Optimizing Your Merchant Center Product Feed: Titles, Descriptions and Attributes

by Francis Rozange | Apr 4, 2026 | Google Ads

optimiser-flux-produit-merchant-center-1

Introduction: the feed is the campaign

In Shopping and Performance Max, the campaign settings barely matter. The bid strategy moves the needle a little. The audience signals nudge the algorithm a few percent. The budget caps the ceiling. None of these decide who wins. The product feed does.

Google’s Shopping algorithm and the Performance Max retail engine read your feed line by line, attribute by attribute, and decide for each query which products deserve to show up, at what position, and against which competitors. A weak feed loses auctions before the bid strategy ever has a chance to act. A strong feed wins auctions you did not even know you were eligible for.

This guide goes deep on the mechanics: the title formula that front-loads the right keywords, the 150-character ceiling versus the 70-character truncation reality on mobile, the description sweet spot between 1000 and 1500 characters, the difference between product type and Google product category, the GTIN/MPN/brand triangle, the variant attributes that drive Performance Max segmentation, the five custom labels that let you bid by margin tier, supplemental feeds for sale prices, image standards, structured data on the source site, and a feed audit playbook with weekly and monthly cadence. We cite Search Engine Land, Tinuiti, Store Growers, DataFeedWatch, and Google Merchant Center Help inline.

Why feed quality is the highest-leverage variable in Shopping and Performance Max

You can A/B test ad copy in Search for weeks and lift CTR by ten percent. That is a respectable gain. You can rewrite a single batch of product titles and lift Shopping CTR by sixty percent in a fortnight. DataFeedWatch reports a 31% average CTR uplift across audited accounts after a single round of title rewrites, and the upper quartile clears 60%. The leverage is not subtle.

The reason is structural. In Search, Google decides eligibility from your keywords and Quality Score, then runs an auction on the bid. In Shopping and Performance Max, Google decides eligibility from your feed itself: titles, descriptions, attributes, image, structured data on the landing page. Search Engine Land calls the feed “the keyword list of Shopping”. There is no manual keyword bidding in Shopping. The feed is your keyword list.

Performance Max amplifies this. The asset group’s product set, audience signals, and creative all funnel into the same engine, but the engine’s primary fuel remains feed data. Tinuiti’s audit of 47 retail PMax accounts found that feed quality explained 64% of the variance in account-level ROAS, more than budget, audience signals, or creative combined. The conclusion is uncomfortable for agencies that sell creative: in retail PMax, the feed is the campaign.

The corollary: a poorly optimized feed cannot be saved by a better bid strategy. Store Growers documents accounts where switching from Manual CPC to Maximize Conversion Value lifted ROAS by 12% on a clean feed and by 2% on a sloppy feed. The bid strategy can only optimize within the eligibility envelope the feed defines. Widen the envelope and every other lever gets more powerful.

The title formula: brand, product type, key attributes, size and color

The product title is the single highest-leverage field in your feed. Google supports up to 150 characters, but the practical reality is harsher: on mobile Shopping ads and most Performance Max placements, only the first 60 to 70 characters render. Everything beyond that is for Google’s matching algorithm, not for the customer.

This split between 70 visible characters and 150 indexable characters is the central design constraint of title writing. The first 70 characters must convince a human to click. The next 80 characters must give Google the keyword surface it needs to match more queries. Most merchants get one or the other right. The winners get both.

The canonical title formula by category

The formula adapts to category, but the skeleton is consistent: front-load the most distinctive identifier, then add the modifiers that filter intent.

Apparel: Brand + Gender + Product Type + Color + Size + Material. Example: “Levi’s Men’s 511 Slim Jeans, Dark Wash, 32×34, Stretch Denim”.
Electronics: Brand + Model + Product Type + Key Spec + Variant. Example: “Sony WH-1000XM5 Wireless Headphones, Active Noise Cancelling, Black”.
Home goods: Brand + Product Type + Key Attribute + Size + Color. Example: “KitchenAid Artisan Stand Mixer, 5-Quart Tilt-Head, Empire Red, 325W”.
Beauty: Brand + Product Line + Product Type + Variant + Volume. Example: “L’Oreal Paris Revitalift Hyaluronic Acid Serum, 30ml, All Skin Types”.
Hard goods and tools: Brand + Product Type + Key Spec + Model. Example: “DeWalt 20V Max Cordless Drill, 1/2-inch Chuck, DCD791D2 Kit”.

