1Introduction

The DTC supplements category has, over the past 24 months, undergone two structural shifts that we believe are not yet reflected in the operating playbook used by most operators on Shopify. First, paid acquisition has flattened: rising CPMs on the major social platforms and the deprecation of granular targeting have compressed paid-acquisition ROAS toward parity with content acquisition[1, 2]. Second, a non-trivial share of category buyer queries are now answered by large language models, with citations either to a small number of authority publications or, increasingly, to brand properties themselves[3].

The first shift is well documented in the trade press. The second is not. To our knowledge, no prior work has measured, at scale, the extent to which a brand’s probability of being cited inside a generative answer translates into measurable changes in organic acquisition. The closest analogues are early studies of organic search position and click-through rate, but those frameworks do not generalise straightforwardly to LLM answers, which are typically synthesised rather than ranked and which surface a small, model-dependent set of citations.

This paper asks a simple empirical question: for DTC supplements brands operating on Shopify, what fraction of brand-level organic growth in the last 90 days is associated with a brand’s probability of being cited inside generative answers to the queries its category buyers ask? We answer this question with a panel of 614 brands, a structural audit on 47 metrics per brand, and a controlled three-model probe of 25 high-intent buyer queries. Section 2 describes the panel. Section 3 develops the estimating model. Section 4 reports the main results and three robustness checks. Section 5 discusses identification, model drift, and selection. Section 6 concludes with a minimal production schedule that, on our estimates, closes the citation gap in two quarters.

Our contribution is threefold. First, we provide the first large-sample estimate of the marginal effect of generative-answer citation on short-run organic acquisition in DTC supplements. Second, we decompose that effect across four operating levers and identify ingredient and use-case content depth as the single strongest correlate of citation. Third, we release the full benchmark methodology as open data alongside this paper, enabling replication and extension by other research groups.

1.1Related work

Three streams of literature inform our analysis. The first, on generative search and discovery, observes that LLM answers disproportionately cite a small set of high-authority publications and that the long tail of cited domains is unstable across model versions[1, 3]. The second, on DTC growth in nutritional supplements, documents the dominance of trust signals (third-party testing, study citations) in conversion[4] and the role of subscription depth in long-run LTV[5]. The third, the most recent, attempts to operationalise an "AI visibility" metric for brands, but to our knowledge has not been tied to an outcome variable on a sample of this scale[3]. We bridge these streams by directly regressing 90-day organic growth on an operationally measurable citation rate, controlling for the standard set of category covariates.

2Data and sample construction

We work with an internal open-data panel of N = 13,400 Shopify merchants spanning 47 verticals. From this panel we isolate the supplements, sports nutrition, and wellness category, yielding n = 614 brands. For each brand we observe (i) a static structural audit on 47 metrics covering content, schema, trust signals, and subscription mechanics, (ii) a monthly traffic estimate from SimilarWeb (Q2 2026 panel), and (iii) the outcome of a probe of 25 high-intent buyer queries against ChatGPT, Perplexity, and Gemini.

2.1The AI citation rate

Denote by Q the set of 25 high-intent buyer queries selected for the category, and by q ∈ Q a single query. For brand b, the citation rate is defined as

where the indicator function is computed by exact-match against the brand’s primary domain in the cited sources surfaced by each model. Sampling is repeated three times per query, two weeks apart, with the maximum taken to limit transient unavailability. We refer to this as the max-of-three protocol; it is intentionally generous to the brand and tends to bias c upward, so the estimates in Section 4 should be interpreted as a conservative lower bound on the marginal effect of an additional cited query.

2.2Query construction

The 25-query set Q is constructed in two stages. First, we collect the universe of category-relevant queries from Google Keyword Planner with a monthly search volume above 1,000, filtered to high commercial intent using a classifier trained on a hand- labelled set of 4,000 queries. Second, we apply a coverage constraint: the final set must include at least three queries from each of five buyer-journey buckets (symptom, ingredient, comparison, dosage, side effects). The full query list is reported in Appendix A and is held fixed across all sampling waves to eliminate composition effects.

2.3Traffic measurement

Organic visits are sourced from the SimilarWeb desktop+mobile panel for the May–July 2026 window. The 90-day growth variable g_b is computed as the log-difference of mean daily organic visits between the trailing 30 days and the leading 30 days of the window. We exclude brands whose panel coverage falls below the SimilarWeb confidence threshold (n = 41 dropped), yielding the final analytic sample of n = 573 brands. Results are unchanged when we re-include the dropped brands using imputed traffic from a sister panel (Similarweb-2).

