Behavior before the act

Moonbrush research note 1

Version

1.0

Published

October 20, 2026

Checked against the product

October 19, 2026

Authors

Moonbrush

Keywordsintention-behavior relation; information search; clickstream; search data; digital footprints; predictive validity

Abstract

Moonbrush reads the behavior that comes before an action, such as what people research and how they shop, and ranks people by it. This note sets out the published evidence for the premise underneath that practice: that behavior before an act carries information about the act.

Three bodies of research bear on it. Meta-analyses in social psychology find that intentions predict behavior substantially but imperfectly, and that changing an intention changes behavior by roughly half as much. Consumer research finds that purchases are preceded by information search, and that the path a person takes through a site separates buying visits from browsing ones. Work on search and behavioral data finds that both forecast outcomes, and that fine-grained behavioral data keep adding predictive value as they grow. The evidence supports the premise as a statement about averages across many people. It does not support claims about any one person, it does not show that research causes buying, and it does not show that a particular model works: those are tested separately.

1. Introduction

Most marketing data record what people have already done. Moonbrush reads what comes first: the subjects people research, how often and how recently, and how that changes over weeks. The bet is that this earlier behavior says something about later behavior.

Stated precisely, the premise is a predictive claim: across many people, those who show a pattern of research before an outcome are more likely than others to reach that outcome. It is not a claim to know what anyone intends, and it is not a causal claim. This note asks how well the published research supports the predictive claim, and where its support ends.

2. From intention to action

2.1 Intentions predict behavior

The most studied account of how deliberate behavior comes about is the theory of planned behavior. It holds that behavior follows from the intention to act and from how much control a person believes they have, and that intentions follow from attitudes, perceived social norms and perceived control (Ajzen, 1991).

The theory has been tested often enough to be summarized by meta-analysis. Across 185 independent studies, its components explained 39% of the variance in intentions and 27% of the variance in behavior (Armitage and Conner, 2001). A review of ten earlier meta-analyses, covering 422 studies, put the average correlation between measured intention and later behavior at 0.53 (Sheeran, 2002). A correlation of that size explains, by its square, about 28% of the variance:

$$r = 0.53 \quad\Rightarrow\quad r^{2} \approx 0.28$$

These studies measure the intentions people state, not their browsing. They establish that intending and acting are linked; sections 3 and 4 turn to behavior that can be observed.

2.2 The gap between intending and doing

Correlations alone cannot show that intentions drive behavior. Experiments can. A meta-analysis of 47 experiments that changed people’s intentions at random found that a medium-to-large change in intention ($d = 0.66$) produced a small-to-medium change in behavior ($d = 0.36$) (Webb and Sheeran, 2006). Intentions matter, and a large share of what people intend does not happen.

The same meta-analysis of the theory of planned behavior found that it explained less of behavior that was observed (an $R^{2}$ of 0.21) than of behavior people reported themselves (0.31) (Armitage and Conner, 2001). Prediction of what people actually do is harder than prediction of what they say they do.

Figure 1. Intentions predict behavior substantially but imperfectly. Panel A: a meta-analysis of 185 independent studies of the theory of planned behavior (Armitage and Conner, 2001). Models explain more of what people intend than of what they do, and less of behavior that was observed than of behavior people reported themselves. Panel B: 47 experiments that changed people's intentions at random (Webb and Sheeran, 2006). A medium-to-large change in intention produced a small-to-medium change in behavior. Values are as reported in the studies.

Values reported in the cited studies

Data behind figure 1

Measure	Value	Source
Intention, variance explained	39%	Armitage and Conner (2001), 185 studies
Behavior, variance explained	27%	Armitage and Conner (2001), 185 studies
Self-reported behavior, variance explained	31%	Armitage and Conner (2001), 185 studies
Observed behavior, variance explained	21%	Armitage and Conner (2001), 185 studies
Change in intention	d = 0.66	Webb and Sheeran (2006), 47 experiments
Change in behavior	d = 0.36	Webb and Sheeran (2006), 47 experiments

3. Search comes before purchase

3.1 Information search

Consumer research has long treated a purchase as the end of a process that begins with looking for information. How much people search before buying varies across product categories and with characteristics of the buyer (Beatty and Smith, 1987). Across channels, many people research in one place and buy in another, a pattern studied as the research-shopper phenomenon (Verhoef et al., 2007).

3.2 What navigation reveals

The path a person takes through a website separates kinds of visit. Classifying store visits by their page-to-page navigation identified distinct shopping strategies, from directed buying through deliberation to hedonic browsing and knowledge-building, which differed in how likely they were to end in a purchase (Moe, 2003). Clickstream research in marketing models browsing, search and purchase from navigation data (Bucklin and Sismeiro, 2009).

4. Behavioral data predict outcomes

4.1 Search data

Search activity forecasts collective outcomes. The volume of web searches predicted the opening-weekend revenue of films, first-month sales of video games and the rank of songs on the charts, days to weeks in advance. Search data improved on baseline models that used other public data, by amounts that ranged from modest to large depending on the domain (Goel et al., 2010). Search query data have also been used to predict the present values of economic indicators (Choi and Varian, 2012).

4.2 Fine-grained behavior

The finer the behavioral data, the more there is to learn from it. In a study of targeted marketing built on 21 million payments by 1.2 million customers to a very large number of individual merchants, adding more traditional structured data produced no appreciable gain, while adding more of the fine-grained behavioral data kept improving prediction across the whole range studied (Martens et al., 2016).

Behavioral records also predict personal attributes. Facebook Likes from 58,466 volunteers distinguished a wide range of attributes, including some that people would consider private, for example with an AUC of 0.93 for gender, and predicted age with a correlation of 0.75 (Kosinski et al., 2013). That finding is evidence of predictive power and also a reason for limits on its use, taken up in note 3.

