The openAPPindex project · openappindex.org Working paper, draft v0.1 · September 2026
Status: preliminary and incomplete. This paper is circulated to invite discussion, replication and criticism. It is not submitted for any degree or peer-reviewed publication. All measurements described here are reproducible from scripts and raw data published at openappindex.org/methode, and the corrections we made to our own classification are documented there as well.
Keywords: search costs, platform markets, price transparency, app stores, gatekeepers JEL codes: D82, D83, L13, L86
We measure two properties of a mobile application marketplace that economic theory treats as preconditions for efficient outcomes: whether the supply of products is discoverable, and whether displayed prices state true costs. Using the full cooking-app category of the German iOS App Store as a test case (measurements taken 19 August to 1 September 2026), we find that 81% of apps whose own store description identifies them as recipe apps (579 of 715) do not appear in any of 25 German-language recipe searches, and that at most 92 of at least 991 apps in the category ever reach the first screen of any query. Ranking positions correlate with indexed metadata rather than with observable quality signals. Separately, 77% of the 907 category apps listed as free (German „Gratis") charge for use through in-app purchases, with a median highest purchase of €39.99. We relate these observations to the literatures on consumer search, shrouded prices and platform gatekeeping, and argue that they are consistent with a structural account in which the operators of distribution and advertising jointly profit from low market transparency. The dataset is narrow: one category, one storefront, one platform, a two-week window. We state this limitation prominently and outline the replication programme the findings require, in particular an extension to Google Play, whose different search architecture makes it a natural comparison case.
Textbook welfare results rest on assumptions that are rarely tested against particular markets: participants can learn what is on offer, and prices convey what buyers will pay. Mobile application stores are a useful setting in which to test both assumptions directly, for three reasons. The supply is enumerable in principle, since every product is a listing with a public page. The discovery mechanism is a single search engine controlled by the market operator. And the displayed price is a standardised label chosen by the operator's interface, not by the seller alone.
This paper reports a small, complete measurement of one category in one store: cooking apps in the German iOS App Store. We attempted to enumerate the entire category rather than sample it, first through a systematic search sweep and then by crawling the store's own recommendation graph, and we then read the published price information of every app found. The exercise produced two findings that we did not expect in their magnitude.
First, search misses most of the market. The large majority of apps that describe themselves as recipe apps are unreachable through recipe-related search, and the set of apps that is reachable is small, stable over time, and largely invariant to the query asked. Second, the displayed price is uninformative for most of the category: roughly three quarters of apps labelled as free charge money once opened, at price points the store's category view never displays.
We are deliberately cautious about generality. These are measurements of one category on one storefront over two weeks, made by a small project, with a category boundary that we had to correct four times before the figures stabilised (Section 3.3). We present them as preliminary evidence and as a method that others can run, not as settled fact. Their interest lies in the pattern they form when read against three established literatures: consumer search theory, which predicts that even modest discovery frictions concentrate outcomes; the economics of shrouded prices, which predicts that hidden charges persist under competition; and the platform literature, which explains why the operator of a two-sided market may not have an incentive to remove either friction. Section 6 sets out what a broader dataset, in particular Google Play data, could confirm or overturn.
Four strands of literature frame the measurements.
Consumer search and rankings. Stigler (1961) treats information as a costly good; Diamond (1971) shows that small search costs can move equilibrium prices to the monopoly level, because no individual consumer finds further search worthwhile. In ranked online environments the friction takes a specific form: position itself influences choices independently of relevance. Ursu (2018) identifies this effect causally in hotel search data. For app markets specifically, Ershov (2024) uses a natural experiment in the Google Play store to show that discovery costs and congestion externalities shape entry and the quality of entrants, and Teng (2026), exploiting a 2019 change in the U.S. Apple App Store's search algorithm, finds that rankings significantly affect both app demand and developers' quality provision, and that the self-preferencing embedded in the pre-change algorithm modestly reduced consumer welfare and third-party profits. An early descriptive antecedent is Zhong and Michahelles (2013), who showed that Google Play adoption follows a superstar rather than long-tail distribution and noted that "the discovery of niche apps is still an intractable task." Industry data are consistent with steep positional decay: Sensor Tower's State of ASO reports roughly 61% of search-driven installs going to the top three results and under 4% beyond rank 30.