Notice the pattern: the brand anchors the search, the product type confirms category match, and the discriminators (size, color, model, capacity) close the loop with the searcher’s specific need. Search Engine Land’s analysis of 12 million Shopping impressions shows that titles starting with brand outperform brand-last titles by 23% on branded queries and underperform by 8% on category queries. If 70% of your traffic is non-branded, lead with category. If 70% is branded, lead with brand. Most retailers should split-test the lead.

Front-loading: the first 30 characters carry the weight

Google’s matching algorithm weights tokens by position. The first word counts more than the second. The first thirty characters carry roughly 60% of the title’s matching power, according to DataFeedWatch’s title-position study across 8 retail verticals. The implication is brutal: if your distinctive keyword sits at character 90, it might as well not exist for matching purposes.

The fix is mechanical. Take any title and ask: if I cut this at character 30, do the remaining tokens still identify a unique product type and brand? If yes, the front is loaded correctly. If no, reorder. “Comfortable, breathable, lightweight running shoes from Nike for men in black, size 10” fails this test catastrophically. “Nike Air Zoom Pegasus 40 Men’s Running Shoe, Black, Size 10” passes.

The 150-character extension: keyword surface for matching

Once the first 70 characters carry the click decision, the remaining 80 characters are pure matching surface. This is where you add the secondary discriminators: material, use case, intended audience, certifications, technical specs. Tinuiti recommends packing the tail with the modifiers that surface in Search Console for the product page: if customers find the page on Google with “vegan running shoe”, and the product is in fact vegan, the tail of the title should say so.

Avoid keyword stuffing. Google detects repetition and demotes feeds that game the field. Two appearances of the same noun is the practical ceiling. “Nike Pegasus running shoe for runners who run” reads exactly as machine-generated as it is.

Description structure and length: the 1000 to 1500 character sweet spot

Descriptions allow up to 5000 characters, but the data is unambiguous: the optimal length sits between 1000 and 1500 characters. DataFeedWatch’s analysis across 2.3 million product descriptions found that descriptions in the 1000 to 1500 character band had a 17% higher conversion rate than descriptions under 500 characters and a 9% higher conversion rate than descriptions over 3000 characters.

The reason is twofold. Below 500 characters, you do not give Google enough text to extract semantic signals: material, intended use, dimensions, compatibility, audience. Above 3000, you dilute the signal with marketing prose that the algorithm discounts and that no customer reads. The 1000 to 1500 window forces you to be specific without being padded.

Description structure that the algorithm rewards

The structure that consistently outperforms in matched-pair tests follows a five-block pattern. Store Growers documents this pattern across audited accounts:

Block 1, identification (first 30 characters): “Nike Air Zoom Pegasus 40 men’s road running shoe.”
Block 2, distinguishing features (next 200 characters): the technical attributes that separate this product from category alternatives.
Block 3, materials and construction (next 300 characters): upper, sole, lining, hardware, fastenings, with named technologies.
Block 4, use case and audience (next 300 characters): who it is for, when to use it, what problems it solves.
Block 5, variant specificity (next 200 characters): exact color, size, weight, drop, fit, and any details that vary across SKUs.

The first block matters disproportionately. Google truncates descriptions at roughly 150 to 200 characters in the expanded product card, so the first sentence has to identify the product completely. After that, you are writing for the algorithm.

Specificity beats marketing language

“Eco-friendly” is a wasted token. “Made from 87% recycled polyester, GRS-certified” is matchable, indexable, and credible. Search Engine Land’s experimentation on 200 SKUs across three retailers found that descriptions with quantified specifications (dimensions, weight, material percentages, compatibility versions) converted at 1.4x the rate of descriptions with generic adjectives. Numbers and named standards ground the description in verifiable claims and signal authenticity to the matching algorithm.

Product type versus Google product category

Two attributes compete for attention here, and confusing them is one of the most common feed errors. They are not interchangeable.

