Working paper · The manifesto · Method & data
A thesis on why the mobile app economy fails the tests of a fair and transparent market — and why that failure is profitable
Tim Meyerdierks · The openAPPindex project · openappindex.org Draft v0.2 · reissued 14 October 2026 (v0.1: 1 September 2026)
Status: working draft, circulated for discussion. Every empirical claim marked "our measurement" is reproducible from the published scripts and carries its measurement date; third-party and company figures are attributed inline; the literature is cited in full at the end. Follows the openAPPindex evidence rules: dated facts, stated denominators, no verdicts on named apps.
The mobile app economy fails the minimum conditions economics sets for a fair and transparent market — symmetric information, contestable entry, and prices that state true costs — and it fails them by construction, not by accident. Its discovery layer, its advertising layer, and its billing layer are operated by the same short list of vertically interlocked firms, which form an oligopoly whose revenues grow with the opacity of the market they intermediate. The resulting equilibrium is allocatively, productively, and dynamically inefficient, and the costs fall on everyone outside the oligopoly: consumers pay in money and attention, developers pay in rents for visibility, and society pays in misdirected innovation. Because the inefficiency is the revenue model, the market cannot be expected to correct itself; correction requires independent measurement infrastructure from outside it.
Four claims, defended in order:
1. Inefficient in information — the discovery mechanism suppresses most of the relevant information (§2). 2. Inefficient in prices — the displayed price systematically misstates the true price (§3). 3. Oligopolistic in structure — distribution is a duopoly and paid discovery a tight oligopoly, run by the same firms that tax the transactions they intermediate (§4). 4. Harmful in aggregate — the equilibrium destroys measurable welfare and cannot self-correct (§5–§6).
The benchmark is old and explicit. The first welfare theorem delivers efficiency only under conditions that include full information and free entry; every deviation from those conditions is a named field of economics with known welfare losses.
Three strands define the test we apply:
- Information economics. Stigler (1961) established that information is a costly good and that search costs alone sustain price dispersion. Diamond (1971) sharpened this into a paradox: with even small search costs, the competitive equilibrium collapses to the monopoly price, because no consumer finds it worthwhile to check one more seller. Akerlof (1970) showed the quality-side analogue: when buyers cannot observe quality, good products are driven out and the market degrades or unravels. - Price transparency. Gabaix and Laibson (2006) showed that firms can shroud attributes — hide part of the price — and that shrouding survives competition: a rival who unshrouds and advertises honest prices educates consumers who then buy the shrouded product anyway, so transparency is punished in equilibrium. Ellison and Ellison (2009) documented deliberate obfuscation raising margins in one of the most price-competitive markets on the internet; Blake, Moshary, Sweeney and Tadelis (2021) showed by randomized experiment on StubHub that revealing fees only at checkout raised consumer spending by roughly 21%. - Platform and attention economics. Rochet and Tirole (2003) and Armstrong (2006) formalized two-sided markets: platforms set structurally skewed prices across the sides they connect, and competition between platforms does not guarantee efficiency on them. Prat and Valletti (2022) modeled the modern version — "attention brokers" that control access to consumers and sell it to producers — and showed that concentration among brokers tightens the attention bottleneck, raises ad prices, excludes entrants, and lowers consumer welfare in the product markets downstream. The policy literature (Crémer, de Montjoye and Schweitzer 2019; Furman et al. 2019; Stigler Center 2019; CMA 2020) converges on the same diagnosis: digital gatekeepers hold entrenched, self-reinforcing positions that ordinary competition does not erode.
A market is fair in this tradition when success correlates with quality and price rather than with privileged access to the buyer; transparent when participants can learn what exists and what it costs at reasonable expense. These are measurable properties. We measured them.