Price opacity. Gabaix and Laibson (2006) show that firms can shroud part of a price and that shrouding survives competition, because a rival who discloses honestly loses the myopic customers it educates. Ellison and Ellison (2009) document deliberate obfuscation raising margins in a highly price-competitive online market. Blake, Moshary, Sweeney and Tadelis (2021) provide experimental magnitudes: on StubHub, revealing fees only at checkout raised consumer spending by approximately 21% and worsened the quality of choices. Dranove and Jin (2010) survey the disclosure and certification literature and find that credible third-party measurement can correct outcomes where self-disclosure fails.
Platforms and gatekeeping. Rochet and Tirole (2003) and Armstrong (2006) establish that two-sided platforms set structurally skewed prices and that competition between platforms does not guarantee efficiency on them. Wang and Wright (2017) analyse ad-valorem platform fees. Prat and Valletti (2022) model platforms as attention brokers and show that concentration among them tightens the bottleneck through which producers reach consumers, raising advertising prices and harming entrants. Empirically, Decarolis and Rovigatti (2021) show that concentration measurably moves online advertising auction outcomes, and the Competition and Markets Authority (2020) found approximately 80% of UK search and display advertising expenditure accruing to two firms. Policy reports (Crémer, de Montjoye and Schweitzer 2019; Furman et al. 2019; Stigler Center 2019) converge on the diagnosis that digital gatekeeper positions are entrenched; under the EU Digital Markets Act (Regulation (EU) 2022/1925), Alphabet, Apple, Meta and ByteDance are designated gatekeepers. Wen and Zhu (2019) and Zhu and Liu (2018) document how platform owners' dual role distorts the behaviour of the complementors who depend on them. For app stores specifically, Jamison, Tęcza and Wang (2026) estimate the effects of first-party entry on direct rivals using a stacked difference-in-differences design over a Sensor Tower panel of roughly 406,900 apps (2012–2021): on iOS, rivals' downloads and revenue fall by 53.7% and 58.6% at their height within two years of Apple's entry, while equivalent entry on Google Play shows no statistically significant effects — evidence both that the platform's conduct moves market outcomes and that the two stores behave differently. On the advertising side, Ju, Zhao and Aral (2026) use a global shutoff of paid install advertising at a major U.S. mobile game developer to show that paid installs raise store rankings, which in turn generate organic installs (a 20–30% organic decline after the shutoff, with the effect fully mediated by rank): in app stores, paid acquisition purchases organic visibility itself. Beyond app stores, regulators have published the structured analyses closest to our question: the Netherlands Authority for Consumers and Markets (2019) examined Apple's and Google's incentive and ability to influence the availability and findability of apps; the CMA (2022) found the two firms hold entrenched market power over native app distribution (over 90% of UK native app downloads in 2020); and the ACCC (2021) reached parallel conclusions for Australia. None of these measures category recall.
Algorithm audits and machine-mediated discovery. Methodologically, this study is an external algorithm audit in the tradition reviewed by Bandy (2021): repeated, dated queries by an outside party against a ranking system whose internals are unobservable. Web search engines and recommender systems have been audited this way for a decade; we are not aware of a published algorithm audit of app store search. The operator's side of the curtain is visible in Christakopoulou et al. (2026), an Apple paper describing the App Store ranker as optimizing behavioral and textual relevance and validating changes by a worldwide A/B test on conversion rate — a published statement that the ranking objective is conversion, with the largest improvements from synthetic relevance labels arising in tail queries where behavioral signals are sparse. Finally, Motger et al. (2025) evaluate five commercial large language models as an emerging app-discovery channel across thousands of queries, finding ranking criteria only partially aligned with store metrics, stable top recommendations within models, and low agreement between models — early evidence on the channel to which the discovery question is migrating.
To our knowledge, no published study enumerates a complete store category and measures the share of it reachable through search, and none measures "free" labelling against published in-app price lists at category scale. We state this as an absence we could not fill by searching, not as a certainty, and would welcome correction. Our contribution is not theoretical: it is a small, complete, dated measurement of recall and price transparency in one real category, with published instruments, intended to connect these literatures to observable magnitudes.
