Definitions
This glossary is for readers who know a little statistics and marketing but have not worked with advertising-attribution systems. Definitions describe how terms are used in this project.
Essential reading path
Start with these ideas:
- Touchpoint
- Customer path
- Conversion
- Attribution
- Multi-Touch Attribution system
- Causal effect and incrementality
- Ground truth
Advertising and journey terms
Attribution
Attribution is the rule or calculation used to divide credit for a business result among earlier marketing activities.
Example: A student sees a shoe advertisement on Monday, clicks a search advertisement on Friday, and buys on Saturday. Attribution decides how much credit Monday and Friday receive.
Conversion
A conversion is a business action the advertiser cares about, such as a purchase, subscription, or completed registration. In this project, conversion usually means a purchase.
Customer path
A customer path is the ordered sequence of touchpoints observed before a conversion or before the measurement window ends.
Example: Display advertisement → brand-search advertisement → product advertisement → purchase.
Funnel
A funnel is a simplified picture of how a large audience becomes a smaller group of customers:
awareness → interest → consideration → purchaseThe shape resembles a funnel because many people may see an advertisement, fewer click, and fewer still purchase.
Touchpoint
A touchpoint is one recorded opportunity for marketing to influence a shopper. In this project, a touchpoint identifies the advertising product, advertisement or inventory type, placement, and creative type.
Example: SPONSORED_PRODUCTS:PRODUCT_AD:TOP_OF_SEARCH:UNSPECIFIED represents a Sponsored Products advertisement shown at the top of search results.
The finalized key does not say whether the shopper viewed or clicked the advertisement. That is a known schema limitation.
Amazon Demand-Side Platform funnel
The Amazon Demand-Side Platform is a system advertisers use to buy display, video, audio, and other advertising inventory programmatically. A demand-side platform chooses among available advertising opportunities according to budgets and targeting rules.
Its funnel describes movement from broad exposure toward purchase:
display or video impression
→ click or product-detail-page visit
→ add to cart
→ purchaseIt is useful for studying upper-funnel awareness and later shopping behavior. Aggregated funnel counts do not reconstruct a specific person's complete journey.
Sponsored Products, Sponsored Brands, and Sponsored Display
- Sponsored Products promote individual products, often in search results.
- Sponsored Brands promote a brand, collection, or branded creative.
- Sponsored Display reaches audiences with display-style advertisements, including retargeting.
They serve different parts of the funnel, so comparing only final clicks can undervalue earlier brand or display activity.
Attribution methods
Multi-Touch Attribution system
A Multi-Touch Attribution system divides conversion credit among multiple advertising touchpoints rather than giving everything to one touchpoint.
It normally has three jobs:
- reconstruct or summarize customer paths;
- estimate how much each touchpoint contributed;
- aggregate credit into campaign, channel, or budget reports.
Example: For a $100 purchase after display → brand search → product search, a system might assign $20, $30, and $50 of attributed revenue. Those numbers describe an allocation rule or estimate; they are not automatically causal truth.
First-touch attribution
First-touch attribution gives all credit to the earliest observed touchpoint. It is easy to explain and can highlight activities that introduce customers to a brand, but it ignores later influence.
Last-touch attribution
Last-touch attribution gives all credit to the final observed touchpoint. It is simple and often favors advertisements close to purchase.
Linear attribution
Linear attribution divides credit equally among all touchpoints. Three touchpoints would each receive one third. Equality is transparent but may not reflect actual influence.
Position-based attribution
Position-based attribution gives chosen shares to the first and last touchpoints and divides the remainder among the middle touches.
Example: A common rule assigns 40 percent to the first, 40 percent to the last, and 20 percent across the middle.
Time-decay attribution
Time-decay attribution gives more credit to touchpoints closer in time to conversion. It requires real timestamps or time intervals. Path order alone is insufficient.
Markov chain and removal effect
A Markov chain models movement between states using transition probabilities.