The finding (ours, DE App Store, cooking-app category, 19 Aug – 1 Sep 2026; scripts public): of 715 apps whose own German store description identifies them as recipe apps — found by a sweep of 25 German-language recipe searches — 81% (579) never appear in the results for the obvious one, „Rezepte". Walking the store's own "similar apps" links instead of searching grew the category by a third in one hop; the closure is at least 991 cooking apps, of which 92 — 9% — ever reach any first screen of any query. Ten apps hold 39% of all first-30 slots across all 25 queries; re-running every query 13 days later left a median 92% of each first screen unchanged. The mechanism is documented, ordinary engineering: iOS search indexes roughly 160 characters of metadata — name, subtitle, keyword field — and does not index the description at all, so an app with 5,507 ratings and the query in its subtitle can be absent from the 168 results for that query while a 12-rating app whose name is the query holds position 8 (both read 19 August 2026).
The replication (ours, same storefront, sleep and meditation apps, 1–2 Sep 2026): a second category chosen to be structurally unlike cooking — anchored by a few large subscription products rather than fragmented — and expected, for that reason, to be more discoverable. It is less. Of the 1,155 apps the sweep itself classified, 88% (1,013) never appear in the results for the obvious query, schlaf, against 81% for rezepte; and of the at least 3,578 apps in the closed category, 68% (2,423) are reachable only through the store's similar-apps graph, against 28% in cooking (second-category findings).
What the literature says this is:
- This is Stigler's search-cost problem with the sign flipped. In Stigler (1961) search costs are a natural friction the market grinds down over time. Here the friction is set by the market's operator: the store chooses what is indexed, how deep results go, and what ranks. The cost of discovering the 91% is not high — it is effectively infinite, because no user action reaches it. Sensor Tower's State of ASO (an industry estimate) puts ~61% of search-driven installs in the top three results and under 4% beyond rank 30; conditional on that behavior, an unranked app does not have a high search cost, it has no channel at all. - The consequences are Diamond (1971) at scale. Diamond showed small search costs hand sellers monopoly power. A discovery layer that is, in practice, three results deep hands the holders of those three slots the corresponding rents — which is precisely why the top slot is auctioned (Apple's own sales material claims "conversion rates over 60%" for it). The rent is real, priced hourly, and collected by the party that created the scarcity. - Rankings are not a neutral mirror of preferences. Ursu (2018) showed causally that position itself — independent of relevance — drives what consumers examine and buy. So a ranking that rewards 160 characters of metadata does not merely reflect a concentrated market; it manufactures one. Our overlap measurements make the point empirically: two different queries' first screens share a median of 5% of their apps, while the same ten apps sit atop everything — the shelf does not respond to the question asked. - Akerlof's lemons dynamic operates on the quality dimension the store does not display. The two signals a buyer sees — a star average and the word "Gratis" — say nothing about whether an app is still maintained. A star average is a monument to the past; it keeps selling an app that has not been updated in years (our index snapshot of 23 August 2026 lists apps marked „Gratis", last updated more than 1,700 days before that reading, that still list paid in-app purchases). When care is unobservable, the return to care falls, and the market selects for what is observable: metadata engineering and ad budgets. That is adverse selection against maintenance — lemons, with extra steps. - Ershov (2024) provides the closest controlled evidence: in Google Play data, discovery frictions and congestion externalities directly shape entry and entrant quality — visibility, not quality, is the binding constraint in app markets. Our measurement of the same phenomenon from the outside: in cooking, the median app found by search has a handful of ratings; the median app found only by following the store's own links has zero. The invisible majority is not a tail of junk; it is simply unlit.
Verdict on claim 1: a market in which 9% of verified supply is ever visible (cooking, the one category where we counted first screens), in which visibility is orthogonal to quality and stable across queries and weeks, and in which a second, structurally different category is less discoverable still, fails the information condition not marginally but categorically. The literature's term for what fills the gap is rent, and §4 shows who collects it.