We define the target population as applications available on the German iOS storefront whose store listing identifies them as cooking or recipe applications. Enumeration proceeded in two stages.
Stage one: search sweep (19 August 2026). We ran 25 German-language queries covering the category's vocabulary (recipes, cooking, baking, meal planning, and appliance- and diet-specific terms) against the store's search interface and recorded every result with its position. The sweep returned 1,312 distinct apps, of which 715 were classified as cooking apps whose own German description identifies them as recipe apps. The best single query surfaced 14% of the apps the whole sweep found; 62.5% of swept apps were reachable through exactly one of the 25 queries.
Stage two: recommendation-graph crawl (21 August – 1 September 2026). Because a search sweep can only find what search returns, we then crawled the store's own "similar apps" links, starting from the swept set and following links until growth was exhausted. The first hop added 243 cooking apps not surfaced by any query; the second hop added 50. Growth fell from +35% to +5% in one step, which we read as approaching closure. The resulting lower bound for the category is 991 cooking apps, of which 941 are served by our index with price data and 50 remain price-unverified.
For each of the 941 indexed apps we read the published in-app purchase list from the store's own product pages (20–23 August 2026, basis corrected 31 August). We record the highest single listed in-app purchase per app. We do not annualise: the store does not reliably publish billing periods, so a figure such as €599.99 is reported as a highest single purchase, never as a yearly price.
Deciding what counts as a cooking app was the largest source of error and we document our failures because they bound the reliability of the figures. Our first classifier matched the German word Rezept, which means both recipe and medical prescription, and wrongly included pharmacy applications. A second pass removed cooking games matching Kochen and Backen; a third caught children's cooking games filed under Education; a fourth (31 August 2026) caught cooking-magazine ePaper apps whose subscription prices had contaminated the price figures. Across the four passes the headline search figure moved from 82.4% to 82.8% to 81.0%, and the price figure remained at 77%. We take the stability of the headlines under reclassification as evidence that the findings do not depend on where the category boundary is drawn, while noting that keyword classification has limits that only manual review or a second rater could quantify.
Denominators differ across findings for a reason we make explicit: the 81% search-absence figure is computed against the 715 search-swept apps, because apps discovered only through the recommendation graph were never in a search result and including them in a "never appears in search" statistic would be circular. Price figures use the 941 indexed apps, since a price is a price regardless of how the app was found.
The store places auctioned advertising slots above organic results (per Apple's developer communications, multiple slots within a single result page since 3 March 2026). These slots are personalised and cannot be independently observed, by us or by anyone outside the operator. All visibility figures below are therefore computed from organic results only and overstate, rather than understate, the visibility of organic listings in the rendered store.
Of the 715 self-described recipe apps found by the sweep, 579 (81.0%) appear in none of the 25 queries' result lists. Extending to the crawled category: across all 25 queries there are 674 first-30 result slots, and 92 of the at least 991 category apps (9.3%) ever occupy any of them. The remaining 897 apps are not ranked low; they are absent.
The apps that search misses are systematically different from those it finds. The median app surfaced by search has a small number of ratings; the median app found only through the recommendation graph has zero. Search does not sample the category; it samples its most established stratum.
The visible set is small, concentrated and stable. Ten apps hold 39% of all first-30 slots across the 25 queries. One app appears in 16 of the 25 first-30 lists, including queries for meal planning and baking, and ranks first for six distinct queries. Two queries' first-30 lists share a median of only 5% of their apps, and 63% of apps that ever enter a first-30 list do so for exactly one query: varying the question varies the periphery, not the core. Re-running all 25 queries after 13 days (1 September 2026) left a median 92% of each first screen unchanged, with an identical seven-app core.
