How to Optimize Google Shopping Titles: Rules-Based and AI Methods

The title is the single field Google reads most when it decides which queries your product can show for. Get it right and you widen the pool of searches you match, lift click-through rate, and lower your cost per click. Get it wrong and you bid on traffic that was never going to convert.

There are two ways to build better titles at scale. The first is rules-based: you concatenate the attributes you already hold into a consistent, query-friendly structure. The second is AI-based: you hand the raw data to a language model and let it write or rewrite the title. Both have a place, and in Feedoptimise you can run either one, or both together, without touching code.

This post walks through each method with examples, and shows where the modifiers and the multi-LLM connectivity in the platform fit.

What Google looks for in a title

Before optimizing anything, line your titles up against three constraints.

Length. The title attribute allows up to 150 characters, and Google's guidance is to use them. Put the attributes that matter for matching and for the click in the first 70, since users usually see only the first 70 or fewer depending on screen size, then add SEO-friendly keywords after that to fill the remaining space. Using the full capacity is the goal, not trimming titles short.

Order. Google reads left to right and weights early words more heavily. Front-load the terms a shopper would type, and re-sequence existing keywords so high-CTR terms sit near the front. If "vinyl" is a strong term for music, "Vinyl, Eminem, Curtain Call: The Hits (2LP)" beats leaving "Vinyl" at the end.

Structure. Google publishes suggested title formats by product category. The patterns most retailers work from:

• Apparel: Brand + Gender + Product Type + Attributes (Color, Size, Material)
• Consumables: Brand + Product Type + Attributes (Weight, Count)
• Hard goods: Brand + Product + Attributes (Size, Weight, Quantity)
• Electronics: Brand + Attribute + Product Type (plus Model Number)
• Books: Title + Type + Format + Author

Google's base electronics pattern stops at product type, but adding the model number is worth it, since shoppers search for specific models by name.

Keep titles unique. Each product variant should have its own title rather than sharing one across sizes or colors. A title that carries size and color already does this, and you can confirm it during setup.

Keep promotional language out. Words like "free shipping", "best price" or "sale" get titles disapproved and add nothing to query matching.

With those rules in mind, here are the two build methods.

Method 1: Rules-based titles through attribute concatenation

The idea is simple. You take structured fields you already have, brand, gender, product type, color, size, material, and join them in a fixed order to produce a clean, predictable title for every product in the catalog.

This is the right method when your source data is reliable and well structured. It is fast, every output is predictable, and you can audit exactly how any title was produced.

The modifiers you use

Feedoptimise ships with more than 120 modifiers, and you chain them in a visual rule builder where the output of one becomes the input of the next. For title concatenation, a handful do most of the work:

• Concatenator: joins two or more strings into one.
• Mapper: maps raw values to cleaner ones, useful for normalizing messy inputs before they reach the title.
• IF: returns different values based on a condition, so the structure can change by product type or by which fields are populated.
• Case formatter: fixes capitalization, so ADIDAS becomes Adidas.
• Trim and Text sanitiser: strip stray spacing and hidden characters.
• Truncate and Text counter: keep the title inside the character budget and tell you how long it is.

Worked example: apparel

Say you sell clothing and hold these fields:

brand        = adidas
gender       = Women's
product_type = Running Shoes
color        = Core Black
size         = US 6

Following Google's apparel pattern, you chain the fields with the Concatenator in this order:

Brand + Gender + Product Type + Color + Size

The output:

Adidas Women's Running Shoes Core Black US 6

Every product in the feed now follows the same shape. A shopper searching "women's adidas running shoes" matches on the words that sit at the front of the title.

Worked example: electronics

Electronics put the defining attribute and the model number to work:

brand        = Samsung
attribute    = 65 inch
product_type = 4K QLED Smart TV
model        = QE65Q80D

Chained as Brand + Attribute + Product Type + Model:

Samsung 65 inch 4K QLED Smart TV QE65Q80D

Adding logic and guardrails

Real catalogs are not tidy, so a few extra steps keep the output clean.

Conditional structure. Use IF so the title format adapts. If gender is empty, drop it from the chain rather than leaving a double space. If product_type is "Books", switch to the book pattern.

Normalize before you join. Run the Mapper so "blk", "Black" and "BLACK" all resolve to "Black" before they enter the title. You do not need to flatten descriptive color names though. "Walnut" or "Royal Sapphire" can stay in the title as long as the separate color attribute carries a standard value. If you only hold color inside the title and not as a separate field, the Color extractor can pull it out, so "Royal Sapphire Cocktail Dress" yields "Blue".

Fill gaps from existing data. The Gender extractor and Age group extractor read the title or category and return the right value, which you can then slot into the structure.

Use the full title, then guard the ceiling. Aim to fill the 150 characters: lead with the attributes that drive matching and clicks, then append relevant keywords after the first 70. Set Truncate to 150 as a hard cap so nothing overflows, and use Text counter during setup to confirm your key attributes land inside the first 70.

