Associated keywords: the missed opportunity in search advertising | R bloggers

Short practical advice about search advertising:

1. Look beyond brand and category keywords
   The article shows that “related keywords” – terms like “license”, “battery”, “plates” – are searched 50 to 80 days before purchase, have a high purchase probability in the category and cost significantly less than brand or category terms.
2. Just a few days of search history significantly improves targeting
   Models that use five days of search history achieve an 8 to 15 percentage point improvement in predicting conversions compared to targeting a single query. The efficiency decreases rapidly after 7 days.
3. Ad effectiveness varies by journey stage and keyword type
   Branded keywords work best at purchase (1.5x lift), but associated keywords drive meaningful lift earlier in the journey, when the gap between category and brand engagement is widest.
4. Focuses on the ‘commitment gap’
   The biggest chance is that people want to buy into the category, but are not yet committed to a specific brand. This gap is widest at the beginning of the journey, which is exactly where advertisers are currently spending the least.

Long version

A recent paper by Rothschild, Needell, Veverka, and Yom-Tov (2025) maps conversion paths using a massive corpus of Bing searches. What I found interesting is how clearly it shows that advertisers are focusing their spend on people who are already likely to buy, while ignoring earlier stages of the journey where the marginal return on advertising could be much higher.

The authors introduce “associated keywords”: terms related to the conversion path, but not to the actual products, brands, or product categories. These terms represent a largely untapped opportunity to target people who are eager to buy in this category, but before they have committed to a specific brand.

Key findings

The article analyzes three product categories: vehicles (cars and trucks), laptops and phones. This is what they found:

The authors also find that:

• Most conversion paths are short and sharp, but some users use multiple paths (search early, pause, then buy)
• Advertisers heavily focus their ads late in the journey, when users are already likely to convert
• Associated keywords appear often and earlier in the conversion process
• Five days of search history improves prediction accuracy by 8-15 percentage points

Associated keywords explained

The article defines associated keywords as terms that: (1) are related to the conversion journey, (2) have a high probability of purchase in the product category, (3) have a low probability of purchase for a specific product, (4) are cheaper than category/brand terms, and (5) occur early in the journey.

Examples from the paper:

The keywords “license” and “battery” are searched 2-3 months before purchase. They indicate a high intent to purchase in the category, but don’t commit to a specific brand – and they cost a fraction of category terms like ‘truck’.

R simulation

Let me simulate some of these insights to visualize the opportunity.

library(ggplot2)
library(dplyr)
library(tidyr)

set.seed(42)

# --- Keyword types and their properties ---
# Based on Table 2 from the paper

keywords <- data.frame(
  keyword = c("Ford", "Honda", "Toyota", "truck", "car", "license", "plates", "dmv", "battery"),
  type = c("Brand", "Brand", "Brand", "Category", "Category", "Associated", "Associated", "Associated", "Associated"),
  ad_cost = c(0.60, 0.70, 0.65, 1.97, 2.20, 0.62, 0.35, 0.40, 1.06),
  category_prob = c(0.76, 0.80, 0.78, 0.50, 0.55, 0.69, 0.84, 0.75, 0.71),
  brand_prob = c(0.58, 0.55, 0.50, 0.18, 0.10, 0.26, 0.60, 0.25, 0.19),
  days_before_purchase = c(0, 0, 0, 0, 0, -80, 0, -60, -85),
  stringsAsFactors = FALSE
)

# Calculate the "commitment gap"
keywords$gap <- keywords$category_prob - keywords$brand_prob

head(keywords)

  keyword      type ad_cost category_prob brand_prob days_before_purchase    gap
    Ford     Brand    0.60         0.76       0.58                   0  0.18
   Honda     Brand    0.70         0.80       0.55                   0  0.25
  Toyota     Brand    0.65         0.78       0.50                   0  0.28
    truck Category    1.97         0.50       0.18                   0  0.32
      car Category    2.20         0.55       0.10                   0  0.45
  license Associated    0.62         0.69       0.26                  -80  0.43

Plot: The associated keyword opportunity

ggplot(keywords, aes(x = days_before_purchase, y = gap,
                     size = ad_cost, color = type, label = keyword)) +
  geom_point(alpha = 0.7) +
  geom_text(hjust = -0.15, vjust = 0.5, size = 3) +
  geom_hline(yintercept = 0, linetype = "dashed", color = "gray60") +
  scale_x_continuous("Expected days before purchase") +
  scale_y_continuous("Category - Brand probability gap") +
  scale_size_continuous("Ad cost ($)", range = c(2, 8)) +
  labs(title = "The Associated Keyword Opportunity",
       subtitle = "High gap = likely to buy in category, but not committed to brand",
       caption = "Keywords with high gap and early timing represent untapped opportunities") +
  theme_minimal(base_size = 12) +
  xlim(-90, 10)

The chart shows what the article describes: associated keywords like “license,” “battery,” and “dmv” appear early in the journey (negative days), have a large gap between category and brand likelihood, and cost significantly less than category terms.