Result quality at the margin is itself informative. On 1 September 2026 the sixth organic result for the query „rezepte app" was an alarm-clock application and the tenth was a dental practice's app, while 885 cooking apps appeared in no first dozen of any query. These are single dated observations; store results move, and ours are timestamped so they can be checked against any later date.
iOS search eligibility derives from an app's name (30 characters), subtitle (30) and a hidden keyword field (100), with some smaller surfaces; the store description is not indexed. The consequences are visible in individual cases, which we report as dated examples and not as judgments of the apps involved. An app with 5,507 ratings and the query term in its subtitle (ZauberMix für Monsieur Cuisine) was absent from the 168 results for „Rezepte" on 19 August 2026; so were apps with 1,879, 827 and 814 ratings whose descriptions or subtitles use the term. An app whose name is nearly identical to the query string held position 8 with 12 ratings. Whatever the ranking optimises, on this evidence it is not proportionate to accumulated user assessment; approximately 160 characters of indexed metadata dominate it. This reading is consistent with the operator's own published account: Christakopoulou et al. (2026) describe the ranker's objective as behavioral plus textual relevance, with changes validated by their effect on conversion rate — an objective in which neither maintenance nor accumulated user assessment appears, and in which tail queries are acknowledged to lack reliable behavioral signal.
Of 907 category apps listed in the store as „Gratis", 697 (77%) charge money through in-app purchases. The median app's highest listed purchase is €39.99. Twenty-six apps list a single purchase above €100, four above €200, and the largest we found is €599.99. All of these prices are published, on pages reached after the install decision is substantially made; none appears in the category or search views where the word „Gratis" does. No individual listing violates store rules. The aggregate effect is that the price signal visible at the point of choice is uninformative for three quarters of the category.
Read against the literature of Section 2, the four results are not independent anomalies but the predicted outcome of a specific structure.
The recall failure (4.1) converts Stigler's search friction into something closer to exclusion: beyond the rendered results there is no channel at higher cost, there is no channel. Given positional install decay of the magnitude the industry reports, a market that is in practice three results deep concentrates demand on whoever holds those results, and the operator auctions the top of them. Apple's own advertising sales material claims conversion rates above 60% for this placement. The stability and query-invariance of the visible set (4.2) indicate that this concentration is not the churn of a responsive ranking but a persistent allocation, and 4.3 indicates the allocation criterion is metadata rather than quality signals. Ursu's ranking results, Teng's demand and quality estimates for this very store, and Ershov's congestion results supply the causal mechanisms by which such an allocation reproduces itself. Ju, Zhao and Aral (2026) close the loop on the paid side: because paid installs raise rankings that generate organic installs, advertising expenditure buys not only the auctioned slot but organic visibility itself, which converts the visibility war into a self-financing feedback rather than a one-off toll.
The price result (4.4) matches the Gabaix–Laibson shrouding equilibrium closely enough to serve as a field observation of it: when 77% of a category shrouds, non-disclosure is the equilibrium strategy, and their model explains why no individual developer can profitably defect to honest labelling. Two features aggravate the standard model here. The shrouding is implemented by the marketplace's display architecture uniformly, not chosen listing by listing; and the operator collects an ad-valorem commission of 15–30% on the shrouded revenue, so the party that could impose transparency is remunerated in proportion to its absence (compare Wang and Wright 2017 on ad-valorem platform fees).
The structural reading follows. Distribution of mobile software is effectively a duopoly; paid reach to mobile users is concentrated among a small group of advertising platforms (industry aggregates attribute roughly two thirds of mobile advertising expenditure to Google, Meta and TikTok; we flag this as a single-source, indicative figure, though company-reported revenues and the CMA's UK findings support the order of magnitude). The store operators sell advertising against their own organic results and tax the transactions that result. Developer expenditure on user acquisition, reported by AppsFlyer at $78 billion in 2025 against roughly $166 billion of consumer spending in the two major stores (Business of Apps), is on this reading not a marketing anomaly but the rent generated by constructed scarcity of visibility, in the sense of Tullock (1967) and, for advertising specifically, Dixit and Norman (1978). Prat and Valletti (2022) describe the welfare consequences of exactly this configuration. We stress that this section is interpretation; the measurements in Section 4 stand independently of it, and alternative readings are discussed below.
The dataset behind this paper is deliberately deep and correspondingly narrow, and we regard the following limitations as binding until replication removes them.
One category. Cooking apps may be unrepresentative: the category is old, crowded and consumer-facing. Categories with professional buyers, fewer sellers or stronger brands may show different recall and price structures. The measurement instruments are category-agnostic and replication across at least several structurally different categories is the immediate next step.
