Structuring a B2B account for offline conversion imports

Turning on OCT in a poorly structured account is like giving better fuel to a car with a broken engine. The data quality improves. The underlying structural problems don't go away.

I've seen this happen. Agency turns on offline conversion tracking. Suddenly Smart Bidding has real revenue data. The algorithm immediately shifts budget toward the campaigns with the most attributed revenue — which happen to be the campaigns with the highest-volume keywords, not necessarily the most efficient ones.

Result: bidding improves at the campaign level while the account structure actively undermines the optimization at the keyword level.

To make offline conversions work, your account structure has to be designed around them.

The core tension

Smart Bidding needs data density to function. It generally needs 30-50 conversions per campaign per month to operate reliably.

In niche B2B with high-ACV products, you might close 3-5 deals per month per campaign. That's not enough for the algorithm to learn on.

But building fewer, larger campaigns to hit the data threshold means combining different products, intents, and match types. Which means your ad copy becomes generic. Which hurts conversion rate. Which reduces the data you're trying to build.

The SKAG approach (one keyword per ad group, maximum specificity) is ideal for ad relevance. It's terrible for Smart Bidding because you're fragmenting conversion data across hundreds of tiny ad groups.

The structural solution

The answer is to separate your ad group architecture from your campaign architecture.

Ad groups stay granular. One primary keyword or small, tightly related keyword cluster per ad group. Specific ad copy. SKAG-adjacent structure. This preserves the ad relevance benefits.

Campaigns consolidate to hit data thresholds. Instead of one campaign per product, you group related product families into single campaigns. Not because the products are interchangeable, but because the consolidated campaign can now accumulate enough conversion data for the algorithm to work with.

For example: instead of separate campaigns for "dot-peen markers," "laser markers," and "scribing systems," one campaign covers all three product types. Within that campaign, each product type has its own ad groups with specific copy.

The algorithm optimizes at the campaign level with real data. The ad copy performs at the ad group level with real specificity. Both priorities are served.

Portfolio bid strategies

The third layer: portfolio bid strategies. You can pool conversion data across multiple campaigns without fully combining them.

A portfolio Target CPA or Target ROAS strategy can learn from conversions across all the campaigns assigned to it. This lets you keep campaigns separated by brand, geography, or product while giving the algorithm a pooled dataset to optimize against.

At $650K/month, I run three portfolio bid strategies:

1. High-margin product campaigns (pooled, aggressive)
2. Mid-margin campaigns (pooled, moderate)
3. Brand search and defensive campaigns (separate, manual CPC)

The max CPC caps inside each portfolio strategy are critical. Without caps, the algorithm will occasionally spike bids dramatically on what it thinks is a high-value opportunity. At $21K/day, an uncapped spike is a real problem.

The setup sequence

1. Build account structure first. Ad groups at keyword level, campaigns at product-family level. Before you touch conversion settings.

2. Implement offline conversion tracking. Get the pipeline working correctly. Validate that actual closed deals are getting attributed.

3. Let the system collect 60-90 days of data. Don't run Smart Bidding during this period. Use manual CPC with enhanced CPC at most.

4. Once you have enough data to see conversion patterns, create portfolio bid strategies and assign campaigns.

5. Set max CPC caps immediately. Monitor daily for the first four weeks.

This sequence takes patience. Most teams want to turn on Smart Bidding and OCT simultaneously and expect immediate results.

The algorithm learns on historical data. If you haven't given it the right structure and enough time to collect that data, it will learn the wrong things.

Do it in the right order. The results compound over 6-12 months.