Model accuracy with search history

The paper shows that just five days of search history dramatically improves prediction accuracy. Let me simulate this:

# Diminishing returns model based on paper's findings
days_history <- c(0, 1, 2, 3, 5, 7, 14, 30, 60, 90)
base_accuracy <- 0.65
max_improvement <- 0.15  # 8-15 pp improvement from paper

accuracy_fn <- function(days) {
  base_accuracy + max_improvement * (1 - exp(-0.3 * days))
}

model_accuracy <- data.frame(
  days_history = days_history,
  accuracy = sapply(days_history, accuracy_fn)
)

ggplot(model_accuracy, aes(x = days_history, y = accuracy)) +
  geom_line(color = "#1f77b4", linewidth = 1.2) +
  geom_point(color = "#1f77b4", size = 3) +
  geom_vline(xintercept = 5, linetype = "dashed", color = "gray50") +
  annotate("text", x = 5, y = 0.68, label = "5 days: ~85% of max benefit",
           hjust = 1.1, vjust = 0, size = 3, color = "gray50") +
  scale_x_continuous("Days of search history considered") +
  scale_y_continuous("Model accuracy in predicting conversion",
                    labels = scales::percent, limits = c(0.60, 0.85)) +
  labs(title = "Prediction Accuracy Improves with Search History",
       subtitle = "Diminishing returns after ~5-7 days of context") +
  theme_minimal(base_size = 12)

The key insight here is that the most benefit comes from just a few days of context. After about 7 days, additional history provides minimal added value.

Ad lift per journey stage

# Simulated ad lift by keyword type and journey stage
# Based on patterns described in the paper

journey_stage <- c("Early (>45 days)", "Mid (15-45 days)", "Late (<15 days)", "At purchase")
ad_lift_by_stage <- data.frame(
  stage = journey_stage,
  brand_keywords = c(1.15, 1.30, 1.45, 1.50),
  category_keywords = c(1.05, 1.12, 1.25, 1.35),
  associated_keywords = c(1.20, 1.18, 1.10, 1.05)
)

p4_long <- ad_lift_by_stage %>%
  tidyr::pivot_longer(cols = c(brand_keywords, category_keywords, associated_keywords),
                     names_to = "keyword_type",
                     values_to = "lift") %>%
  mutate(keyword_type = factor(keyword_type,
                            levels = c("associated_keywords", "category_keywords", "brand_keywords"),
                            labels = c("Associated", "Category", "Brand")),
         stage = factor(stage, levels = c("Early (>45 days)", "Mid (15-45 days)", "Late (<15 days)", "At purchase")))

ggplot(p4_long, aes(x = stage, y = lift, fill = keyword_type, group = keyword_type)) +
  geom_line(linewidth = 0.8) +
  geom_point(size = 3.5, shape = 21, color = "white", stroke = 1.5) +
  geom_hline(yintercept = 1.0, linetype = "dashed", color = "gray60") +
  scale_y_continuous("Conversion lift multiplier", breaks = seq(1.0, 1.6, 0.1)) +
  scale_fill_manual("Keyword type", values = c("#2ca02c", "#ff7f0e", "#1f77b4")) +
  labs(title = "Ad Lift Varies by Keyword Type and Journey Stage",
       subtitle = "Brand keywords strongest at purchase; Associated keywords valuable early") +
  theme_minimal(base_size = 12) +
  theme(axis.text.x = element_text(angle = 45, hjust = 1),
        legend.position = "bottom")

The pattern here shows that branded keywords work best at the point of purchase, but associated keywords deliver meaningful improvement earlier in the journey, when the customer is still undecided about the brand.

Implications for LLM-based search

The article ends with an interesting observation: as search evolves toward LLM-based conversational interfaces, platforms will have more context than just a single query. Multi-phase conversations provide rich signals about where a user is in their conversion journey.

This could significantly change the advertising landscape. If the platform can understand the context of a conversation rather than just a keyword, the entire targeting model shifts from ‘who searched for X’ to ‘who is in stage Y of their journey for product category Z’.

Limits

The paper acknowledges several limitations: it uses observational and natural experiments rather than RCTs, focuses on three product categories that are purchased multiple times over a lifetime, and analyzes data from Bing (which has a smaller market share than Google).

My simulation has additional limitations: it is based on published summary statistics rather than raw data, and the ad lift per phase is illustrative rather than estimated directly from the article.

Takeaways

The most useful insight for me is the ‘commitment gap’: identifying users who are willing to buy into the category, but not yet committed to a brand. This is where the associated keywords provide the most value, and where advertisers are currently spending the least.

If you’re running search campaigns, consider the following:

1. Identify associated keywords in your category
2. Testing cost-effectiveness early in the funnel
3. Use minimal search history (5-7 days) to improve targeting without significant privacy costs

Sources

Rothschild, D. M., Needell, C., Veverka, J., and Yom-Tov, E. (2025). Search conversion paths and the missed opportunity of associated keywords. Frontiers in Communication, 10: 1560141.

Related