5. How Moonbrush applies the premise

Moonbrush records behavior as events. An event is one person’s presence against one topic on one day, and as of July 31, 2026 there were 62.6B resolved events. Topics are organized in one tree: 10,312 topics in 26 categories as of August 3, 2026. Each topic is read with two qualities, recency and strength, so a recent burst of research and a long-standing interest are different signals. See Signals and intent topics.

The product reflects the evidence above in two ways.

• Rankings, not predictions about a person. Consistent with a link that holds on average and loosely for any one person, a Gravity score is a rank, and audience groups are the reliable unit. See Reading a score.
• Observed outcomes, not reported ones. Models are evaluated against what people did, which section 2.2 shows is the harder test.

6. What the evidence supports, and what it does not

Table 1. The premise of this note, separated into claims, with the strength of the evidence for each. Grades are defined on the research overview.

Claim	Grade	Basis
Stated intentions predict behavior, substantially but imperfectly	Established	Meta-analyses of hundreds of studies (Armitage and Conner, 2001; Sheeran, 2002)
Changing intentions changes behavior, by less	Established	Meta-analysis of 47 experiments (Webb and Sheeran, 2006)
Information search precedes purchase, and navigation separates buying from browsing	Supported	Consumer and clickstream research (Beatty and Smith, 1987; Moe, 2003)
Search and fine-grained behavioral data forecast outcomes	Supported	Field data across several domains (Goel et al., 2010; Martens et al., 2016)
A given person who researched will act	Not supported	The evidence is about averages and does not license predictions for any one person
Research causes buying, or reaching researchers causes action	Not addressed	A causal question; see note 7
A particular Moonbrush model ranks well	Not addressed by this literature	Tested by evaluation; see note 6

7. Limits and threats to validity

• Coverage. Only behavior that reaches the feed is visible. People with little visible browsing are under-represented, and a model’s coverage is reported on its card.
• Aggregate and individual. The strongest forecasting results above are about aggregates, such as total search volume for a film. Individual-level prediction is noisier, which is why scores are ranks and groups are the unit.
• Reported and observed. Much of the intention literature relies on self-reported behavior. Effects on observed behavior are smaller.
• Decay. Research loses information as it ages. How Moonbrush treats recent episodes and long-run habits differently is the subject of note 2.

8. Questions an auditor should ask

1. For a given model, what share of the population can it see, and how does its coverage differ by age band or region?
2. How does the model’s ranking compare with a baseline that uses only record facts such as age band and geography?
3. How quickly does a model’s ranking lose accuracy as the behavior it reads gets older?

9. Conclusion

The research supports the premise as a statement about averages: behavior before an act carries information about the act, and fine-grained behavior carries more of it as it accumulates. It also sets the limits Moonbrush writes into the product: the link is loose for any one person, it is not causal, and whether a specific model captures it is a matter for evaluation, not for the literature.

References

1. Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2), 179–211. doi.org/10.1016/0749-5978(91)90020-T
2. Armitage, C. J., & Conner, M. (2001). Efficacy of the theory of planned behaviour: A meta-analytic review. British Journal of Social Psychology, 40, 471–499. doi.org/10.1348/014466601164939
3. Beatty, S. E., & Smith, S. M. (1987). External search effort: An investigation across several product categories. Journal of Consumer Research, 14(1), 83–95. doi.org/10.1086/209095
4. Bucklin, R. E., & Sismeiro, C. (2009). Click here for Internet insight: Advances in clickstream data analysis in marketing. Journal of Interactive Marketing, 23(1), 35–48. doi.org/10.1016/j.intmar.2008.10.004
5. Choi, H., & Varian, H. (2012). Predicting the present with Google Trends. Economic Record, 88(S1), 2–9. doi.org/10.1111/j.1475-4932.2012.00809.x
6. Goel, S., Hofman, J. M., Lahaie, S., Pennock, D. M., & Watts, D. J. (2010). Predicting consumer behavior with Web search. Proceedings of the National Academy of Sciences, 107(41), 17486–17490. doi.org/10.1073/pnas.1005962107
7. Kosinski, M., Stillwell, D., & Graepel, T. (2013). Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences, 110(15), 5802–5805. doi.org/10.1073/pnas.1218772110
8. Martens, D., Provost, F., Clark, J., & Junqué de Fortuny, E. (2016). Mining massive fine-grained behavior data to improve predictive analytics. MIS Quarterly, 40(4), 869–888. doi.org/10.25300/MISQ/2016/40.4.04
9. Moe, W. W. (2003). Buying, searching, or browsing: Differentiating between online shoppers using in-store navigational clickstream. Journal of Consumer Psychology, 13(1–2).
10. Sheeran, P. (2002). Intention–behavior relations: A conceptual and empirical review. European Review of Social Psychology, 12(1), 1–36. doi.org/10.1080/14792772143000003
11. Verhoef, P. C., Neslin, S. A., & Vroomen, B. (2007). Multichannel customer management: Understanding the research-shopper phenomenon. International Journal of Research in Marketing, 24(2), 129–148. doi.org/10.1016/j.ijresmar.2006.11.002
12. Webb, T. L., & Sheeran, P. (2006). Does changing behavioral intentions engender behavior change? A meta-analysis of the experimental evidence. Psychological Bulletin, 132(2), 249–268. doi.org/10.1037/0033-2909.132.2.249

How to cite this note

Moonbrush. (2026). Behavior before the act (Research note 1, version 1.0). https://docs.moonbrush.com/research/behavior-before-action/