Every one of these steps previews live against real products from your catalog before you publish, so you see the finished titles, and catch the edge cases, before anything reaches Google.

Where rules-based falls short

Concatenation only rearranges what you already have. If your source titles are thin, inconsistent, or written for humans rather than search, joining fields produces a tidy title that still misses the terms shoppers use. That is where the second method earns its place.

Method 2: AI-based titles with multi-LLM connectivity

Instead of assembling fields in a fixed order, you give a language model the product data and a prompt, and it writes a title that reads naturally and includes the terms people search for. The model can infer attributes that are missing, reword awkward source copy, and adapt phrasing per category.

The platform treats this as a modifier like any other, so it slots into the same visual flow as the rules above.

Choose your model, or mix several

Feedoptimise connects to all major LLMs such as: OpenAI ChatGPT, Google Gemini, Anthropic Claude and xAI Grok, alongside our own hosted open models. Two things make this practical at feed scale rather than a novelty:

You are not locked to one provider. Each has a dedicated Title Optimiser modifier (OpenAI Title Optimiser, Google Gemini Title Optimiser, Anthropic Claude Title Optimiser, xAI Grok Title Optimiser) plus the hosted AI Title Optimiser. Swap the model without rebuilding the rule, and pick the one that gives you the best titles for the price.

You can chain different models in one flow. Because each prompt interface is its own modifier, you can use one model to generate, a second to check, and a third to translate. A worked sequence:

1. Anthropic Claude Prompt rewrites the title from the product attributes.
2. Google Gemini Prompt rates that title from 1 to 10 and returns only the number.
3. An IF modifier keeps the AI title where the score clears your threshold, and falls back to the rules-based title where it does not.

That is a quality gate built from two providers, assembled visually, running across the whole catalog.

Full control over the prompt

The prompt is not a black box. A WYSIWYG prompt template editor lets you drop your feed attributes straight into the instruction, so the model works from real product data rather than guesswork. A workable title prompt:

Write a Google Shopping product title, up to 150 characters,
and use the space. Put the attributes that drive matching and
clicks in the first 70 characters, then add relevant search
keywords after that. Lead with brand, then product type, then
the most important attributes. Use this order for apparel:
brand, gender, type, color, size. Front-load the highest-CTR
term. Keep each variant title unique. No promotional words.
No ALL CAPS. Return the new title only.

Brand: {{brand}}
Type: {{product_type}}
Color: {{color}}
Size: {{size}}
Description: {{description}}

Fill the gaps the rules can't reach

The AI modifiers do more than rewrite. The AI Attributes Extractor pulls attributes like color, material and gender out of the description when they are missing as fields. The Image Attributes Extractor (available on OpenAI and Gemini) reads the product image and returns the same, so you can recover the brand from a label in a photo. The AI Product Highlights modifier extracts the standout features from the description, which you can feed into the title for the attributes that win clicks.

So the AI route fixes the exact weakness of the rules route: it manufactures clean, structured attributes where your source data has none, then writes them into a title.

Enrich titles with real search-term data

Google's own advice is to add high-impression search terms to titles, broad terms to win impressions ("patio", "shoe") and granular terms to win clicks ("wicker", "Energy Star"), using sources like Google Trends, the Search Terms report and Keyword Planner. The trick is feeding that data into the title rule rather than guessing.

This is where the platform does something a plain rules build or a generic AI prompt cannot. The Report modifier pulls data from Google Ads and Google Analytics, and the Analysers surface what is performing, so you can route real impression and click data into the title flow. Identify the terms shoppers actually use, then either map them into the rules-based structure or pass them to the AI prompt as the keywords to feature. You enrich titles from evidence, not assumption.

Cost and testing

Two features keep this affordable and measurable.

• Caching. Feedoptimise caches the AI responses and only refreshes them when your settings or the underlying product data change, so you are not paying the model to regenerate the same title on every feed run.
• A/B testing. Feedoptimise lets you generate two title variants, a rules-based one and an AI one, or two different AI prompts, then schedule an A/B test that surfaces the version performing best on the channel. You optimize on results rather than opinion.

Which method, and when

• Use rules-based concatenation when your attribute data is clean and complete. It is cheaper, fully predictable, and trivial to audit.
• Use AI when source titles are weak, attributes are missing, or you want phrasing that tracks how shoppers actually search.

In practice the strongest setup combines them. Extract and clean attributes with AI, concatenate them with rules for a predictable structure, validate the length with Truncate and Text counter, then A/B test the rules-based title against a fully AI-written one to see which the channel rewards. Because every step lives in the same visual flow, switching between, or blending, the two methods is a matter of adding a modifier, not rebuilding a pipeline.

Getting started

If you already run feeds through Feedoptimise, open the modifier library and build the rules-based title first. It gives you a reliable baseline in minutes. Then layer an AI Title Optimiser on top, point it at a single model to start, and A/B test the two against each other. Let the channel data decide which wins, per category if you want to go further.

If you are not on the platform yet, this is the kind of work the visual rule builder and the multi-LLM modifiers were made for. You can try it on your own catalog with the free trial.