

All three produce the same output field. They differ in where the words come from and who is accountable for them.

What the channels actually enforce

Whichever approach you choose, the output has to fit the channel specifications. These are the rules we verified against each channel's own documentation.

One discrepancy is worth knowing about. GoDataFeed's disapproval guide states that TikTok enforces a 25 to 200 character range. TikTok's public catalog parameter page does not state any range. We treat 25 to 200 as a safe operating band and note that we could not confirm it from TikTok's own text. The rest of the numbers above come from Google and Meta documentation directly.

Where template rules break

Templates fail in four predictable ways, and every one of them is visible in a feed export before it reaches a channel.

1. Empty slots. A template of brand + type + color + size produces "Acme Outdoor Base Layer Top - - M" when color is blank. Always drop empty slots and their separators.
2. The wrong "brand" field. If the source maps your store name into the brand slot, you violate Google's rule against adding the company name to the title.
3. Source text that already contains promotion. A template that appends attributes to an existing title keeps "FREE SHIPPING" and "Summer Sale 20% off" in place. Build from attributes, not from the legacy title.
4. Casing functions that ignore language. This one matters for Turkish catalogs. In Turkish, capital I lowercases to dotless ı and dotted İ lowercases to i; the Unicode SpecialCasing table defines these as language-specific mappings. A generic lowercase function does not apply them, so "KIRMIZI" (red) becomes "kirmizi" instead of "kırmızı". The result is a title that looks wrong to a Turkish shopper and does not match how they search. Use a locale-aware casing routine for tr feeds.

Where AI titles break

A model is useful exactly where templates fail, on messy text. It fails in a different way: it adds words that nobody verified.

A worked row

The row below is illustrative, not a client feed. The source title is the kind of supplier text many catalogs inherit.

The template output is 64 characters, so everything that identifies the product sits inside the first 70. The AI output is longer, and it introduces "Ultra-Warm", "Thermal", "Hiking" and "Skiing". None of those words exist in the source attributes. If the garment is a summer-weight liner, the title now makes claims the landing page does not support. The first 70 characters of the AI version also end mid-word at "for Hiki", which pushes color and size beyond the window Google says users usually notice.

Our own earlier piece on QA for AI-enriched product attributes covers the same failure at attribute level. The title version is more visible because shoppers see it first.

The Google labelling rule

Google requires titles created with generative AI to be submitted through structured_title instead of title. The structured attribute has two sub-attributes: digital_source_type with the value trained_algorithmic_media, and content, 1 to 150 characters. Google also states that if you provide both structured_title and title, only title is used. Two consequences follow. A pipeline that writes AI text into title is non-compliant with the AI-content guidance. And a pipeline that sends both fields silently discards the structured one.

Google's AI-content page, as we read it, does not describe the penalty for non-compliance or whether a label is shown to shoppers. Treat the labelling as a compliance requirement and do not promise a client what it will do to performance.

Score your catalog before you choose

Score each criterion from 0 to 2. A low score favours rules, a high score favours AI involvement. This rubric is our own heuristic, not an industry standard, and the thresholds are a starting point to adjust after a test.

Total 0 to 4: use template rules. Total 5 to 8: use the hybrid. Total 9 to 12: AI-led generation is reasonable, with the validator and the structured_title labelling still mandatory. Note that the last criterion works in the opposite direction from intuition on purpose: the more a wrong word can hurt, the less freedom a model should have.

Build the hybrid in six steps

1. Measure truncation first. Export titles and count how many lose their distinguishing attribute after character 70. Google's "Text too long" notice also appears in the Needs attention tab when titles exceed limits.
2. Write one formula per category. For apparel, follow Google's own order: brand, product type, then size type, gender, age group, color, size, material. Put the identifying words inside the first 70 characters.
3. Build from attributes, never from the legacy title. Drop empty slots with their separators and strip anything that matches your promotional-word list.
4. Let AI fill only missing attributes. Give the model the product image and description and ask it to return the missing color or material as a value, not a sentence. A model returning a field is easier to validate than a model writing a title.
5. Validate before export. Reject any title over 150 characters, with sale or shipping words, with ALL-CAPS runs, or with a word that is absent from source attributes and not in an approved vocabulary list.
6. Label and diff. Route any AI-written title through structured_title with trained_algorithmic_media. Keep the previous title and compare impressions and clicks per SKU group over a fixed window before rolling out.

Steps 2 to 5 are the kind of work a rule and enrichment layer exists to do. Feedance combines feed enrichment with rules and AI before the feed is exported to a channel such as Google Merchant Center. For description-side limits, see our guide to writing product descriptions that pass feed review.

What this guide doesn't cover, and where it breaks down

We did not find an independent, controlled study that measures AI titles against template titles on click-through or conversion. Vendor case studies exist, but we could not verify their sample sizes or methods, so this guide gives no performance uplift figure. If you see a precise percentage quoted for title optimization, ask for the test design.

The rubric is a judgement tool, not a measured model. The 70-character guidance is Google's description of what users usually notice, not a hard cut. Channel specifications change, and the TikTok range discrepancy above is an example of a rule you should recheck in your own account.

If your whole need is concatenation rules on a single feed and you are not using AI, a lightweight rule tool is enough. Established rule-based tools such as DataFeedWatch and Channable are a natural fit for that job, and we do not claim our rule builder beats them for a template-only setup. The hybrid approach earns its complexity when source data is uneven, when you serve several markets, or when creative and feed work happen in the same pipeline.

Frequently asked questions

How long should a Google Shopping title be?

Google allows up to 150 characters and says users usually notice only the first 70 or fewer. Put the brand, product type and the attribute that separates this item from its siblings inside that window. Use the remaining characters for secondary detail.

Do I have to label AI-generated titles?

For Google, yes. Titles created with generative AI must go in structured_title with digital_source_type set to trained_algorithmic_media. Other channels' pages we read did not state an equivalent rule, but check each one in your account.

What happens if I send both title and structured_title?

Google states it will use only the title attribute. If you intend the AI-generated title to be shown, do not send a standard title for the same item.

Can I include the brand name in a title?

Yes, and Google's apparel guidance puts it first. What Google prohibits is adding your company name, meaning the store, as extra information. The two differ whenever you resell other brands.

Is 25 characters really TikTok's minimum?

One feed vendor's guide says so, citing a 25 to 200 range. The TikTok catalog parameter page we read states no minimum or maximum. Keep titles in that range anyway, since it costs nothing.

Why does my Turkish title look wrong after automatic capitalization?

Generic casing functions do not apply Turkish dotted and dotless I mappings. Use a locale-aware function for tr feeds and test it on words such as KIRMIZI and ISPARTA.

Should I rewrite titles for every SKU or only the weak ones?

Start with SKUs that lose their key attribute after character 70, then expand. A scoped change is also easier to measure against an unchanged control group.

Does Meta need shorter titles than Google?

Meta accepts up to 200 characters but recommends 65 to avoid cut-off. A title built for Google's 70-character window is a good fit for both.

Check where your titles stand

Run your feed through the free product feed audit tool to see completeness and quality scores for the attributes your title formula depends on.

Sources

• Google Merchant Center Help: Title [title] and structured title [structured_title]
• Google Merchant Center Help: AI-generated content
• Google Merchant Center Help: How to fix: Text too long
• Google Merchant Center Help: Best practices for clothing and accessories
• Meta for Developers: Catalog reference
• TikTok Ads Help: Catalog product parameters
• GoDataFeed: TikTok product title under 25 character minimum
• Unicode Character Database: SpecialCasing.txt