{"id":"db7a18a9-8148-4ddc-aaf3-6724735f71c8","arxiv_id":"1908.07489","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Under multinomial logit demand with seller-set prices, the revenue-maximizing display rule is to show only the top-quality products, while the welfare-maximizing rule is to show all products; Cournot competition also yields top-quality rules.","lead":"This paper studies how an online platform should choose which sellers to show when sellers, not the platform, set prices. It finds that showing every product maximizes total welfare, while showing only the top-quality products maximizes platform revenue, and the best cutoff can be computed quickly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dominant-seller revenue proof relies on figure-based inequalities and invokes Lemma 5 outside its q1<0.5 hypothesis; Theorem 4 is conditionally supported.","rationale":"The paper's social-welfare result (Theorem 3) is well supported, and the Cournot results (Theorems 6-7) are proved with a clean quasi-convexity argument on g(w). The scientific value of the paper is the revenue theorem for Bertrand competition (Theorem 4), which is the basis of the claimed linear-time quality-order mechanism. That theorem's proof has two layers: Lemma 5 for q1<0.5, and Appendix H for q1>=0.5. The first layer already contains a 'by plotting h(q1)' step, but the second layer is more fragile: Lemma 9 relies directly on a figure for its root count, and Theorem 9 applies Lemma 5 in a regime where its q1<0.5 assumption is false. This is exactly the case the authors singled out as needing more complicated analysis, so the gap is not peripheral. I therefore agree with the reader that the verdict should be conditional: accept the structure and the non-dominant-seller case, but require completion of the dominant-seller proof before full acceptance. The numerical check is decisive in that a counterexample would refute the theorem, while a clean pass would justify the extra effort of converting the figure-based steps into analytic inequalities.","tokens_in":27953,"tokens_out":7908,"duration_ms":73590,"concrete_test":"Run a high-precision numerical search over the dominant-seller domain, sampling many combinations of theta1 and q0 in [q0^min,q0^max], and compute the sign pattern of re'(q0) to test Lemma 9's asserted at-most-one/at-most-two intersection structure; also compute min hbar(q1) on [0,0.5] with exact interval arithmetic. If any sampled point violates the asserted quasi-convexity or gives hbar<0, Theorem 4 fails; if none does, replace the figure citations with analytic inequalities and re-prove Theorem 9 without citing Lemma 5 in the q1>=0.5 region.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim Theorem 4 depends on quasi-convexity of the equilibrium revenue re(q0) in the dominant-seller regime q1>=0.5; Appendix H is supposed to supply this but does not do so rigorously. Lemma 9's root-count argument for re'(q0)=0 is settled by 'observe from the figure' (Fig. 4), and the inequalities hbar(q1)>0 (Fig. 5) and W(x)>=2x/(e+x) (Fig. 6) are asserted as 'tedious calculations' backed by plots rather than derived. More than a stylistic issue: in the proof of Theorem 9, Case B, the authors construct a virtual product j' and write 'We have shown in Lemma 5 that such revenue function reb(q0) is quasi-convex' over [q0^min', q0^max']. Lemma 5's own hypothesis is q1<0.5, but the interval [q0*, q0max'] on which the claim is then used is exactly where q1>=0.5 (q0* is defined by q1=0.5). Thus the proof invokes a lemma outside its stated domain. The top-k theorem may still be true and repairable, but as written the dominant-seller branch is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies a platform's search segmentation problem under a discriminatory control model: the platform chooses which products to display to a representative buyer, while sellers set prices and engage in Bertrand competition. Buyer demand follows the multinomial logit (MNL) choice model. The authors prove that for social welfare maximization, displaying all products is optimal; for revenue maximization, the optimal mechanism is to display the top k* products by quality, where k* is computable in linear time. They extend the analysis to Cournot competition, where both objectives lead to top-k threshold mechanisms. The results rely on a closed-form characterization of the Bertrand equilibrium using a V function related to the Lambert W function, and on quasi-convexity of equilibrium revenue as a function of the outside-option demand.","tokens_in":28195,"tokens_out":17520,"duration_ms":153011,"significance":"If the central theorems were fully established, the paper would make a valuable contribution: it reduces a combinatorial display-optimization problem with an exponential search space to a linear-time threshold rule and gives a clean economic interpretation via the quality-order mechanism. The equilibrium characterization