How do we test a product feed change (new titles, images or attributes) on Google Shopping and Performance Max: what control, how long, and which metric shows it worked?
Our answer
Change one thing on one group of products and leave a matched group as the control. Once Merchant Center shows the new values, run about four weeks plus 7 days for late orders, judged on conversion value or profit.
- Confidence
- Low
- Strongest proof
- 1 controlled test
- 8 Google rules, 1 dataset, 7 practitioners, 5 opinions or claims
- Sources
- 12 sources
- 22 statements
- G
PMZ
P
New here? Take the 1-minute tour
Why this is the answer
3 reasons and our judgmentA matched group of unchanged products is the control. Product Metrics and Productsup test a change on one group against a matched group left alone, one variable at a time. A before and after read also picks up season, price and stock (Product Metrics).
Feed changes show slowly. Merchant Center recrawls in 1 to 3 days, then results take 7 to 14 days on Standard Shopping, 14 to 30 on Performance Max (Adgenix). Common Thread Collective runs AI Max tests four weeks. WPMarketingRobot adds 7 days for late orders.
Clicks can rise while sales don't. Product Metrics reads impressions and clicks, then conversions, then contribution after ad cost, and its example shows a title adding clicks that don't buy. Common Thread Collective judges AI Max tests on ROAS. Google says choose the metric first.
Our judgment
- Split products, not traffic, into test and control groups, and tag each group with a custom label the feed change doesn't touch. Google's custom labels are built for monitoring and reporting, but attribute changes affect product group targeting. The only Google experiment here that tests the product feed is the Demand Gen one, and Google's Experiment Center splits users or geography, not products.
- Make one change through a supplemental data source that keeps each product's ID, and change nothing else. That keeps the control untouched, the test reversible and each product's history intact (Google). ZenoX and WPMarketingRobot use a supplemental feed as the testing layer. ZenoX's jewelry account changed titles, labels, the campaign split and the ROAS target together, and couldn't credit the feed work.
- Treat about four weeks plus 7 days as the minimum for a sales read, and expect to need longer. Read clicks after at least two full weeks (Product Metrics), consistent with Product feed changes that pay's four weeks. Lily AI calls 28 days a defensible floor, and Optiversal's sales gains hadn't reached significance after about eight weeks while its clicks had.
- Decide on the test group's change in conversion value or ROAS against the control. Use contribution after ad cost where margins are known, and treat impressions and click-through rate as the early read. Jetfuel Agency sees impression share move within 3 to 7 days of a recrawl. This matches Setting up and reading a Google Ads experiment.
- When a catalog is too small for a matched control, use a before and after read as direction only. Run WPMarketingRobot's 30 days before and 30 after on one group of 20 to 50 products, with no sale or promotion in either window, and check the same weeks a year earlier as Hustle Marketers does.
Our judgment, built from Product Metrics, Productsup, smec (Smarter Ecommerce), Common Thread Collective, Google, ZenoX, WPMarketingRobot, Adgenix, Jetfuel Agency, Hustle Marketers, Lily AI and Optiversal: our reading of these sources, not a finding from any one of them.
What the sources say
22 statements from 12 sourcesGoogle's Experiment Center splits an eligible audience by users or geography and lists no product data or feed experiment type, so a feed change has to be tested by splitting products yourself.
“split an eligible audience into two groups based on users or geography”
Google warns that editing product IDs, titles, descriptions or identifiers can make Shopping performance fluctuate, and attribute changes can affect product group targeting. It gives no duration.
“Editing ID, titles, descriptions and/or unique product identifiers may also cause performance fluctuations”
Google Merchant Center says a supplemental data source can only add or update attributes for existing products, matched by exact product ID; only primary sources can add or remove products.
“Primary data sources are the only data source type that can add or remove products.”
Google's Merchant Center help tells merchants to run experiments on product data, such as adding color to apparel titles, and keep each product's ID constant to preserve its history. No design or results given.
“Using the same, consistent ID [id] with each product will preserve its historical data.”
Product Metrics says to test a new Shopping title on one product group against a similar unchanged group over the same weeks, judging orders and contribution after ad cost. Its 40-product example is illustrative.
“Change one title variable at a time.”
Google lets Shopping and Performance Max campaigns split products with up to 5 custom labels (such as season, margin or clearance), one value per product each, for reporting and bidding.
