DHM framework: delight, hard to copy, margin-enhancing

DHM is not a checklist. It is a single compound question: does this initiative delight customers in a way that is structurally hard to replicate and that improves our economics over time? If you cannot answer yes to all three simultaneously, you have a feature, not a strategy. Gibson Biddle coined it while VP of Product at Netflix and later applied it as CPO at Chegg. It is his personal operating model, not a generic consulting artifact.

The three pillars

D: delight means the experience is genuinely better in a way users feel, not just measure. Not “users are satisfied” but “users would notice and miss it if it were gone.” The test is behavioral: did usage change? Netflix’s personalization algorithm grew from 2% to 80% of what members watched. That is delight you can measure.

H: hard to copy is the strategy pillar. Something is hard to copy when a competitor cannot replicate it in 6 to 12 months by throwing resources at it. The structural types map directly to Hamilton Helmer’s 7 Powers: scale economies, network effects, counter-positioning, switching costs, branding, cornered resource, process power. Biddle adds captured resources (key patent positions, proprietary data, key hires) as an eighth. Naming the moat type is required. “Our algorithm is hard to copy” is not an answer. “Our algorithm is hard to copy because it has been trained on 10 years of behavioral data that competitors cannot access without first acquiring our user base” is an answer.

M: margin-enhancing means the initiative improves the unit economics of the business, not just revenue. It might reduce CAC by improving retention and referral, reduce content or inventory cost by predicting demand better, or increase pricing power through brand or switching costs. The margin logic should be traceable: what specifically improves, and why does this initiative cause that?

The dependency most candidates miss

A strong D does not automatically give you H. Most features delight temporarily and can be copied in under a year. That is the critical asymmetry: the delight column is where PMs spend most of their thinking, and the hard-to-copy column is where strategy actually lives. The Netflix personalization flywheel worked because the more users watched, the better the recommendations, the more they watched. A competitor starting from zero could not replicate that without years of accumulated behavior data. It passed D, H, and M simultaneously, which is what makes it strategy rather than a product bet.

The canonical failure is Netflix’s “Friends” social feature. It was designed as a network effect play (H) and an engagement play (D), but only 6% of members used it. It failed the D test in practice despite looking plausible in theory. Biddle uses this example to make a specific point: a strategy is a hypothesis, and hypotheses can be wrong. Netflix tested 4-6 product strategies per year, each run by a dedicated cross-functional pod. The discipline is to formulate the hypothesis in DHM terms, run the test, and let the result update your model.

The strategy statement format

Biddle’s preferred format: “We believe that [initiative] will delight customers by [mechanism], is hard to copy because [moat type], and will enhance margin by [economic logic].”

Written out for Netflix personalization: “We believe that personalized content recommendations will delight members by reducing the time they spend searching and improving match between their tastes and available content, is hard to copy because it requires a behavioral data flywheel that compounds with every viewing interaction and cannot be bootstrapped by a competitor without years of member-scale usage, and will enhance margin by reducing content licensing cost through demand prediction and improving retention which lowers the CAC-to-LTV ratio.”

That is what the format is for: forcing you to be specific about the mechanism of each pillar so you can test whether it actually holds.

The 2026 reframe: H is doing more work than ever

When feasibility was a real constraint (pre-AI-era tooling), a technically difficult feature could claim H on complexity alone. A competitor simply could not build it fast enough. That moat no longer exists. Any team can ship a working version of a complex feature in weeks using current AI development tooling. Feasibility is no longer the binding constraint, which means the H pillar cannot rely on technical difficulty.

The real moats in 2026 are behavioral data flywheels (the more users interact, the better the model, the more users stay, and a new entrant without data cannot close that gap), proprietary workflow integration (switching cost moat: your product is embedded in how teams operate, and replacing it means retraining, remigration, and re-approval), and trust and brand in high-stakes domains (users will not risk a cheaper alternative when the cost of failure is high). A strong DHM answer for an AI product should name which of these the H creates and trace why it is structural rather than temporary.

The M pillar is also being rewritten. AI inference costs are high but falling. Margin strategy for AI products is now about whether your usage model improves unit economics as scale grows, or whether costs scale linearly with every query. A product where user behavior generates training signal that reduces inference cost over time has a margin flywheel. One where every query costs the same regardless of accumulated usage does not.

GEM: the prioritization layer above DHM

DHM generates strategy hypotheses. GEM (Growth, Engagement, Monetization) is the layer Biddle uses to force-rank them against company stage. A pre-product-market-fit company should weight growth hypotheses most heavily. A product with strong retention but weak monetization should weight M hypotheses. A product with high acquisition but poor retention should weight E hypotheses. DHM tells you whether something qualifies as strategy. GEM tells you which strategies to run first.

Applying DHM outside Netflix

DHM is easier to apply in high-engagement consumer subscription products with clear behavioral signals. It requires deliberate adaptation elsewhere.

For B2B SaaS (Stripe, Databricks): the D metric is time-to-value and workflow fit, not consumer-style engagement. The H is typically switching cost (deep API integration, compliance workflows, institutional knowledge baked into configuration) or data network effects (your product aggregates signals across many enterprise customers that no single customer can see). The M logic often runs through expansion revenue within accounts rather than volume growth.

For platforms and marketplaces: D applies to both sides. The H is almost always network density. The M follows from take rate and cost efficiency as density grows. DHM for a marketplace requires writing the hypothesis twice: once for supply, once for demand, and checking whether both sides of the flywheel are covered.

Chegg application (Biddle’s second major case): personalized tutoring recommendations (D: students find the right resource without wading through irrelevant material), proprietary student behavior dataset built over years of usage that no new entrant can replicate at launch (H: data scale moat), reduced CAC from direct student relationships rather than paid acquisition (M).

Use it, do not recite it

---

DHM connects naturally to jobs-to-be-done for grounding the D pillar in specific customer outcomes, and to north star metric for operationalizing the GEM layer. For why the H pillar carries more weight when feasibility is cheap, see feasibility is free.