3Model

Let g_b denote the 90-day growth in organic visits to brand b. We estimate the linear model

where x_b is a vector of controls (log baseline traffic, brand age, sub-category fixed effects) and ε_b is an idiosyncratic shock. Identification rests on the assumption that citation rate is, conditional on x_b, weakly exogenous to short-run growth. We return to this assumption in Section 5.

The parameter of interest is β, the marginal effect of an additional percentage point of citation rate on 90-day organic growth. Under the null hypothesis that LLM citations are a passive consequence of pre-existing brand authority and have no causal effect on acquisition, we expect β = 0. Under the alternative that citations route incremental demand to the cited brand, β > 0. The magnitude of β is economically informative: at β = 0.41, moving from the category-median citation rate (4%) to the 90th-percentile rate (41%) is associated with a ≈15 percentage-point lift in 90-day organic growth, holding controls fixed.

We estimate (2) by ordinary least squares with heteroskedasticity-consistent standard errors clustered at the sub-category level. As an auxiliary specification we also report a decomposition of c_b into its four operating- lever inputs (Section 4.2), which sheds light on which content investments most predict downstream citation.

4Results

4.1Baseline estimate

Across the 614 brands in the sample, the citation rate c_b is highly skewed: the median brand is cited in 1 of 25 queries, while the 90th-percentile brand is cited in 10 of 25. The unconditional distribution is well-approximated by a Beta distribution with shape parameters (0.6, 8.4), reflecting a long tail of zero-citation brands and a thin head of heavily-cited incumbents. Pooled across the sample we obtain a point estimate β̂ = 0.41 (s.e. 0.04), α̂ = 0.06, and R² = 0.62. The intercept α̂ is economically small and statistically distinguishable from zero only at the 10% level, which we interpret as evidence that brands with zero category citation experience essentially no organic growth tailwind over the 90-day window.

Figure 1 plots binned means of the relationship between c_b and g_b. The fit is close to linear across the full support of the citation rate, with a mild concavity above c ≈ 0.65 consistent with a saturation effect at very high citation density.

4.2Lever decomposition

Table 1 reports the marginal correlation between each lever score and the citation rate. Ingredient and use-case content depth is the single strongest correlate of citation, consistent with the structural observation that LLMs preferentially surface comparison and ingredient pages[3, 4]. A first-stage regression of c_b on the four lever scores delivers an R² of 0.51, with content depth alone accounting for 48% of the explained variance, trust signals 27%, subscription depth 14%, and the residual share split across schema markup and brand age.

The asymmetry across levers is itself a contribution: it suggests that operators face a clear ordering of marginal returns to investment. Each additional dedicated ingredient page is, in our sample, associated with a 0.012 increase in citation rate (p < 0.01), meaning that the median brand could close the gap to the 90th-percentile citation rate by publishing approximately 31 additional ingredient and use-case pages — a target that is within reach of a single full-time content operator over a two-quarter horizon.

4.3Robustness

We perform three robustness checks. (i) Tail trimming: dropping the top decile of citation rate yields β̂ = 0.38 (s.e. 0.05), confirming that the result is not driven by a handful of heavily-cited incumbents. (ii) Quantile regression: estimating the median regression yields β̂_0.5 = 0.36 (s.e. 0.05). (iii) Placebo: estimating the same regression on a citation-orthogonal query set (25 queries drawn from the office-supplies category) yields β̂ = 0.02 (s.e. 0.03), statistically indistinguishable from zero, which we read as evidence against a generic "popular brands get cited everywhere" story.

5Discussion and limitations

Three threats to identification warrant note. (i) Reverse causality: fast-growing brands may attract more inbound links and therefore more citations. We partially address this with a placebo regression on a category-orthogonal query set, where β̂ is statistically indistinguishable from zero, but we cannot fully rule out a confounding flow of unobserved brand quality that drives both citations and growth. (ii) Model drift: LLM retrieval is updated continuously; our snapshot reflects Q2 2026, and we make no claim about the stability of the estimated effect into future quarters. A repeat measurement is planned for Q4 2026. (iii) Selection: the panel covers only brands operating on the Shopify platform. Brands on alternative platforms may have a systematically different relationship between citation and growth, although we see no theoretical reason this should be the case.

A fourth concern is measurement error in c_b arising from non-deterministic LLM outputs. Repeated probes show a test–retest correlation of 0.91 over two-week windows, suggesting that the noise is modest. Attenuation bias in the OLS estimate is therefore expected to be small (≈9%); the true β may be slightly larger than the 0.41 we report.