“Each custom label can only have one value per product.”
Google's Demand Gen A/B experiments test creative, audience, feed or bidding, not budget, with a 50% split. Results need at least 50 conversions per arm, so Google suggests shallow goals like Add to Cart.
“We don't recommend testing the budget as a variable at this time.”
Google says that from June 2026 product reporting covers all Performance Max networks plus Video, App and Demand Gen campaigns, which may cause a one-time jump in Performance Max impressions and clicks.
“This may cause a one-time increase in metrics such as impressions, clicks, and more for Performance Max campaigns only.”
Google's experiments advice: set a clear hypothesis, test one variable at a time, pick one or two success metrics in advance, avoid changing base campaigns unless sync is on, and keep records.
“When it’s time to end a test, let that metric tell you the winner.”
Product Metrics, which sells product segmentation software, advises testing a Shopping feed change on one input type in a defined product group; a before/after comparison does not isolate a title change.
“A simple before/after comparison does not isolate the effect of a title change.”
Hustle Marketers, which sells feed audits, advises testing feed changes in blocks, not per SKU: change one product type and compare with the prior period and the same period last year. No data.
“Test in blocks, not per SKU.”
Optiversal, a feed vendor, ran an 8-week product split: clicks rose 12.4% and CPC fell 3.5%, both significant, while conversions, value and ROAS rose but were not confirmed. How control was chosen is unstated.
“Not every conversion and ROAS metric reached statistical significance inside the test window, but directional movement was positive across every major indicator”
Lily AI says it has run its feed test design over 1,000 times: a 28-day holdout against an untouched control with spend and all else frozen. Its example showed +28% Shopping revenue, without product counts.
“Yes, but only when measured against a control.”
Jetfuel, which sells feed audits, says product titles are the most valuable Shopping feed fix (front-load the main search term), missing GTINs cost impression share and ratings, and changes show in 3 to 7 days.
“Product title structure is the single most valuable feed optimization. Front-load the most relevant search term.”
Common Thread Collective suggests recording a baseline when Google's product click columns appear, auditing product click rate for 30 days, and testing AI images one variant per campaign for 2 weeks. No data.
“Test one variant per campaign over 2 weeks.”
CTC, an agency, relays Google's claim that AI Max for Shopping averages 5% more conversions at similar cost, and advises filling every feed attribute and testing four weeks against a control.
“Run a four-week test against a control campaign and measure conversion rate and ROAS at the campaign level before expanding.”
Productsup, which sells the tool, announced feed experimentation that splits eligible products 50/50 into two variants and advises testing one variable at a time. No results or test length are given.
“Testing one variable at a time helps isolate what is actually driving performance improvements.”
ZenoX, which sells Shopping feed management, says a supplemental feed in a Google Sheet is free and still the best way to test title changes, since each change can be rolled back by deleting rows.
“A supplemental feed in a Google Sheet is free, fully in your control, and still the best testing layer for title changes.”
WPMarketingRobot, which sells a feed plugin, tests Shopping titles on 20 to 50 products: 30 days on old titles, then 30 on new. It calls the result directional, not a controlled experiment.
“Treat the result as a directional signal, not a controlled experiment.”
Adgenix, which sells Shopping audits, says feed overhaul effects take 7 to 14 days in Standard Shopping and 14 to 30 in Performance Max, after a 1 to 3 day recrawl. These are stated timings, not test results.
“Performance changes from improved targeting: 7-14 days for Standard Shopping, 14-30 days for PMax (PMax needs longer to recalibrate).”
ZenoX, which sells Shopping management, reports one jewelry account spending €120k a month cut CPA 38% by week 9 after feed and ROAS target changes. It says the case does not separate feed work from bid changes.
“This case does not isolate feed work from bid changes.”
smec, which sells Shopping and Performance Max tools, says AI Max text customization is only as good as the feed and PMax spends on low-margin items with easy clicks; test it on one segment first.
“If your feed is messy, your AI-generated copy will be garbage.”
In the full answer
For readers who sign in- What to do this week3 steps
- The numbers behind it3 figures from cited sources
- Where each source standsA chart placing each source on the question
- The strongest disagreementThe best case against the answer, from a cited source
- How sure we areWhat its confidence rests on, and what would change the answer