Google product category is a fixed taxonomy maintained by Google. It currently lists roughly 6000 nodes, organized as a hierarchical tree from “Apparel and Accessories” down to “Apparel and Accessories > Clothing > Activewear > Bicycle Activewear > Bicycle Jerseys”. You pick the deepest applicable node. Google uses this to enforce category-specific policies and to constrain matching to the right vertical. If you submit a swimsuit under “Footwear”, you will not show up for swimsuit queries no matter how good your title is.

Product type is your own taxonomy. You define the values. Google uses it as a secondary signal for matching and, critically, as a campaign segmentation lever. You can target Shopping campaigns or PMax asset groups by product type without restructuring your feed. The value is up to you, but the convention is to use a slash-delimited path that mirrors your site’s category structure: “Men’s Apparel > Outerwear > Down Jackets > Lightweight”.

The error to avoid: using Google product category as your campaign segmentation. It is a fixed taxonomy, you cannot bid differently on its sub-nodes, and using it for segmentation locks you into Google’s view of your catalog. Use product type for segmentation, Google product category for eligibility.

Product type for PMax segmentation

Performance Max asset groups can target by product type path. This is the underused superpower of the attribute. Tinuiti recommends structuring product type as a path that encodes both category and a segmentation dimension. For a multi-brand retailer: “Apparel > Brand-A > Outerwear > Down Jackets” lets you target an asset group by brand and category in a single rule. For a margin-based segmentation, you can append a margin tier: “Apparel > High-Margin > Outerwear > Down Jackets”. This works, but it pollutes the taxonomy. Custom labels are cleaner for margin segmentation.

GTIN, MPN, and brand: the identifier triangle

For products manufactured by a third party (the majority of retail SKUs), Google requires two of three identifiers: GTIN (UPC, EAN, JAN, ISBN), MPN (manufacturer part number), and brand. The official rule: branded products with a known GTIN must submit it. Branded products without a GTIN must submit MPN and brand. Private-label products without a GTIN or MPN can use the identifier_exists attribute set to “no”.

The GTIN matters disproportionately. Google’s own published data shows that products with valid GTINs see up to 20% higher conversion rates than identical products without. The reason is the matching algorithm: with a GTIN, Google can confidently cross-reference your product with manufacturer specifications, customer reviews aggregated across retailers, and price comparison signals. Without it, Google has to guess.

The implication for retail accounts is direct: if your product database does not store GTINs, fixing that is the highest-ROI feed project you can run. Most enterprise PIM systems support GTIN as a first-class field. If you sell on Amazon, you already have GTINs in the Amazon backend. DataFeedWatch reports cases where adding GTINs to a 5000-SKU catalog over two weeks lifted Shopping ROAS by 18% within the following month, with no other changes.

Brand: include even for private label

Brand is required for products with a known manufacturer. For private-label products, Google accepts your store name as the brand value. Some retailers leave brand blank for private label, hoping to avoid associating the product with a less-known name. This is a mistake. A blank brand reduces matching confidence. Use your store name. Search Engine Land documents private-label retailers who saw 11% impression share gains after populating brand with their store name across the catalog.

Variant attributes: color, size, material, age_group, gender

Variant attributes are the dimensions on which a product splits into multiple SKUs. Google requires each variant to ship as a separate feed entry with its own ID, but tied together by an item_group_id that signals “these SKUs are variants of one product”.

Color and size: the apparel non-negotiables

For apparel, color and size are mandatory and create the most common variant explosion. A T-shirt in 3 colors and 6 sizes is 18 feed entries, not 1 with 18 variants. Each entry has its own GTIN, its own image (ideally showing that specific color), its own availability, and its own price.

The discipline that separates good apparel feeds from bad ones is image-per-color. Submitting the same image for all 3 colors of a T-shirt is a missed opportunity that DataFeedWatch quantifies at a 14% CTR penalty. The customer searching for “blue T-shirt” needs to see a blue T-shirt in the ad thumbnail. Generic catalog photography on a white background works for one color; for the others you need a per-color shot.

Material, pattern, age_group, gender

Material is required for jewelry and recommended for apparel and home goods. The convention: name the dominant material and percentage if you know it. “100% cotton”, “80% wool, 20% polyamide”, “stainless steel 316L”. Vague values like “fabric” or “metal” trigger data quality warnings.

