Key Takeaways

• Amazon’s ad engine now runs on intent-inference, not keyword matching, making hyper-segmented campaign structures actively harmful to performance.
• Any campaign generating fewer than 30 conversions in 30 days is in data starvation, meaning Amazon’s algorithm cannot optimise its bids.
• Consolidating campaigns by intent cluster, not match type, gives the algorithm the conversion volume it needs to exit permanent learning mode.
• Negative keywords replace match-type restriction as the primary control mechanism in a 2026-compliant account structure.
• Pausing campaigns and restarting them to fix performance typically makes the problem worse by triggering a learning reset and spiking CPCs.

Amazon’s advertising platform is no longer the keyword-matching system it was five years ago. The launch of COSMO, Amazon’s intent-inference engine, changed how ads are matched to queries. Products now surface based on inferred buyer intent, co-purchase patterns, and contextual signals, not keyword overlap. The campaign structures most accounts were built on in 2019 and 2020 assumed a different platform. That assumption is now costing sellers money. Accounts built around one keyword per ad group, complete match-type isolation, and exact match as the primary scaling mechanism are keeping campaigns in permanent learning mode. The algorithm cannot optimise without conversion data. When budgets are split across 50 fragmented campaigns, each one generating 3 to 5 conversions per month, none of them accumulates enough data to function. Performance plateaus. Spend becomes erratic. The fix is not tighter control. The fix is restructuring campaigns to feed the algorithm the conversion volume it actually needs.

Amazon’s intent-inference engine COSMO surfaces products based on buyer behaviour and contextual signals, not literal keyword matches. Any campaign generating fewer than 30 conversions in 30 days cannot activate rule-based bidding and stays in permanent learning mode. Hyper-segmented campaign structures spread budget thin across too many campaigns, starving each one of the data it needs to optimise. The 2026 account architecture separates campaigns by intent layer, not match type, with discovery, validation, and performance as the three core buckets. Negative keywords, not match-type restriction, form the control layer that keeps broader targeting profitable.

The shift from keyword-matching to intent-matching on Amazon mirrors a broader change in how machine-learning systems operate at scale. Algorithms optimise against patterns, and patterns require volume. When an account’s conversion data is fractured across dozens of isolated campaigns, no single campaign accumulates enough signal to learn from. The result is a system that looks controlled but functions poorly. Consolidation is not about losing precision. It is about applying precision at the level where it actually produces results: intent, not syntax. Accounts that understand this distinction stop fighting the algorithm and start feeding it. That is when spend becomes predictable and performance starts to scale.

Your Amazon Ad Structure Is Starving the Algorithm — Here’s the Fix

Why Did Amazon’s Advertising Platform Change So Drastically?

Amazon’s platform shifted from keyword-matching to intent-inference when it deployed COSMO, its AI-driven intent engine. COSMO does not match ads to queries based on words. It infers what a buyer actually wants from behavioural signals: search history, co-purchase patterns, review engagement, and contextual relationships between products and use cases.

A search for “shoes for a wedding” no longer retrieves products containing those exact words. It surfaces formal dress shoes because COSMO has inferred the buyer’s intent. The literal text of the query is secondary to what the buyer’s behaviour suggests they need.

Amazon’s ad engine is moving in the same direction. Broad match in 2026 surfaces ads for semantically related terms and intent-matched queries that go well beyond the literal keyword entered. The discovery function that used to live exclusively in auto campaigns now operates inside broad match manual campaigns too.

Manual control over match types matters less than it did. The fragmented structures built to maximise that control are now creating a different problem. The question to ask about any account is not “how tightly am I controlling match types?” It is “how much conversion data is each campaign actually accumulating?”

What Is Data Starvation and Why Does It Destroy Campaign Performance?

Data starvation is what happens when a campaign generates fewer than 30 conversions in 30 days, the threshold Amazon itself requires before activating rule-based ROAS bidding. Below that threshold, the algorithm stays in permanent learning mode.

In learning mode, a campaign cannot build statistical patterns. It bids inefficiently, sometimes overpaying for placements, sometimes losing competitive auctions unpredictably. Spend becomes erratic rather than optimised. The campaign looks active. It is not performing.

