In short
Group products by what should change the bid, not by how the catalogue is organised. Margin, price band, proven performance and seasonality are the useful divisions, and custom labels in the feed are how you express them. Keep the number of campaigns small enough that each one gathers sufficient conversion data for automated bidding to work, separate only what genuinely needs a different target, and give the long tail and new products a place to earn a test.
Segment by economics, not by taxonomy
The instinct is to mirror the site: a campaign per department, then per category, then per subcategory. It produces a tidy structure that answers the wrong question, because the reason to separate two products is that they need different bidding, and category membership rarely determines that.
What determines it is margin, price, conversion behaviour and how much you want to sell the item. A high-margin accessory and a low-margin appliance in the same department should not be bid the same way. Two products in different departments with the same margin and conversion rate should.
So start from the economics. Which products can afford an aggressive target, which need protecting, which are strategic even at a weak return, which are seasonal, which are new and unproven. Those groupings are stable, they map onto decisions you actually make, and they are expressible in the feed.
There is a second reason to resist a deep structure. Automated bidding needs conversion volume to work with, and a structure split forty ways divides the same conversions into forty thin pools. Fewer, better-populated campaigns generally outperform many precise ones on a large catalogue.
The custom label scheme that does most of the work
Five custom label slots are available in the feed, and a consistent scheme across them removes most structural problems. A workable allocation looks like this.
The first label carries margin tier, calculated from your own cost data and bucketed into three or four bands. This is the single most useful segmentation available and the one most stores never implement, because it requires joining cost data to the feed.
The second carries price band, since low-value and high-value items behave differently in the auction and often need different targets even at the same margin.
The third carries performance tier, refreshed on a schedule from your own data: proven sellers, occasional sellers, and products with no conversion history. This one changes over time and should be regenerated rather than set once.
The fourth carries seasonality or collection, so a summer range can be pushed and pulled without restructuring anything.
The fifth is best kept free for whatever the business needs this year: clearance, exclusives, a supplier deal, stock depth.
Generate these in the feed pipeline or a supplemental feed rather than by hand, and document what each slot means somewhere the whole team can see, because an undocumented label scheme becomes unusable within a year.
How many campaigns, and where the lines go
Fewer than instinct suggests. A large catalogue can usually be run well with a small number of campaigns divided by the tiers above, plus separations that are genuinely necessary rather than merely tidy.
The separations that earn their place are: products needing a materially different return target, products with different budget ownership within the business, markets or countries with different economics, and a distinct treatment for new or untested products so they are not starved by better-performing lines in the same pool.
Within a campaign, use listing groups to subdivide for reporting and, where the campaign type supports it, for targeting. That gives visibility without fragmenting the data that bidding relies on.
Decide deliberately how automated and manually structured campaign types coexist. Running several types over the same products changes which one serves, so either separate the products they cover or accept and monitor the overlap rather than discovering it in the reporting.
And leave room for the exceptions your business actually has: a supplier-funded push, a clearance block, a launch. Those should be expressible with a label change rather than a rebuild, which is the whole point of putting the segmentation in the feed.
The long tail, new products and the review routine
Every large catalogue has products that have never had a click. Left in a general campaign they stay invisible, because the system spends where it has evidence. Give them a defined place with its own budget, accept a weaker return there, and treat it as a test budget whose job is to generate evidence rather than profit.
New products need the same treatment for a limited period. Set a rule for how long an item is treated as new, what it is allowed to spend during that period, and where it moves afterwards based on what it did. That rule is what stops a catalogue calcifying around last year's winners.
Build the review routine around the labels. Weekly, look at spend and return by tier and by campaign rather than by product, and act on the tiers that have drifted. Monthly, regenerate the performance tier labels so products move between pools on evidence. Quarterly, revisit margin bands, because costs change and a scheme built on last year's margins misallocates money quietly.
And keep the feed diagnostics in the same routine. Structure cannot compensate for products that are disapproved, out of stock or missing attributes, and on a large catalogue those problems are continuous rather than occasional.
Related questions
Should each product category have its own campaign?
Only where categories genuinely differ in margin, target return or budget ownership. Splitting by category alone divides conversion data without changing any decision, which makes automated bidding less effective. Use listing groups within a campaign to get category-level reporting instead, and reserve separate campaigns for economic differences.
Is it worth separating brand and non-brand traffic in Shopping?
Standard Shopping campaigns give you negative keywords and campaign priority to influence this, and stores with strong brand demand often want to see the two separately because their returns differ so much. Whether it is worth the complexity depends on how much of your volume is brand searches, which your search terms report will show.
What should be done with products that never convert?
Give them a bounded test with a defined budget and a defined period, then decide. Some are genuinely unsellable through the channel and should be excluded. Others have a data problem, a poor image or an uncompetitive price, and the exclusion decision should come after checking those rather than before.