Pattern is recommended for apparel: “solid”, “striped”, “floral”, “checked”. It is a small attribute but it surfaces in many fashion queries.

Age_group has five enumerated values: newborn, infant, toddler, kids, adult. For children’s apparel, this attribute is required and disambiguates from adult apparel of similar size.

Gender takes three values: male, female, unisex. For non-gendered products (most home goods, electronics), omit the attribute rather than submit “unisex” by default. For apparel, gender is required.

Custom labels: five slots for bid-strategy segmentation

Custom labels (custom_label_0 through custom_label_4) are five free-text fields that do nothing for matching but everything for campaign segmentation. They are the merchant’s escape hatch from Google’s taxonomy. Store Growers’ canonical guide proposes a slot allocation that has become an industry default and is worth treating as a starting point.

The recommended slot allocation

custom_label_0: margin tier. Values: high (over 40% gross margin), mid (20% to 40%), low (under 20%), loss-leader. This lets you set higher target ROAS on high-margin SKUs and lower target ROAS on loss-leaders that pull traffic.
custom_label_1: performance bucket. Values: top-seller, mid-seller, slow-mover, new-arrival. Recompute weekly from the last 30 days of conversion data. Lets you bid up top-sellers and protect new arrivals from being starved by Smart Bidding’s preference for proven SKUs.
custom_label_2: seasonality. Values: spring, summer, fall, winter, all-year, holiday. Lets you build seasonal campaigns and pause out-of-season SKUs without restructuring.
custom_label_3: price tier. Values: under-25, 25-100, 100-500, over-500. Useful for free-shipping segmentation and for matching bid strategy to AOV.
custom_label_4: promotion or campaign hook. Free text for ad-hoc campaign tags. “blackfriday-2026”, “clearance-q4”, “newsletter-feature”.

The discipline: pick a convention, document it, recompute the dynamic labels (margin, performance bucket) on an automated schedule, and never let an analyst hand-edit individual SKU labels. The moment custom labels drift from the convention, your segmentation breaks silently.

Bid-strategy segmentation in PMax

Performance Max asset groups can filter the product set by custom label. Tinuiti’s PMax playbook recommends running parallel asset groups: one for high-margin SKUs at a higher target ROAS (say, 600%), one for mid-margin at 400%, one for loss-leaders at 200%. The asset groups share creative themes but pursue different profitability ceilings. The data shows this beats a single asset group with a blended ROAS target by 12% to 18% in account-level profit, even when revenue is flat.

Supplemental feeds: sale prices and seasonal overrides without breaking the primary feed

The primary feed is your source of truth. Touching it for every promotion is dangerous: a botched sale price update can corrupt 5000 prices in one upload. Supplemental feeds solve this. A supplemental feed updates a subset of attributes for a subset of SKUs without touching the primary feed.

Use cases for supplemental feeds

Sale prices: the sale_price and sale_price_effective_date attributes communicate a discounted price with start and end timestamps. Google displays the strikethrough price in Shopping ads if the discount is at least 5% and lasts at least 30 days. Push sale_price via supplemental feed, not via the primary feed, so a sale rollback is one feed delete instead of a full primary re-upload.
Custom label overrides: recompute custom_label_1 (performance bucket) weekly via a supplemental feed. The primary feed stays clean.
Seasonal title and description boosts: append “Christmas Gift Guide 2026” to the title for Q4 via a supplemental feed, then remove in January with another upload.
Inventory overrides for click-and-collect: append local availability for stores via the local inventory feed, which is structurally a supplemental feed.

DataFeedWatch’s case study on a fashion retailer shows that moving sale price updates from the primary feed to a supplemental feed reduced disapproval incidents from 4 per month to 0.3 per month and cut average promotion-launch lead time from 24 hours to 90 minutes.

Image requirements: white background versus lifestyle, by category

Google Merchant Center publishes strict image policies: minimum 100×100 pixels for non-apparel and 250×250 for apparel, no logos or watermarks, no promotional text overlay, no borders. Below the policy floor, the practical question is: white background or lifestyle?

The answer depends on category. For most non-apparel categories (electronics, hard goods, beauty, kitchen), the canonical image is a clean catalog shot on white background, 800×800 or larger, with the product centered. This is what Google’s matching algorithm expects, and it is what customers compare across retailers in the Shopping grid. Search Engine Land reports that for electronics, white-background images outperform lifestyle by 22% on CTR.

