Shopping, Performance Max and YouTube

In a Search campaign, you choose keywords. In a Shopping campaign you choose nothing of the sort — Google reads your product feed and decides which queries each product should appear for.

That single difference is the whole discipline. Your product feed is your keyword list. Every hour spent adjusting bids on a poor feed is an hour spent optimising the delivery of a message Google cannot understand.

Most Shopping accounts I am asked to look at are bidding aggressively against a feed that prevents Google understanding half the catalogue, and the diagnosis is usually mistaken for a bidding problem because that is where the controls are.

The title is the highest-leverage field there is

Product titles carry more matching weight than any other attribute. A title that reads the way a manufacturer's internal SKU description reads will match almost nothing a customer types.

The pattern that works, adapted to the category:

Brand + Product Type + Key Attribute + Model/Size/Colour

• Weak: SKU-4471-BLK
• Weak: Premium Comfort Range — Now Available
• Strong: Nike Air Zoom Pegasus 41 Running Shoes — Men's, Black, UK 9

Front-load the important words. Titles are truncated in display, and matching weight skews toward the beginning.

The attributes that decide whether you appear at all

• GTIN — where one exists, supply it. Missing GTINs restrict eligibility for some surfaces and remove Google's ability to match your product to its own catalogue.
• Brand — required for most categories, and used heavily in matching.
• google_product_category — set it explicitly. Google will guess otherwise, and its guesses on ambiguous products are frequently wrong in ways that quietly suppress delivery.
• product_type — your own taxonomy. Google does not use it for matching, but it is the field you will use to build campaign structure, so populate it thoughtfully.
• availability and price — must match the landing page exactly. Mismatches cause disapprovals, and repeated mismatches cause account-level warnings.
• image_link — a clean product image on plain background. Promotional overlays and watermarks are a policy violation and get items disapproved.

The audit that finds most of it

Sort your feed by product-level impressions, ascending. Items with a great deal of stock and almost no impressions are usually not losing an auction — they are failing to enter one. Check those items' titles, categories and disapprovals first.

Merchant Center will show you a great many warnings. Most are noise. Three categories are worth acting on immediately, and in this order.

1. Disapprovals. An item that is disapproved is not selling. This is the only category that is unambiguously urgent. The frequent causes are price or availability mismatch against the landing page, missing GTIN in a category that requires one, image policy violations, and landing pages that fail Google's automated checks — commonly a broken SSL certificate or a page that redirects.

2. Limited performance warnings. Usually missing recommended attributes. These do not stop the item serving but they narrow where it can serve.

3. Item-level issues on high-value products. Triage by revenue rather than by warning count. Fifty warnings on discontinued lines matter less than one on your best seller.

Price and availability accuracy

This causes more disapprovals than anything else, and it is almost always a synchronisation problem rather than a data problem.

If your feed updates once a day and your prices change more often, you will be permanently out of sync. The fix is automatic item updates or a scheduled fetch frequent enough to match how often your prices actually move. For a store running frequent promotions, a daily feed is not sufficient.

Do not neglect the free listings

Free product listings across the Shopping tab and other surfaces use the same feed. Feed work therefore pays twice — once in paid performance, once in unpaid visibility. This is a genuinely underrated argument for spending time on the feed rather than the bids.

Performance Max runs one campaign across Search, Shopping, Display, YouTube, Discover, Gmail and Maps, with Google allocating between them. It absorbed Smart Shopping in 2022 and became, for a period, the least transparent campaign type Google had ever shipped.

That criticism is now substantially out of date, and if your mental model of PMax was formed in 2022 or 2023, it is worth rebuilding.

What 2026 changed

• Negative keywords at campaign level. This was the single largest complaint about PMax for years and it is now resolved — campaign-level negative keyword support, with a substantially raised limit, plus account-level negative lists.
• Channel-level reporting. You can now see performance broken down by Search, Display, YouTube, Discover and Maps, rather than one blended number. This is what makes PMax diagnosable at all.
• Search terms visibility for search-originated traffic, and placement reports for Display and YouTube inventory.
• Search themes doubled to 50 per asset group.
• Brand exclusions at campaign level, so PMax can be prevented from serving on your own brand queries.
• Asset-level performance ratings and audience insights showing who is converting.

The practical consequence is that the old advice — "run PMax but keep a Standard Shopping campaign alongside so you can see what is happening" — is a weaker argument than it was.

Working with it

Asset groups are your structure. Treat them the way you would treat ad groups: one per theme that deserves its own creative and its own landing page. A single asset group covering the whole catalogue gives Google no way to vary the message.

Search themes are direction, not keywords. They tell the system what queries you consider relevant. They do not restrict it to those queries.

Feed quality still decides retail performance. Everything in the previous section applies unchanged. PMax does not rescue a bad feed; it obscures one slightly less than it used to.

