Digital Marketing Exam 2

[Wk 8] Three versions of the same AI request

Bare (generic), with context (who you are, who buys, how you talk), grounded (adds real material to read).

[Wk 8] Grounded prompt

A prompt that includes real material for the model to read (menu, reviews, competitor page). Reduces hallucination.

[Wk 8] Three levers to improve AI output

Context, retrieval, agency.

[Wk 8] Context

What the model knows about you. It only knows what you typed.

[Wk 8] Retrieval

What the model can read before answering (documents, search results). Also called RAG.

[Wk 8] Agency

What the model can do beyond writing, using tools like search, browsing, and running code.

[Wk 8] Prompt like a brief

Write the prompt like a creative brief: clear task, audience, and context.

[Wk 8] Give models examples

Show examples of what you want instead of just describing it. Also called few-shot prompting.

[Wk 8] Adversarial reasoning

Never accept the first draft. Make the model argue against you, or have another model find flaws.

[Wk 8] AI agent

A language model that can use tools and act on its own.

[Wk 9] Two kinds of data

Observational (collected by watching) and experimental/A/B (result of an intervention).

[Wk 9] Big data

Very large, constantly generated, often unstructured (over a million rows, images, video).

[Wk 9] Correlation vs. causation

Correlation does not imply causation, but causation always implies correlation.

[Wk 9] Three conditions for causation

1) Time order (X before Y). 2) Correlation. 3) No other cause explains the change.

[Wk 10] Treatment vs. control

Treatment gets the change (music room). Control doesn't (quiet room).

[Wk 10] Selection bias

When a person or rule picks who gets the treatment.

[Wk 10] Random assignment

Chance, not a person, decides who gets which version. This removes selection bias.

[Wk 10] Why no individual treatment effect?

Each unit is in only one world, so we never see both outcomes.

[Wk 10] Average Treatment Effect (ATE)

The average of the individual effects. Estimated as treatment average minus control average.

[Wk 10] Weakened control

A control that differs in more than the treatment (e.g., music AND light), so you can't tell the cause.

[Wk 10] Spillover (treatment leakage)

Treatment reaches the control group and shrinks the gap. Fix: cluster randomization.

[Wk 10] Cluster randomization

Randomize groups instead of individuals. One coin flip = one cluster.

[Wk 10] Assigned vs. received treatment

Being in the treatment group doesn't mean you got it (e.g., you never saw the ad).

[Wk 10] Difference and lift (5% vs. 5.5%)

Difference = 0.50 points. Lift = 0.5 / 5 = 10%.

[Wk 10] Standard error (SE)

How random your estimate is. Smaller SE means more reliable. It shrinks as sample size grows.

[Wk 10] Null vs. alternative hypothesis

H0: no effect (Pb = Pa). H1: there is an effect (Pb does not equal Pa).

[Wk 10] Z statistic

The gap measured in standard errors. Below 1.96 = not significant.

[Wk 10] p-value

If there were no effect, how often would we see a gap this big? Smaller p = stronger evidence.

[Wk 10] Confidence interval

If it includes 0, the result is not significant (effect unknown).

[Wk 10] Power

The chance the test detects a real effect. 90% power = 90% chance.

[Wk 10] Two types of errors

Type I: false positive (see an effect that isn't there). Type II: false negative (miss a real effect).

[Wk 11] One unit in an AI experiment

One fresh chat.

[Wk 11] Running Version A and B

In separate chats. Only the treatment differs.

[Wk 11] Final project treatment

Exactly one change (e.g., a picture, color, or heading).

[Wk 11] AI disclosure line and records

Name the exact model, not just the product. Save every answer in Sheets or Excel.