The finding (ours, 20–23 Aug 2026, basis corrected 31 Aug; denominator: the 907 cooking apps listed as „Gratis" among the 941 the index serves): 77% of apps labeled „Gratis" (697 of 907) charge money once opened. The median app's highest single in-app purchase is €39.99; twenty-six exceed €100; the largest is €599.99.
Replicated (ours, sleep and meditation apps, 1–2 Sep 2026; denominator: the 3,351 of the 3,417 „Gratis" apps among 3,578 whose product pages could be read — the other 66 are recorded as not collected and excluded, not counted either way): 77% (2,581 of 3,351) charge money once opened, and the median highest single in-app purchase is €39.99 — to the cent the cooking median. The largest is €1,199.99, for an item its developer labels a yearly subscription; it is reported as the highest single in-app purchase, not converted to a yearly price.
No individual developer is lying — the prices are published, on a page reached after the choice is substantially made. The store displays the word and not the number, and takes a 15–30% commission on the difference between them.
What the literature says this is:
- A textbook shrouded attribute (Gabaix and Laibson 2006). The base good is advertised at zero; the true cost sits in an add-on revealed late. Gabaix and Laibson's central result explains the 77%: shrouding is an equilibrium, not a scattering of bad actors. A developer who prices honestly up front loses the install to a "Gratis" rival before the buyer ever sees either price list; unshrouding is punished, so honest pricing is selected against. When 77% of a category shrouds — and 77% of a second category chosen to be unlike it — that is the theory's prediction confirmed, not exceeded. - The magnitude of the harm is experimentally established. Blake et al. (2021) randomized fee disclosure timing on StubHub: back-loading fees raised spending ~21% and degraded the quality of choices. The app store's architecture back-loads essentially all of the price for 77% of the category — and unlike StubHub's checkout, the subscription price arrives after onboarding, inside the product, at maximum psychological commitment. - Ellison and Ellison (2009) supply the strategic frame: obfuscation is an investment that raises margins by frustrating comparison. Here the obfuscation is not even the developer's investment — it is the store's display choice, uniform across two million apps, and the store is paid in proportion to what obfuscation extracts. An intermediary that earns an ad-valorem commission on shrouded prices has a direct, mechanical stake in shrouding (on the distortive potential of ad-valorem platform fees, see Wang and Wright 2017).
Verdict on claim 2: the single most decision-relevant fact about a product — what it costs — is systematically misstated at the point of decision, in an equilibrium the literature predicted, under a fee structure that pays the market operator for the misstatement. The transparency condition fails, and the failure has an owner.
Strip the market to its plumbing and count the firms.
- Distribution is a duopoly. Two stores mediate effectively all mobile software distribution in the West; on iOS, one. The policy literature is unambiguous that these positions are entrenched and not contestable (Crémer et al. 2019; Furman et al. 2019; CMA 2020), and the EU has made the finding legal fact: Alphabet, Apple, Meta and ByteDance are all designated gatekeepers under the Digital Markets Act (Regulation (EU) 2022/1925) — the same four firms this section counts. - Paid discovery is a tight oligopoly. Google, Meta and TikTok together take roughly two thirds of mobile ad spend (industry aggregate; single secondary source, treat as indicative); Google booked about $295bn in advertising in 2025 and Meta $196bn (company reporting), TikTok around $32bn (a third-party estimate, eMarketer). Apple auctions the top of its own search results, and since March 2026 places multiple ad slots inside a single result page (Apple developer communication). The CMA (2020) found ~80% of UK search and display spend going to two firms and concluded that consumers suffer from weak competition in these markets; Prat and Valletti (2022) supply the mechanism — concentrated attention brokers tighten the bottleneck, raise ad prices, and exclude entrants from product markets. (Decarolis and Rovigatti 2021 show, from the other side, how much auction outcomes move with concentration: market power in ad markets is real enough that buyer-side consolidation measurably shifts prices.) - The same firms hold multiple roles at once. Apple and Google each: operate the store, design the search that makes organic visibility scarce, sell the ad slots that relieve the scarcity, and take 15–30% of the revenue the purchased visibility produces. This is the conflicted dual role — referee, bookmaker, and toll collector — that the self-preferencing literature identifies as structurally corrosive: platform owners that compete with or tax their complementors distort complementor entry and innovation even without overt abuse (Zhu and Liu 2018; Wen and Zhu 2019). Meta and TikTok hold no store at all and need none: they charge for access to the users the stores made hard to reach organically. The four firms are not a cartel and do not need to be one; their incentives are aligned by construction, each collecting a different toll on the same forced detour.