and the social-welfare result under Bertrand are solid in structure, and the Cournot extension broadens the scope. The paper is transparent about its assumptions and clearly indicates where the q1<0.5 condition is relaxed. However, the revenue-maximization theorem in the dominant-seller regime (q1>=0.5) rests on proof steps in Appendix H that are either asserted from figures or invoke lemmas outside their stated domain. Until those steps are supplied with complete analytic arguments, the headline revenue result is conditional.","major_comments":[{"comment":"In the proof of Lemma 9, after the derivative expression around Eq. (39), the authors assert that at q0 = q0^min, \"the demands satisfy the relation: 2 x q1 = 1 - q0^min\". This equality holds only when the two displayed products have identical demand, i.e., identical quality. For a generic quality vector, q2 differs from q1, and the subsequent derivation of re'(q0^min) = A(-3 q0^min - (q0^min)^3) does not follow. Because this endpoint evaluation is the first step toward proving quasi-convexity of the revenue function for the k* = 2 case, the proof of Lemma 9 is invalid as written.","section":"Appendix H, Lemma 9"},{"comment":"The appendix establishes load-bearing facts by appeal to figures rather than analytic derivation: Lemma 9 counts intersections of q1(q0) and V(q0 exp(theta1-1)) by \"observing\" Figure 4; Lemma 10 asserts hbar(qi) > 0 for 0 <= qi < 0.5 with a plot in Figure 5; and Theorem 9 asserts W(x) >= 2x/(e+x) on [exp(-1), 2 exp(2)] after \"tedious calculations\" supported only by Figure 6. These inequalities and root counts are necessary for quasi-convexity in the dominant-seller regime and for excluding the interval [q0*, q0^max]. The authors should replace each of these graphical arguments with a complete analytic proof.","section":"Appendix H, Lemmas 9-10 and Theorem 9"},{"comment":"In Case B, the authors construct a virtual product j' and state that Lemma 5 shows the corresponding revenue function reb(q0) is quasi-convex on [q0^min', q0^max'], and then use this statement on [q0*, q0^max']. Lemma 5's own hypothesis is q1 < 0.5. On [q0*, q0^max'], however, q0 >= q0*, so q1 = V(q0 exp(theta1-1)) >= 0.5; the interval is precisely the regime excluded by Lemma 5. Thus a lemma is invoked outside its stated domain. The proof must either extend the quasi-convexity statement to the q1 >= 0.5 regime or provide a different argument for this interval.","section":"Appendix H, Theorem 9, Case B"}],"minor_comments":[{"comment":"The bullet list says \"re(q0) is quasi-concave in q0\" but the surrounding text and Lemma 5 state \"quasi-convex\"; this inconsistency should be corrected.","section":"Section 4.2, design rationale"},{"comment":"The sentence introducing the ordering says \"non-decreasing order\" but the displayed inequality is theta1 >= theta2 >= ... >= thetan, which is non-increasing. The terminology should match the inequality.","section":"Section 2"},{"comment":"The proof relies on two properties of f(q) (f(q)+f(1-q)=1 and the inequalities f(q) <= q and f(q) <= f2(q)) that are stated after plotting Figure 2 but not proved. These properties are used to bound the expression in (31); provide analytic proofs for completeness.","section":"Appendix D, proof of Lemma 3"},{"comment":"The property that g(q) is decreasing on [0,0.5] with g(0)=1 and g(0.5)=-1/3 is asserted with \"We can verify this property by showing the first derivative g'(q) is negative\" but the derivative computation is not shown. Include the computation or a reference.","section":"Appendix F, proof of Lemma 5"},{"comment":"The closed-form expression for q1(q0) (the root of re'(q0)=0) is typeset ambiguously: the denominator 2(q0+2) appears to multiply only the square-root term. Clarify that it multiplies the entire right-hand side.","section":"Lemma 9 proof"},{"comment":"The proof claims that G(p) is an ordinal potential and that best-response dynamics converge in a finite number of steps because the potential has a finite value. The ordinal potential property (30) is asserted without proof, and a strictly increasing sequence in a finite-valued potential need not be finite in a continuous strategy space. A rigorous convergence argument is needed.","section":"Lemma 2"}],"recommendation":"major_revision","confidential_remarks":"Dear Editor, the paper addresses an interesting and important question, and the main results are plausibly correct. However, the proof of the headline revenue theorem in the dominant-seller regime currently has a concrete algebraic error in Lemma 9 and several appeals to figures in place of analytic derivations. These issues are fixable but require substantial additional appendix material. The social-welfare theorems are in better shape. I recommend major revision rather than rejection. No concerns about novelty or attribution; the citation practice appears normal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper has a clean and plausible headline result — in a two-sided market with MNL demand and Bertrand sellers, a revenue-maximizing platform should display the top k* products by quality, and k* is computable in linear time. If it holds, that genuinely extends the assortment-optimization literature, where prices are fixed and assortments are ranked by price rather than quality. The social-welfare half (display all products) is proved in detail, and the Cournot extension is clean. The q1 < 0.5 revenue case is also convincing: Lemma 5's quasi-convexity proof is analytic and Lemma 6 uses it correctly. Citations look appropriate, and there are no fitted parameters or circular steps.