We close with an operational interpretation. The estimated coefficient implies that, on the margin, a brand currently cited in 1 of 25 category queries that invests sufficient content to reach the 90th-percentile citation rate of 10 of 25 should expect a 90-day organic growth tailwind of roughly 14.8 percentage points, conditional on the controls in (2). We provide a worked example of such a content production schedule in Appendix B.

6Conclusion

For DTC supplements brands operating on Shopify, the brand-level probability of being cited inside generative answers is, on the margin, associated with a 0.41-percentage-point lift in 90-day organic growth per percentage-point of citation rate. The lever most closely associated with citation is ingredient and use-case content depth, not paid acquisition spend. Operators are likely under-investing in the production of comparison and ingredient pages relative to the marginal return, and the gap is largest among brands that have historically over-indexed on paid social.

The result is, in our view, not a temporary artefact of the current generation of retrieval-augmented LLMs but a structural feature of the shift from ranked search to synthesised search. We expect the weight on citation in the brand-growth function to increase, not decrease, as generative interfaces are integrated more deeply into the buyer journey.

7Acknowledgments

We thank the founders of the brands in our panel for permitting the use of their public-facing data in this study, and Ryze Research Labs’ data engineering team for maintaining the open-data benchmark pipeline. All errors are our own. The views expressed are those of the authors and do not necessarily reflect those of Ryze Research Labs LLC.

References

1. Anderson, C. (2024). Generative search and the new shape of discovery. JMR, 61(3), 411–429.
2. Bansal, R., & Patel, M. (2025). The decline of paid social as a DTC growth lever. WSJ Strategy Notes.
3. Garcia, L., et al. (2025). Probing LLM retrieval: a methodology for brand citation audits. arXiv:2502.10421.
4. Iyer, V., & Smith, P. (2024). Trust signals and conversion in nutritional supplements. JCR, 51(2), 217–238.
5. Liu, Y., & Hofmann, R. (2026). Subscription depth and lifetime value in DTC categories. MSI Report 26-103.
6. Ryze Research Labs (2026). The 13,400-store Shopify open-data benchmark, Q1–Q2 2026. RRL-DATA-2026-002.
7. Nakamura, K., & Olafsson, B. (2025). Synthesised answers and the decay of organic clicks. SearchEng. Letters, 12(4), 89–104.
8. Park, J., et al. (2025). Brand citation stability across LLM versions: a longitudinal study. arXiv:2509.04812.
9. Rivera, T., & Schmid, F. (2024). Beta regression for bounded share data. J. Econometrics, 240(1), 55–78.
10. SimilarWeb (2026). Traffic panel methodology, v9. Tech. Note SW-2026-MET-09.
11. Tran, N. (2025). Schema markup and machine readability of e-commerce pages. Web Conf. ’25, 1144–1156.
12. Walsh, P., & Kim, H. (2026). Quantile regression with clustered data: a practitioner’s guide. Stata Journal, 26(1), 12–34.

AAppendix · The 25-query probe set

The probe set Q is held fixed across all sampling waves. Selection criteria are described in Section 2.2. A representative subset is reproduced below; the full list is published with the replication archive at the citation above. Symptom queries: "best probiotic for bloating", "magnesium for sleep", "what helps inflammation"; Ingredient queries: "ashwagandha vs rhodiola", "is quercetin safe long term"; Comparison queries: "Symprove vs Bio-Kult", "best probiotic UK 2026"; Dosage queries: "how much vitamin D3 per day", "creatine loading protocol"; Side-effect queries: "ashwagandha side effects", "magnesium glycinate vs citrate side effects".

BAppendix · Worked example: closing the citation gap

Consider a median brand with c = 0.04 (1 of 25 queries), targeting the 90th-percentile rate c = 0.41 (10 of 25) over two quarters. Using the lever decomposition in Section 4.2 and the marginal citation lift of 0.012 per ingredient page, the arithmetic gap closes at ≈31 dedicated ingredient and use-case pages, assuming linear extrapolation. At a sustainable production cadence of 2 well-researched pages per week per content operator, the gap closes in approximately 16 weeks. Under the baseline estimate β̂ = 0.41, the associated 90-day organic growth tailwind, conditional on controls, is roughly 14.8 percentage points. We caution that this is an in-sample extrapolation and not a causal forecast; the realised tailwind for any single operator will depend on the quality of the pages published, sub-category dynamics, and the state of the underlying LLM corpora during the measurement window.