Formally: the discovery layer is a two-sided market (Rochet–Tirole 2003; Armstrong 2006) in which the platform's profit-maximizing structure is exactly what we measure — organic visibility kept scarce and unresponsive (§2), the scarcity sold by auction, and the seller side taxed ad valorem. Two-sided-market theory is explicit that this can be privately optimal and socially inefficient at once; the sides do not internalize each other's losses, and the platform internalizes only what it can bill.
Verdict on claim 3: "oligopoly" is not rhetoric here; it is the count. Two firms control distribution, three to four control paid reach, the intersection of those sets designs the information environment of §2 and §3, and each is a legally designated gatekeeper. The market's opacity is not a bug in this structure. It is the product.
The aggregate accounting (2025 figures; third-party industry estimates, attributed): consumers spent roughly $166bn across the App Store and Google Play (Business of Apps / store reporting). Developers spent $78bn on user acquisition — $109bn on app marketing in total, up 13% year over year (AppsFlyer). In-app advertising inside apps was worth a further $151bn (Business of Apps). For every €2 a person spends on apps, close to €1 is spent by somebody trying to be found — money that never touches the software.
The literature gives each part of this a name:
1. Rent-seeking waste. Visibility spend does not create value; it reallocates a fixed prize — the same finite installs — between contestants (Tullock 1967). Advertising that is persuasive and positional rather than informative is exactly the case where oligopoly generates excessive advertising relative to the social optimum (Dixit and Norman 1978). A $78bn arms race over ~$166bn of demand, with the arms dealer also running the battlefield, is rent dissipation at a ratio the rent-seeking literature would call spectacular. 2. Allocative inefficiency. §2's information failure means the match between user and app is made on 160 characters of metadata and ad budget, not on quality or fit; §3's price failure means the buyer's willingness-to-pay is elicited against a false price. Both mechanisms misallocate — the wrong apps are chosen, at prices that would not have been accepted if stated. Blake et al.'s 21% is a measured lower bound on the second distortion in a milder setting. 3. Productive inefficiency and quality erosion. The developer must recover the visibility spend from a higher price or more in-app advertising — the $151bn attention invoice. Product effort is substituted away from maintenance (unobservable, unrewarded — §2's lemons dynamic) toward metadata engineering and ad operations. The app gets worse in order to afford being found. Ershov (2024) documents the mechanism's grip on entry and entrant quality; Rosen's (1981) superstar economics explains the outcome distribution — tiny visibility advantages compound into extreme revenue concentration disconnected from proportionate quality differences. 4. Dynamic harm. Entry is deterred not by incumbents' efficiency but by the toll structure: a new entrant of arbitrary quality starts invisible (§2), must buy reach from the oligopoly (§4), and surrenders 15–30% of the proceeds. The gatekeeper reports (Crémer et al. 2019; Furman et al. 2019; Stigler Center 2019) all identify this innovation tax as the deepest cost — the products never built because the year of engineering was, as our manifesto puts it, the cheap part. This sits inside the broader documented rise of market power and markups across the economy (De Loecker, Eeckhout and Unger 2020), of which platform gatekeeping is the sharpest instance. 5. The harm is regressive and attention-denominated. The consumer pays three times: in the shrouded price, in the ad load that funds the visibility war, and in the degraded quality of the match. The attention cost falls hardest on users of "free" apps — disproportionately those with the least money — which is the distributional pattern the attention-broker literature predicts (Prat and Valletti 2022).