One storefront and one platform. All figures come from the German iOS storefront. Search behaviour, category vocabulary and price levels differ across countries; more importantly, the entire analysis covers only one of the two dominant stores. The extension we consider most informative is to Google Play, for a substantive rather than merely additive reason: Play's search architecture is generally reported to index the full store description, which iOS search does not. The two stores therefore differ on precisely the variable our recall findings turn on. If Play, with description indexing, shows materially higher category recall for equivalent queries, that would support the architectural explanation of Section 5; if recall is similarly low despite the broader index, the explanation would shift towards ranking behaviour and positional demand decay. Either outcome is informative, which is what makes the comparison worth running. There is direct evidence that the two stores are not interchangeable: Jamison, Tęcza and Wang (2026) find large first-party entry effects on iOS and none on Play from the same design, so results from one store cannot be presumed for the other. A Play extension would also test the price-opacity result under a different fee and labelling regime, and Ershov (2024) demonstrates that Play data can support quasi-experimental designs beyond our descriptive ones. A further extension follows from Motger et al. (2025): the same recall measurement can be run against language-model assistants as a discovery channel, asking what fraction of the enumerated category any assistant ever recommends — the natural sequel once the category closure exists.
A two-week window. Our stability result covers 13 days. Longer longitudinal tracking is needed to distinguish a persistently frozen visible set from slow rotation, and to observe how rankings respond to app updates, price changes and seasonal demand.
Classification error. Section 3.3 documents four correction passes. The headline figures were stable across passes, but a keyword classifier has an error rate we have not formally quantified; a manually labelled validation sample would put confidence bounds on the 81% and 77% figures. We note that the store's own results exhibit the same German homonym error we first made, returning a prescription-related result at position one for a recipe query, which we record without inferring anything beyond the observation.
Unobservable advertising slots. Because personalised ad placements cannot be observed externally, our visibility figures describe the organic layer only. Any full welfare statement about the rendered store requires data only the operators hold. This asymmetry of observability is itself a finding of sorts, but it caps what outside measurement can establish.
Interpretation is underdetermined. The structural reading of Section 5 is consistent with our data but not compelled by it. Alternative accounts deserve statement: store curation solves real problems of malware and payment fraud that pre-store distribution did not, and nothing in our data weighs that benefit; a defender of current search design could argue that shallow, stable results economise on consumer attention for the typical user, at the cost of the marginal one; and advertising can be informative rather than purely positional (Becker and Murphy 1993), although a paid slot above organic results the same firm ranks fits the positional case better. Deciding among these accounts requires the broader dataset described above, and ideally operator data that external researchers currently cannot obtain.
In the one corner of the mobile software market we measured completely, the two informational preconditions of an efficient market fail together: most of the supply is invisible to search, and most displayed prices misstate the cost of use. Both failures are consistent with the economics of search, shrouding and platform intermediation, and both are located in design choices of the market operator rather than in the conduct of individual sellers. The magnitudes are large enough that, if they replicate across categories and stores, they bear on current regulatory debates about gatekeeper obligations for transparency and data access.
We circulate these results at a deliberately early stage because their chief value is methodological: the measurements are cheap, the instruments are published, and any reader can re-run them against the store as it stands today. Independent, dated, reproducible measurement of marketplaces that have so far been measured only by their operators seems to us a reasonable public good regardless of how our particular interpretation fares, and we would consider the paper successful if it is primarily the method, and the criticism of it, that travels.
Industry and company sources quoted: AppsFlyer (user-acquisition and app-marketing expenditure, 2025); Business of Apps and store reporting (consumer spend; in-app advertising, 2025); Sensor Tower, State of ASO (positional install shares); Apple Ads sales material (top-of-search conversion claim); Apple App Store Transparency Report 2025 (store size); Apple and Google published commission terms; Apple developer communications on result-page ad slots (March 2026); eMarketer and company reports (advertising revenues, 2025).
Data and code: all queries, crawl scripts, classification rules (including the four documented correction passes) and raw result lists: openappindex.org/methode.
Correspondence: via openappindex.org. This draft is circulated for discussion. Comments, corrections and attempted refutations are explicitly welcome and will be published alongside the data.