Exclude your brand unless you want it. Without brand exclusions, PMax will happily serve on people searching your name, absorb those cheap conversions, and report excellent blended results — the same flattery problem as mixing brand into a generic Search campaign, one layer less visible.

Give it conversion data worth optimising toward. PMax is more dependent on clean conversion signal than any other campaign type, because it is making allocation decisions across seven surfaces on your behalf. Broken measurement plus PMax is the worst combination in the platform.

The honest answer is that it depends on catalogue size, account maturity and how much control you actually want to exercise.

Standard Shopping suits advertisers who want explicit control over which products get which budget, who have a structured campaign architecture by margin or category, and who are prepared to do the ongoing work that control implies.

Performance Max suits advertisers with clean conversion tracking, reasonable conversion volume, and either a large catalogue or a small team. It genuinely does find demand across surfaces that a Shopping-only setup will not reach.

Running both was the standard defensive recommendation for several years. Note that where they overlap on the same products, PMax has historically taken priority in the auction — so "both" in practice often meant "PMax, plus a Standard Shopping campaign that serves less than you think". With channel reporting and negative keywords now available in PMax, the diagnostic argument for the belt-and-braces approach is weaker.

Structuring by margin, not by category

The most useful Shopping structure most retailers never build is one organised by profitability rather than by product type.

Group products into high, medium and low margin. Give each its own campaign and its own target. A 4× ROAS target applied uniformly across a catalogue where margins range from 10% to 60% is a target that is simultaneously too aggressive for some products and far too soft for others.

This requires margin data in the feed, usually via a custom label. It is a genuine piece of work and it is the single highest-return structural change available to most ecommerce accounts.

Search captures demand that already exists. Someone typing "banquet hall garhwa" has a need and is looking for a supplier. YouTube and Demand Gen do something categorically different: they interrupt someone who was not looking for you.

This distinction sounds obvious and is routinely ignored in measurement. Video campaigns judged on last-click conversions will look terrible, because that is not how they work. The conversion, when it comes, usually arrives days later through a brand search — which the Search campaign then takes credit for.

Two conclusions follow.

Do not run video as a cheap way to inflate impressions. If it is not being measured properly, it is a donation.

Measure it as demand generation. The signals worth watching are brand search volume, direct traffic, view-through conversions and — where volume allows — a genuine incrementality test.

If you are going to run video

Creative decides most of the outcome, far more than targeting. The first five seconds carry it, because that is what a skippable ad gets. Lead with the problem, not the brand. Most viewers watch without sound, so the message has to survive muted.

Match format to job. Skippable in-stream for a considered message; bumpers for reinforcement; in-feed for people already browsing related content.

Audiences over demographics. Custom segments built from search behaviour — people who have recently searched terms relevant to your category — are usually the strongest available targeting, because they carry an intent signal into a non-intent surface.

Demand Gen replaced Discovery campaigns and now sits alongside PMax as a sibling channel. It is the right home for image and video creative aimed at generating interest, and it reports more transparently than the equivalent inventory inside PMax.

For most local service businesses

Be sceptical. A banquet hall, a clinic or a local contractor almost always has unexhausted search demand in its own catchment. Buying awareness through video before you have saturated the people already searching for you is an expensive way to be ambitious.

The order is: capture existing demand first, prove the unit economics, and only then spend money creating demand.

The uncomfortable question underneath everything in this guide is whether the conversions you are being credited with would have happened anyway.

For brand search, the answer is often partly yes. For remarketing, frequently yes — you are showing ads to people who already visited and were reasonably likely to return. For a Shopping campaign on a product nobody else stocks, largely no; those are genuinely incremental.

Attribution models do not answer this question. They distribute credit among the touchpoints that happened. They cannot tell you what would have happened in their absence, because they never observe that world.

How to actually find out

Geographic holdout. Turn the campaign off in one region and leave it running in a comparable one. Compare total business outcomes — not platform-reported conversions — over a period long enough to clear your sales cycle. This is crude, it requires enough volume in both regions to see through the noise, and it is the most honest tool available to most advertisers.

Google's built-in experiments. Campaign experiments and conversion lift studies do this more rigorously where you have the volume to qualify.

The pause test. For brand campaigns specifically: pause for two weeks and watch total conversions, including organic. If total conversions do not fall, the paid brand clicks were largely cannibalising free ones. If they do fall, someone else was catching those searchers.

Why this matters commercially

A reported 6× ROAS that is 60% incremental is a real 3.6×. That is still a good number. But the difference between the two decides whether scaling the budget makes sense, and scaling a channel on its reported figure is how accounts hit a wall that nobody can explain — spend doubles, reported ROAS holds, and total revenue barely moves.

You do not need to run incrementality tests continuously. You need to run one before any significant increase in budget, so that the decision to scale is made against a number that means something.