For apparel and home decor, the relationship inverts. Lifestyle imagery, the product in context (a model wearing the jacket, a sofa in a styled living room) outperforms white-background by 11% to 19% on CTR. The customer is buying a feeling, not a spec. The catalog shot looks sterile in the Shopping grid surrounded by competitors’ lifestyle shots.

The compromise that works for both: primary image lifestyle, additional_image_link populated with white-background catalog shots and detail crops. Google rotates additional images in some placements and uses them for matching even when they are not displayed.

Resolution and aspect ratio

Google’s stated minimum is 100×100 (250×250 for apparel), but the practical floor is 800×800. Below that, images render fuzzy on high-density mobile screens. For Performance Max, the algorithm benefits from 1200×1200 or larger, which gives the AI room to crop for various placement aspect ratios. Tinuiti’s PMax audits consistently find that accounts uploading 1200+ pixel images outperform accounts at 800 pixel by 8% to 12% on conversion rate, attributed to better cropping in the Discover and Gmail placements.

Structured data on the source site: JSON-LD Product schema feeding the auto-extract

Google Merchant Center can auto-extract product data from your website using schema.org Product structured data embedded as JSON-LD. This is no longer an exotic technique. For Shopping Free Listings and for Merchant Center automated feeds, schema is now the primary ingestion path.

The minimum viable Product schema

The minimum viable Product schema includes name, image, description, sku, brand, offers (with price, priceCurrency, availability, priceValidUntil), and aggregateRating if available. Google’s Product schema reference lists the full schema, but in practice the fields above carry 90% of the matching weight.

The field that retailers most often miss: priceValidUntil. Without it, Google flags the offer as ambiguous and demotes the listing in some surfaces. Set it to a rolling 30-day future date, regenerated nightly.

The other commonly missed field: gtin13 (or gtin12, gtin8, gtin14 depending on barcode type). Schema GTIN should match feed GTIN exactly. Mismatches between feed GTIN and schema GTIN trigger reconciliation warnings and can suppress the listing entirely.

Schema as the canonical source

The most robust architecture treats the schema on the product page as the canonical product representation, with the feed generated from the same data layer. This eliminates the most common feed problem: feed and site drift. When the marketing team updates a product description on the site, the schema updates, and the feed regenerates from the schema. DataFeedWatch’s case study on a 12000-SKU retailer shows feed disapprovals dropping from 8% to 0.6% after consolidating site, schema, and feed on a single product data layer.

Feed quality and Performance Max ROAS: the correlation data

The case for feed investment in PMax is now quantitative. Tinuiti’s 47-account audit measured five feed-quality dimensions on a 100-point scale: title quality, description completeness, attribute coverage, image quality, and identifier validity. Accounts in the top quartile of feed quality averaged ROAS of 6.8x. Accounts in the bottom quartile averaged 2.4x. The correlation coefficient between feed-quality score and account ROAS was 0.71, an extraordinary number for a single-variable analysis in advertising.

Optmyzr’s parallel analysis on 230 PMax accounts found a similar effect, with feed quality explaining 58% of variance in cost per acquisition after controlling for budget and category. The implication: if you have to choose between hiring a PMax specialist and hiring a feed analyst, the feed analyst is the higher-ROI hire on a typical retail account.

The compounding effect over time

Feed quality compounds. DataFeedWatch’s longitudinal study on 88 accounts that ran a six-week feed improvement program found average ROAS uplift of 22% in month 1, 38% by month 3, and 51% by month 6. The reason: Smart Bidding accumulates conversion data on the cleaner feed, learns better, and bids more accurately. The feed improvement does not just add to revenue, it improves the rate at which Smart Bidding learns. The early gain compounds.

The feed audit playbook: weekly and monthly cadence

Feed maintenance is not a one-off project. It is a discipline with a cadence. The playbook below is what high-performing retail teams run.

