Tracksuit technical interview

Can Search Data Complement Brand Tracking?

Internet search data is cheap, behavioral, and available at high frequency. If it moves before brand-tracking outcomes, it may help identify where brand health is changing before surveys fully catch up.

Tracksuit Brand Data

Monthly syndicated surveys of Australian category buyers. For each category, roughly 200 qualified respondents answer brand-funnel questions for the brands in that competitor set.

I use the three provided funnel metrics: prompted awareness, consideration, and preference.

--categories

--unique brand-category combinations

Share of Search from Google Trends

Clean Search Terms Resolve ambiguous brand terms before comparison. Fruits → Fruits Hair
Monday → Monday Hair
Avène → Avene
Collect Pull monthly Google Trends interest by brand and category.
Bridge Use a reference brand to connect five-brand request batches on a common scale.
Compute Share of Search Divide brand interest by total category interest. ${SoS}_{b c t} = \frac{{Search}_{b c t}}{\sum_{j \in c} {Search}_{j c t}}$

A Month × Brand × Category Panel

Link monthly Tracksuit outcomes to Google Trends share of search by matching each brand-category pair to the same calendar month. Each cell is one brand in one category in one month.

This linkage enables three analyses.

Static Do higher-search brands have higher brand metrics?
Dynamic Do search changes correlate with metric changes?
Shocks Do large search moves lead later outcomes?

Static Correlation

Do brands with higher share of search tend to have higher awareness, consideration, and preference?

Noise Reduction To reduce noise, I average both survey outcomes and share of search over the latest six months.

Regression I estimate cross-sectional regressions of brand outcomes on share of search.

Inference Standard errors are clustered at the category level.

{Awareness}_{b c} = α + β \log ({ShareSearch}_{b c}) + ε_{b c}

{Consideration}_{b c} = α + β \log ({ShareSearch}_{b c}) + ε_{b c}

{Preference}_{b c} = α + β \log ({ShareSearch}_{b c}) + ε_{b c}

Interpretation

β represents the difference in brand outcomes associated with a proportional increase in share of search.

Prompted Awareness

Same axes, same brand-category points. As you scroll, the y-value changes from awareness to consideration to preference.

Dynamic Correlation

Static differences are useful, but for a given brand the present share of search is fixed. The actionable question is what happens when a brand increases that share.

Do year-over-year changes in share of search correlate with year-over-year changes in brand metrics?

Noise Reduction I average both survey outcomes and share of search over six-month windows.

Change Measure I compare Oct 2024-Mar 2025 to the same six-month window one year earlier: Oct 2023-Mar 2024.

Inference Standard errors are clustered at the category level.

Δ {Awareness}_{b c} = α + β Δ \log ({ShareSearch}_{b c}) + ε_{b c}

Δ {Consideration}_{b c} = α + β Δ \log ({ShareSearch}_{b c}) + ε_{b c}

Δ {Preference}_{b c} = α + β Δ \log ({ShareSearch}_{b c}) + ε_{b c}

Interpretation

β represents the expected change in the brand metric associated with a proportional change in share of search.

Dynamic · Results

Prompted Awareness

Same axes, same brand-category points. As you scroll, the y-value changes from awareness change to consideration change to preference change.

Do Search Shocks Lead Brand Metrics?

Can large search shocks help identify whether search moves before brand-tracking outcomes?

Motivation Small share-of-search movements may be noise, but larger jumps may reflect meaningful changes in brand activity or attention.

Shock Definition Define a shock as the first brand-category month where the log of six-month average share of search changes by at least 0.25.

Stack Construction Each shock becomes a stack from month -6 to +12. Treated brands are compared with same-category brands that never cross the shock threshold.

High-Dimensional Fixed Effects The model includes brand-by-stack fixed effects and stack-by-month fixed effects, with standard errors clustered by stack.

Stacked event-study design following Cengiz, Dube, Lindner, and Zipperer (QJE, 2019).

Y_{b c s m} = α_{b c s} + λ_{s m} + \sum_{k \neq - 1} β_{k} 1 [m = k] \times {Treated}_{b c s} + ε_{b c s m}

Interpretation

β_k traces how brand outcomes move around the search shock, relative to the month before the shock.

Case study

Case Study: Cocobella Yoghurt

Shock list

Identified Share-of-Search Shocks

Shocks are first brand-category months where the log of six-month average share of search changes by at least 0.25.

Event-study results

Positive Search Shocks Lead Brand Metrics

Event-study results

Negative Search Shocks Do Not Move Later Metrics

Share of Search May Contain Useful Signal

These results are preliminary and suggestive. They point to the possibility that search may be a useful indicator of brand momentum that appears later in survey metrics.

Static +0.11 pp awareness per 1% higher share of search

Dynamic +0.02 pp awareness change per 1% higher search change

Positive Shocks +0.02 pp 0.008 / (0.012 / 0.028) = 0.019 pp per 1% search lift

Across static differences, year-over-year changes, and large positive search shocks, the awareness estimates point in a similar direction. That consistency is the useful signal: search looks promising as an early indication, not as a standalone replacement for brand tracking.

Limitations and Extensions

Search terms are imperfect. Some brands are generic words, and some categories need brand-category-specific intent that Google Trends may not isolate.
Reasons behind changes in share of search are ambiguous. A spike may be good, bad, or simply news-driven.
Shock timing needs more context. The event-study setup would be stronger if the shock dates could be linked to campaign spend, distribution changes, product launches, or earned-media events.
Where search has the most signal. The search-brand relationship may differ across markets, categories, and audience segments, so heterogeneity is central to deciding when the signal is most useful.