Example: Historical paths may show that people move from display advertising to product advertising 30 percent of the time. The model combines such transitions to estimate the chance of eventually converting.
A removal effect asks: “How much does the estimated conversion probability fall if this touchpoint is removed from the transition network?” A larger fall leads to more attribution credit.
Markov removal is data-driven, but it still learns association from observed paths. It does not prove causality.
Shapley value and Shapley marginal contribution
A Shapley value comes from cooperative game theory. Imagine touchpoints as players on a team and conversion lift as the team's prize. The Shapley method considers every order in which players could join and averages how much each player adds.
A marginal contribution is the change produced when one touchpoint is added to a particular group of existing touchpoints.
Example: Suppose display alone adds 2 sales, search alone adds 5, and together they add 10. The extra value of adding display after search is 5, not 2, because the pair has a useful interaction. Shapley values average contributions like this across possible joining orders.
Shapley allocation handles interactions more symmetrically than a fixed rule, but its result depends on how the value of every touchpoint group is estimated.
Evidence and causal reasoning
Causal effect and incrementality
A causal effect is the difference caused by an action, compared with what would have happened without that action.
Incrementality is the additional outcome caused by advertising.
Example: If 120 people purchase after an advertisement but an equivalent no-ad group would have produced 100 purchases, the estimated incremental effect is 20 purchases—not 120.
Observational data
Observational data records what happened without randomly deciding who receives advertising. It can reveal patterns but may confuse customer intent with advertising influence.
Example: People already planning a purchase may search for the product and then convert. Search and conversion are strongly associated, but the search advertisement may not have caused all those purchases.
Randomized controlled trial and holdout group
A randomized controlled trial randomly assigns eligible people or regions to different conditions.
- The treatment group can receive the advertising.
- The holdout group is prevented from receiving it.
Random assignment makes the groups comparable on average, allowing differences in outcomes to estimate incrementality.
Ground truth
Ground truth is a reference answer used to evaluate a model. In real advertising, perfect causal ground truth is rare.
In this project, simulation_ground_truth.csv records the simulator's known response mechanism. It is valid for testing synthetic experiments only. It is not Amazon data and must not become a model input.
Inference
An inference is a conclusion supported by evidence but not directly observed or explicitly stated. A good inference identifies its assumptions.
Recommendation
A recommendation is a proposed choice or next step. It remains open to review and is not an existing repository fact.
Data and model terms
Amazon Marketing Cloud
Amazon Marketing Cloud is a privacy-focused analytics environment where eligible advertisers run queries over pseudonymized advertising and shopping events. Exported results are aggregated and must satisfy privacy rules.
Amazon Marketing Stream
Amazon Marketing Stream delivers advertising performance and management messages, often at hourly granularity. It supports monitoring and data-warehouse ingestion but does not independently reveal complete user-level cross-channel paths.
Baseline model
A baseline model is a simple method used as a comparison point. A complex model should demonstrate a clear benefit over a baseline before its extra complexity is accepted.
Calibration
Calibration asks whether predicted probabilities match observed frequencies.
Example: Among journeys assigned a 20 percent conversion probability, approximately 20 percent should convert.
Causal calibration can also mean adjusting observational estimates to better match experimental lift.
Carryover or adstock
Carryover means advertising can influence outcomes after the day it runs. Marketing Mix Modeling often represents this delayed memory with an adstock transformation.
Click-through rate
Click-through rate is clicks divided by impressions. If 1,000 impressions produce 20 clicks, the click-through rate is 2 percent.
Conversion rate
Conversion rate is conversions divided by a relevant opportunity count, such as clicks or users. The denominator must always be stated.
Binomial distribution
A binomial distribution describes how many successes occur in a fixed number of independent yes-or-no trials when every trial has the same success probability.
Example: If 1,000 impressions each have a 2 percent click probability, a binomial draw produces a realistic click count near 20 instead of forcing the count to equal exactly 20.
Data leakage
Data leakage occurs when model training accidentally receives information that would not be available when making a real prediction.