Hyper-segmented structures create this problem directly. Split a budget across 50 campaigns and each one receives a fraction of the available conversion data. A well-funded account running 50 campaigns might generate 3 to 5 conversions per campaign per month. None of them can learn. All of them stay inefficient.

The frustration compounds when sellers try to fix performance by pausing campaigns and restarting them. In account audits run by Clear Ads, CPCs spike 15 to 30% during active hours after a campaign restart. The restart triggers a new learning period. The campaign needed more data. It got less. The fix made the problem worse.

How Do You Diagnose Whether Your Account Has a Data Starvation Problem?

Pull your campaign report for the last 30 days and filter by conversions. Every campaign sitting below 30 conversions is in data starvation territory. That is the starting point, not bid adjustments or match type changes.

Count how many campaigns are below the threshold. In most accounts audited by Clear Ads, the majority of campaigns fall into this category. The problem is not a few underperforming campaigns. It is the structural architecture of the account.

An account with 200 keywords across 50 campaigns looks disciplined. To Amazon’s algorithm, it looks like 50 campaigns that cannot learn. The structure that felt like control is the source of the inefficiency.

What Does a 2026-Compliant Amazon Campaign Structure Look Like?

The 2026 account architecture organises campaigns by intent layer, not match type. Three buckets form the structure: discovery, validation, and performance.

The discovery bucket runs auto and broad match campaigns with conservative bids and down-only bid adjustments. This bucket is the data engine. Its job is to surface new intent signals, accumulate conversion data across a wide range of queries, and feed the algorithm the volume it needs to learn.

The validation layer uses phrase match campaigns at moderate bids aligned to target ACoS. This layer tests whether the intent signals from discovery hold up under tighter targeting. Keywords that convert consistently here earn promotion to the performance core.

The performance core uses exact match for two purposes only: brand defence keywords that need to be owned, and terms with documented conversion history from the discovery and validation layers above. Exact match is not the starting point for every keyword. It is a lock-in mechanism for terms that have already earned it.

Budget allocation shifts accordingly. In the old model, exact match received the largest share because it felt safest. In the 2026 model, discovery and validation receive the largest share because that is where the conversion data comes from that makes the performance core work. Exact match without conversion history behind it is not precision. It is expensive guessing.

Should You Negate Broad Match Winners When You Promote Them to Exact?

No, not automatically. A keyword performing well in broad match is often winning different placements than the same keyword in exact match. Both can be profitable simultaneously.

Test both before negating. Removing a broad match winner that is earning profitable placements does not give you control. It removes revenue. Only negate when data confirms the broad match version is cannibalising the exact match version rather than complementing it.

How Do You Control Spend With Broader Match Types?

Negative keywords replace match-type restriction as the control layer. In a 2026 account structure, negatives are not cleanup. They are architecture.

The practical system runs on weekly Search Term Report audits. Flag any term that has hit the break-even spend threshold with zero conversions. For a product at average price, that threshold is around $35, though it scales with AOV and target ACoS. A $10 product and a $150 product have very different break-even points.

Use negative phrase match at campaign level for category-level irrelevance: wrong product type, wrong demographic. Use negative exact at ad group level for surgical blocking of specific high-volume non-converters.

Accounts running this discipline consistently show 15 to 25% reductions in wasted spend. Wider targeting with disciplined negative keywords outperforms exact-match-only structures in most accounts at most budget levels. The negative keyword list is what makes wider targeting safe.

What Should You Do With Your Amazon Ad Account Right Now?

Pull the campaign report. Filter by conversions over the last 30 days. Count how many campaigns sit below 30. Each one is a campaign Amazon cannot optimise. The budget going into it is working against you.

The exact-match-only structure made sense on a keyword-matching platform. Amazon is now an intent-matching platform. The structure built for control in 2022 is creating data starvation in 2026. More campaigns is not more control. It is a slower algorithm and erratic spend.

Structure for data first. Then structure for precision. The algorithm will do its job once you give it what it needs to learn.