Verdict on claim 4: by every standard welfare category — allocative, productive, dynamic, distributional — the equilibrium destroys value for every participant outside the oligopoly. The sums are not marginal: the visibility war plus the attention invoice ($78bn + $151bn) are together larger than the consumer spend they contest ($166bn).
Three self-sealing mechanisms, each independently documented:
- The behavioral loop. A search that is three results deep teaches users not to scroll; users who do not scroll make the top slots worth everything; slots worth everything get auctioned; auctioned results get worse, which teaches users — correctly — not to scroll. Our 13-day stability measurement (median 92% of each first screen unchanged; the same seven-app core) shows the loop at rest: there is nothing for an explorer to find. - The shrouded-price trap. Gabaix–Laibson equilibrium: no developer can profitably defect to honest pricing, and the intermediary that could impose transparency is paid ad valorem on the shrouding. The only actor able to fix §3 is the one actor paid not to. - The corpus inheritance. The question which app should I use? is migrating to AI assistants — the natural reset. But an assistant recommends from a corpus, and the available corpus is the one the store already distorted: incumbents loud enough to be in the training data, and vendors loud enough to be in the search results. Our assistant measurements (23–26 August 2026) find the assistants naming a different minority from the store's, but one just as narrow — 53 apps of a 951-app cooking corpus, 5.6% — selected largely by competitor and vendor marketing pages rather than by the store's ranking (assistant recall findings). Absent an alternative corpus, the gatekeeper is replaced and the bias is inherited — harder to see, and no longer even auditable query by query.
The economics of disclosure points at the remedy. Dranove and Jin (2010), surveying the certification and disclosure literature, find that credible third-party measurement changes seller behavior and market outcomes where self-disclosure and platform disclosure fail — and that the credibility of the certifier (independence from what it certifies, method transparency) is the binding constraint. That is the institutional gap openAPPindex is built to fill. No placement is for sale, at any price. Who funds us is published (funding policy). What the index builds is not another ranking inside the distorted candidate set, but the candidate set itself — recall before ranking — with every fact dated, sourced, denominated, and reproducible, readable by people and by the assistants that will make the next decade's recommendations. The stores became bottlenecks because nobody outside could measure them. The thesis of this paper is that the measurement itself is the missing market institution.
Stated per our evidence rules, because a thesis that hides its weaknesses commits the failure it describes.
- Scope. Our measurements cover two categories (cooking, 19 Aug – 1 Sep 2026; sleep and meditation, 1–2 Sep 2026), one storefront (DE), one platform (iOS). The mechanisms — 160 indexed characters, auctioned top slots, "Gratis" labeling, 15–30% commission — are store-wide by construction, but the magnitudes are category-specific until replicated. The second category was chosen to break the first result. The price magnitudes replicated exactly (77%, €39.99 median); the recall magnitudes did not, because they were worse (88% against 81% for the obvious query; 68% against 28% reachable only through the graph), which refutes the hypothesis we went in with that concentration makes a category more discoverable. The 9% first-screen figure is measured for cooking only. It is two categories, not twenty, both consumer-facing, both on one storefront. Replication in further categories — ideally chosen by someone other than us — remains the immediate research program. - The curation defense. Stores did solve real problems: malware, payment fraud, distribution cost. The pre-2008 counterfactual was worse on those axes. The thesis does not deny this; it observes that security curation does not require opaque discovery or shrouded pricing — the bundle is chosen, not forced. - Advertising as information. Becker and Murphy (1993) model advertising as a complement to the good, not waste. This defense fits informative advertising in open media; it fits poorly a positional auction for a slot above organic results the same firm ranks, where the spend's function is displacement — the Dixit–Norman and Tullock cases, not the Becker–Murphy one. - Two-sided pricing skew can be efficient. Rochet–Tirole warn against naive one-sided intuitions; a zero consumer price with seller-side fees is not inefficient per se. Granted — the thesis's claim is not that sellers pay and buyers don't, but that the information environment both sides transact in is degraded by the party pricing it. - Concentration numbers vary by source. The "two thirds of mobile ad spend" figure is a single secondary aggregate and is labeled indicative here as everywhere we use it, and the TikTok figure is a third-party estimate; the company-reported Google and Meta figures and the CMA's 80% UK finding carry the structural argument on their own.