Using simulation_ground_truth.csv as a training feature would be direct leakage because it contains the answer the model is supposed to estimate.
Economic willingness multiplier
An economic willingness multiplier is this project's experimental adjustment from economy-wide household consumption growth to simulated purchase odds.
Example: 1.02 means conversion odds are multiplied by 1.02 before advertising effects. It does not mean conversion probability rises by two percentage points.
This value is a modeling proxy, not a measured willingness score for Amazon shoppers.
Gini coefficient
The Gini coefficient summarizes how unevenly income is distributed. Zero represents complete equality; one represents maximum inequality in the mathematical limit.
It describes a population distribution. It does not reveal any one shopper's income or prove that a premium product will sell more.
Hidden variable
A hidden variable affects simulated behavior but is not written into the model-input tables.
Example: A simulated economic mood can change conversion probability while remaining unavailable to the attribution model. Hidden variables help test whether a model incorrectly credits advertising for something it could not observe.
Internet reach rate
Internet reach rate is the fraction of a country's population reported as using the internet. It is an upper-level accessibility measure, not the fraction using Amazon or seeing a specific campaign.
Logistic response
A logistic response converts any real-valued score into a probability between zero and one. Logistic regression often models conversion log odds as a sum of feature effects.
Log odds
For probability (p), odds are (p/(1-p)), and log odds are the natural logarithm of those odds. Adding a positive log-odds effect increases conversion probability, although the probability increase depends on the starting level.
Multiplicative noise
Multiplicative noise changes a quantity by a random ratio rather than a random addition.
Example: Multiplying traffic by 1.08 raises it by 8 percent, while multiplying by 0.92 lowers it by 8 percent. Using positive multipliers prevents traffic or revenue from becoming negative.
Marketing Mix Modeling
Marketing Mix Modeling uses aggregated time-series data to estimate how marketing channels, prices, seasonality, and other factors relate to sales. It is useful when user-level paths are unavailable, but it answers a different question from path-level attribution.
Normalized touchpoint key
A normalized touchpoint key is a consistent four-part label that lets the path table match the performance table:
adProduct : adType-or-inventoryType : placement : creativeTypeMissing structural segments become UNSPECIFIED inside this key only.
Response curve and saturation
A response curve connects advertising input, such as spend, to an expected outcome.
Saturation means each additional dollar may produce less benefit after the most valuable opportunities have already been reached.
Seasonality
Seasonality is a repeating time pattern, such as higher weekend shopping or holiday demand. A model that ignores seasonality may incorrectly credit advertising for changes caused by the calendar.
Random seed
A random seed is the starting value for a pseudorandom number generator. The same seed, algorithm, and configuration reproduce the same sequence of simulated draws.
Changing the seed creates a different sample from the same assumptions; it does not change those assumptions.
Standard deviation
Standard deviation describes how widely values vary around their average. A larger simulation standard deviation produces less stable daily results.
Structural null value
A structural null value means a field does not apply, rather than being accidentally missing.
Example: placement is structurally null for an Amazon Demand-Side Platform touchpoint in the finalized schema. It remains null in the raw dimension column but becomes UNSPECIFIED inside the normalized key.
Target audience density and sparsity
Target audience density is the assumed fraction of reachable people who match campaign conditions such as age, interests, or income. Sparsity describes the opposite idea: a sparse target audience has low density.
Example: Density 0.02 means the simulation treats 2 out of 100 online people as eligible prospects. Without platform audience evidence, this remains a low-confidence assumption.
Traffic multiplier
A traffic multiplier scales campaign volume relative to a reference.
Example: If the reference produces 100 visits, multiplier 1.0 also starts from 100, while multiplier 0.25 starts from 25 before other adjustments.
It is a simulation scale, not proof of actual country traffic.
Weighted log loss
Log loss evaluates predicted probabilities by penalizing confident wrong answers heavily. A weighted version lets one aggregated row represent many users instead of expanding it into individual records.