DMI Expert cheat sheet

Porter's Five Forces: rivalry, supplier power, buyer power, threat of entry, threat of substitutes; a force is 'strong' when it depresses industry margins below the cost of capital.

SWOT is an input, not a strategy: pair internal S/W with external O/T to generate SO (attack), WO (build), ST (defend), WT (avoid) initiatives.

SMART goals require a metric, a baseline, a target, and a deadline; 'increase brand awareness' fails until a measurable indicator (e.g., aided recall %) is attached.

OKR rule: Objectives are qualitative and time-boxed; Key Results are 3-5 quantitative outcomes, not task lists; a KR score of 0.6-0.7 is commonly treated as healthy stretch.

Ansoff matrix: market penetration (existing product/market), product development, market development, diversification - diversification carries the highest risk.

Digital maturity models typically stage: ad hoc → emerging → structured → integrated → optimized; investment in governance and data precedes scale-up.

Pareto / 80-20 prioritization: rank initiatives by expected impact ÷ effort; the top ~20% of levers usually drive ~80% of incremental value in a portfolio review.

RACI assigns one Accountable per decision; multiple Accountable owners is a governance defect, not a collaboration win.

Unit economics gate: LTV:CAC ≥ 3:1 and CAC payback ≤ 12 months are common thresholds before scaling paid acquisition spend.

Quality Score (0-10) is built from expected CTR, ad relevance, and landing page experience; higher QS lowers CPC because Ad Rank = bid × quality.

Ad Rank = bid × Quality Score (+ expected impact of extensions and ad formats); you can outrank a higher bidder with better quality.

Match types: exact [keyword], phrase "keyword", broad keyword; broad match modifier was retired, so use negative keywords and audience signals to control broad reach.

Negative keywords are essential on broad and phrase; add search terms with high spend and zero conversions to a negative list at least weekly.

ROAS = revenue ÷ ad spend; CPA = cost ÷ conversions; breakeven ROAS = 1 ÷ gross margin (e.g., 25% margin → 4.0 breakeven ROAS).

Impression share = impressions ÷ eligible impressions; lost IS (budget) vs lost IS (rank) diagnose whether to raise budget or improve quality/bid.

SEO crawl budget: robots.txt controls crawling, meta robots/noindex controls indexing, canonical tags consolidate duplicate URLs; sitemaps aid discovery, not ranking.

Core Web Vitals thresholds: LCP ≤ 2.5s, INP ≤ 200ms, CLS ≤ 0.1 (75th percentile) - the page-experience signal set used in ranking and ad landing page quality.

Search intent tiers: informational, navigational, commercial investigation, transactional; map each to page type (guide, brand page, comparison, product/checkout).

Engagement rate = (likes + comments + shares + saves) ÷ reach (or ÷ impressions, or ÷ followers); state the denominator - the number is not comparable across definitions.

Reach vs impressions: reach counts unique accounts; frequency = impressions ÷ reach; high frequency with flat reach signals creative fatigue.

Funnel mapping: awareness → reach/video views; consideration → saves, shares, profile visits, link clicks; conversion → leads, purchases, ROAS.

Platform-native formats: short vertical video for discovery feeds, carousels for education, Stories for time-limited offers, live for community and launches.

Community management SLAs: first response within 1-4 business hours for public complaints; escalate to legal/PR for crisis keywords.

UGC and creator briefs must specify deliverables, usage rights, disclosure (#ad / paid partnership label), and exclusivity windows.

A/B test one variable at a time (hook, thumbnail, caption, CTA); require minimum sample before declaring a winner - early spikes are noise.

Social listening metrics: share of voice = brand mentions ÷ total category mentions; sentiment = positive ÷ (positive + negative) mentions.

Paid social bidding: CPM for reach, CPC for traffic, CPA/ROAS for conversion; optimize to the event closest to revenue that has enough volume.

Supervised vs unsupervised vs reinforcement learning: labels for prediction, clustering for segmentation, reward signals for sequential bidding/optimization.

Training/validation/test split prevents overfitting; a model that scores well on training data but poorly on held-out data has memorized, not learned.

Precision = TP ÷ (TP + FP); recall = TP ÷ (TP + FN); for churn or fraud, choose the threshold based on the cost of false positives vs false negatives.

Generative AI output requires human review for factual claims, brand voice, and legal/compliance; never publish unverified statistics or claims.

Prompt structure: role + task + context + constraints + output format + examples (few-shot); specify format (JSON, table, word count) to reduce rework.

Retrieval-augmented generation (RAG) grounds answers in your own documents; it reduces hallucination but does not eliminate it - cite sources.

Personalization needs a consent basis and data minimization; segment-level personalization is lower risk than 1:1 profiling under privacy rules.

Model evaluation: hold out a test set, track precision/recall or MAE, and monitor drift after deployment; retrain when performance degrades beyond a set threshold.

Attribution and incrementality: AI bid strategies optimize to modeled conversions; validate with geo holdouts or lift tests before reallocating budget.