None of these limits touches the core: the welfare theorems' conditions are violated, the violations are measurable, the violators are four designated gatekeepers, and the violation is their revenue.
Our measurements — openAPPindex, DE App Store, iOS: cooking category, 19 Aug – 1 Sep 2026; sleep and meditation category, 1–2 Sep 2026. Method, scripts, raw data and published corrections: openappindex.org/methode. Denominators and dates as stated inline; the full write-up of both categories is the working paper, the replication is in second-category findings, the assistant measurements in assistant recall findings, and the claim-by-claim register is the source table of the manifesto.
Literature
Industry figures quoted (attributed inline where used): AppsFlyer ($78bn UA / $109bn app marketing, 2025); Business of Apps / store reporting ($166bn consumer spend; $151bn in-app advertising, 2025); company reporting (Google ~$295bn, Meta $196bn advertising revenue, 2025) and eMarketer (TikTok ~$32bn, estimate, 2025); Apple Ads sales material ("conversion rates over 60%"); Apple 2025 App Store Transparency Report (2,172,472 apps); Sensor Tower State of ASO (~61% of search installs to top three; <4% past rank 30); Apple and Google published terms (15–30% commission).
Dated 2026-10-14.
- Figures checked against the current sources (working paper, second-category findings, manifesto source table). No cooking figure changed; the 77% now carries its count (697 of 907), and the ratings examples in §2 carry their measurement date. - Corrected: what the 81% measures (§2). v0.1 said the 579 apps appear in none of the 25 searches; they were found by that sweep, so they cannot. The 81% is absence from the results for the obvious query, „Rezepte", as the manifesto and the second-category findings state it. - Second category added where the argument rested on one: sleep and meditation, 1–2 Sep 2026, in §2 (recall: 88%, 1,013 of 1,155; 68%, 2,423 of 3,578), §3 (price: 77%, 2,581 of 3,351; median €39.99; largest €1,199.99) and §7 (scope). Two categories, not twenty. - Removed: Apple's ad business "estimated near $7.4bn" (§4). Here it stood unattributed. It is an eMarketer forecast, now attributed as such in the manifesto; the argument here does not need it. - Corrected: the CMA (2020) "quotation" (§4). The words "weak to the detriment of consumers" were in quotation marks, but we could not find them in the report. The sentence now paraphrases the finding without quotation marks. - Labelled as estimates inline: TikTok's ~$32bn (eMarketer), Sensor Tower's positional install shares, and the 2025 AppsFlyer and Business of Apps aggregates in §5. - Corrected: assistants "recommending the same visible minority from memory" (§6). Our own assistant measurements found otherwise: assistants name a different minority from the store's, just as narrow, picked largely from vendor marketing. The sentence now says what was measured. - Reworded: "apps last updated 1,700+ days ago billing subscriptions today" (§2). Billing is not something we observe; the sentence now states what the dataset shows (still listing paid in-app purchases, at a dated snapshot), and no longer infers that a developer "stopped caring". - Reworded: the certifier's credibility "(… no seller payments)" (§6). It did not match our funding policy, which allows disclosed support from app publishers. The line now reads: no placement is for sale, at any price; who funds us is published.
Draft v0.2. The thesis attacks the system and names companies only via their own published figures or their legal designation as gatekeepers. Named apps appear only as dated, measured examples. Where a figure rests on a single secondary source it is labeled indicative inline.