Weekly tasks

Monday morning, disapproval review. Open Merchant Center Diagnostics, sort disapprovals by item count, fix the top three issue types. Most weeks this is GTIN mismatches, image issues, or price discrepancies. A 30-minute weekly habit prevents the disapproval backlog that cripples Q4.
Recompute dynamic custom labels. Pull last-30-days conversion data, recompute custom_label_1 (performance bucket) for the catalog, push via supplemental feed. Automate this if you can; manual is fine if catalog is under 2000 SKUs.
Spot-check ten random products. Pick ten SKUs at random, open the feed entry and the live product page side by side. Read both titles, both descriptions, both prices. Mismatches you spot manually are mismatches the algorithm will spot too.

Monthly tasks

Title rewrite for the bottom decile. Pull the bottom 10% of SKUs by impression share. Audit titles for keyword presence, front-loading, and category match. Rewrite. Most accounts find that the bottom decile contains structural title problems, not pricing or relevance issues.
Identifier coverage audit. Report the percentage of SKUs with valid GTIN, MPN, brand. The target is 100% for products with manufacturers. Track the trend month over month.
Image quality audit. Spot-check the bottom decile by CTR for image issues: low resolution, watermarks, white-background where lifestyle would win, generic stock images.
Custom label allocation review. Verify that the slot conventions still match business reality. Margin tiers may have shifted. Performance buckets may have aged.
Search terms report against title vocabulary. Pull the search terms report from PMax and Shopping. Identify high-converting terms that do not appear in your title vocabulary. Inject them where natural.

Quarterly tasks

Full Google product category audit. Verify category assignments against the latest taxonomy, which Google updates twice a year.
Schema audit on the source site. Crawl the site, validate JSON-LD against schema.org Product, fix mismatches with the feed.
Supplemental feed cleanup. Retire stale supplemental feeds. Old promotional feeds left running create silent overrides.

Common mistakes that drain Shopping and PMax budgets

Treating the feed as a static export. The feed exports nightly from your e-commerce platform with no review, no enrichment, no rewriting. The catalog leaves the platform with terrible titles (“Product 4729-blue”) and arrives at Google unchanged. Performance reflects what you submit.

Putting promotional language in titles. “SALE”, “FREE SHIPPING”, “BEST PRICE” in the title triggers policy violations and disapprovals. Use the sale_price attribute and the promotion feed instead.

Identical title and description. If your description starts by repeating the title verbatim, you wasted 70 characters. The description should expand, not echo.

Submitting one image for all color variants. The customer searching for “red dress” needs to see a red dress, not the black one because that was the catalog hero shot.

Custom labels populated by hand and never refreshed. An analyst tags 800 SKUs as “high-margin” in February and forgets. By August, 200 of them have been repriced down. The labels lie.

Mixing Google product category and product type semantics. Using Google product category for segmentation locks you into Google’s taxonomy. Using product type for eligibility leaves Google to guess your category.

Feed and site drift. Marketing updates the description on the site for a launch. The feed exports the old description. Google’s mismatch detection flags the listing. Disapproval, then a two-week recovery.

Ignoring schema on the source site. Schema is not optional anymore. Free Listings and Merchant Center automated feeds depend on it. Skipping schema is leaving impressions on the table.

Skipping GTINs because the PIM does not store them. The 20% conversion lift Google documents is not theoretical. Skipping GTINs costs more than the work to add them.

Conclusion: the feed is the merchandiser of the algorithmic shelf

Google Shopping and Performance Max do not show your products to people. They show your feed to people. The catalog photo on Shopping is your image_link. The headline a customer reads is your title. The detail page snippet is your description. Every visible element of the customer’s encounter with your brand on Google’s shelves comes straight from your feed.

Treating the feed as IT plumbing is the most expensive misallocation of attention in retail digital marketing. Treating it as merchandising, with the same care a store manager gives to the front window, is the cheapest path to better Shopping and PMax performance. The work is unglamorous: title rewrites, GTIN backfills, custom label conventions, image audits, schema validation, weekly disapproval cleanup. None of it photographs well for an agency case study. All of it shows up in the ROAS line.

Start with the highest-leverage move. Audit the bottom decile of titles. Add GTINs where missing. Set up your five custom labels with a documented convention. Move sale prices to a supplemental feed. Validate JSON-LD schema on your top 100 SKUs. That sequence will lift account-level ROAS by 15% to 30% in most retail accounts within 90 days, with no other change.

The feed is the campaign. Run it like one.