\n\nThe soft spot is the dominant-seller case, q1 >= 0.5, which is load-bearing for Theorem 4. Lemma 9 in Appendix H claims quasi-convexity of the revenue function but verifies the key root-count and inequality facts by referring to Figures 4 and 6 rather than deriving them. The inequality W(x) >= 2x/(e+x) is asserted with 'tedious calculations' plus a plot. Then in Theorem 9, Case B, the proof invokes Lemma 5 to get quasi-convexity of the virtual-product revenue function on an interval where q1 >= 0.5 — exactly outside Lemma 5's stated hypothesis. That is a real proof gap, not a stylistic issue. The fix is probably easy: Lemma 9 already handles the k=2 revenue function, and the same argument should cover the virtual product, or the analytic inequality can be supplied. But as written, the dominant-seller branch is not established.\n\nOne minor side issue: Lemma 2 claims best-response dynamics converge in a finite number of steps, but the proof only shows ordinal potential structure with a continuous potential, which yields convergence, not finite-step convergence. This is peripheral and does not affect the main theorems.\n\nWho this is for: people working on platform market design, assortment planning with endogenous prices, and competitive revenue management. It deserves a serious referee. The central result is likely true and worth publishing, but the referee should require a complete proof of the q1 >= 0.5 case before acceptance.","headline":"A likely-true top-k revenue theorem with a real proof gap in the dominant-seller case; the rest is solid and the paper deserves engagement.","tokens_in":28681,"tokens_out":4182,"would_cite":true,"duration_ms":40076,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An online platform that can choose which products to show should display all products to maximize social welfare, but only the top $k^*$ quality products to maximize revenue; the paper proves both under Bertrand competition with…","keywords":["search segmentation mechanisms","online platform markets","Bertrand competition","multinomial logit demand","revenue maximization","social welfare maximization","quality-order mechanism","Cournot competition"],"falsifier":"Construct a Bertrand-MNL instance with $n=4$ products where the top product's equilibrium share satisfies $\\bar q_1\\ge 0.5$ (e.g., choose the highest quality $\\theta_1$ large), compute $re(\\bar q_0)$ from Theorem 2 over the feasible interval $[\\bar q_0^{\\min},\\bar q_0^{\\max}]$, and check whether any interior stationary point has negative second derivative; if so, quasi-convexity fails and the replacement argument in Theorem 4 no longer goes through. Equivalently, enumerate all $2^n$ subsets for such quality vectors and look for a case where the best revenue set is not a quality-ordered prefix.","tokens_in":27735,"feed_emoji":"🛒","tokens_out":9598,"duration_ms":86743,"temperature":0.7,"pith_summary":"The paper studies an online platform that controls only which products appear in a buyer's search results, while sellers set prices competitively and buyers choose according to multinomial logit demand. It claims that for social welfare, the platform should display every product, but for platform revenue it should display only the top $k^*$ products ranked by quality, where $k^*$ is determined by the full quality vector and computable in linear time. Under Cournot competition, the paper argues, the same quality-threshold form is optimal for both objectives, though welfare maximization there may also omit low-quality products. If correct, the combinatorial problem of choosing a display set from $2^n$ possibilities collapses to checking $n$ quality-ordered prefixes, and a revenue-maximizing interface would never show a lower-quality product while hiding a higher-quality one.","feed_headline":"Displaying only top-quality products maximizes platform revenue","feed_subtitle":"With seller price competition, showing fewer, better listings maximizes revenue; showing everything maximizes welfare.","key_machinery":"The load-bearing object is the equilibrium revenue function written in terms of the outside option's demand, $re(\\bar q_0)=\\sum_{i\\in S}\\bar q_i/(1-\\bar q_i)+1/(\\bar q_0+\\sum_{i\\in S}\\bar q_i)-1$, together with the demand-link function $V(x)$, the unique $v\\in(0,1)$ satisfying $v e^{v/(1-v)}=x$. Because $V$ is increasing, a higher-quality candidate product is equivalent to a smaller feasible $\\bar q_0$, so the decision of whether and which product to add becomes a one-dimensional optimisation over $\\bar q_0$. Quasi-convexity of $re(\\bar q_0)$ forces the optimum to an endpoint of the feasible interval, meaning the best choice is either to add nothing or to add the highest-quality unselected product; this endpoint logic is what converts subset choice into quality-ordered prefixes. In the Cournot extension the analogous role is played by the Lambert W function through $w_i=W(e^{\\theta_i-1})$, and quasi-convexity of a scalar function $g(w)$ in the proof of Lemma 7 drives the top-$k$ replacement argument.","core_discovery":"The central claim is that under Bertrand price competition with MNL demand, the equilibrium outcome of any displayed set $S$ can be described by a single outside-option demand $\\bar q_0$ solving $\\sum_{i\\in S} V(\\bar q_0 e^{\\theta_i-1})=1-\\bar q_0$, where $V(x)$ is the unique $v\\in(0,1)$ with $v e^{v/(1-v)}=x$. Expressing equilibrium revenue as a function $re(\\bar q_0)$, the paper proves it is quasi-convex over the feasible interval, so its maximum over choices of which product to add occurs at an endpoint: either add no product or add the highest-quality available one. Iterating this replacement argument yields Theorem 4: revenue maximization displays the top $k^*$ products, and $k^*$ is found in linear time by evaluating the revenue of each quality-ordered prefix. For social welfare in the Bertrand game, the paper proves the welfare function decreases in $\\bar q_0$, so adding any product improves welfare and all products should be displayed. The Cournot results use the same top-$k$ structure, now mediated by Lambert-W weights $w_i=W(e^{\\theta_i-1})$, with Theorems 6 and 7.","pith_inferences":["The same quality-prefix logic suggests a testable design rule for real platforms using price competition: when the objective is take-rate or commission revenue, ranking or displaying by quality alone should beat display rules based on predicted revenue scores, because prices already incorporate surplus extraction.","If quasi-convexity fails for some quality vector with a dominant seller, the optimal display might require checking non-prefix sets; a natural computational experiment is to enumerate all subsets for $n\\le 8$ and compare the best subset against the top-$k^*$ rule across random quality vectors, especially with large $\\theta_1$.","The welfare/revenue divergence has a regulatory reading: a revenue-maximizing platform will systematically hide low-quality sellers, while a surplus-maximizing regulator would want them shown; comparing displayed sets to quality rankings could reveal which objective the platform is optimizing."],"forward_implications":["A revenue-maximizing platform can determine its display set in $O(n)$ time: compute the equilibrium revenue for each prefix $\\{1,\\ldots,k\\}$ and pick the best $k^*$.","No display set that maximizes revenue can skip a high-quality product in favor of a lower-quality one; the optimal display is always a quality-ordered prefix.","For social welfare under Bertrand competition, search segmentation never helps: hiding any product can only reduce total welfare, so the platform should show all sellers.","Under Cournot competition, welfare maximization may hide some low-quality products, but the displayed set remains a quality-ordered prefix rather than an arbitrary subset.","Because quality determines inclusion, sellers have a long-run incentive to improve product quality in order to enter the displayed set, which increases equilibrium revenue and welfare."],"supporting_citations":[{"why":"Supplies the multinomial logit model that yields the softmax demand probabilities from which seller payoffs and equilibrium conditions are derived.","marker":"[28]"},{"why":"Establishes existence and uniqueness of the Bertrand-Nash equilibrium and the first-order-condition characterization used as the equilibrium constraint.","marker":"[21]"},{"why":"Introduces the Lambert W function used in the closed-form equilibrium expressions for prices and demands in both Bertrand and Cournot settings.","marker":"[14]"},{"why":"Provides the concave-game equilibrium existence and uniqueness result used for the Cournot competition extension.","marker":"[33]"},{"why":"Solves assortment optimization under MNL with fixed prices, the closest benchmark that this paper extends by making prices endogenous and ranking by quality.","marker":"[37]"},{"why":"Derives the MNL choice probabilities under i.i.d. Gumbel utility shocks, grounding the demand model.","marker":"[4]"}],"fun_headline_variants":["Top-quality listings maximize revenue, not welfare","Revenue peaks when platforms show only top sellers","Quality threshold for revenue, full display for welfare","Optimal search: filter quality for revenue, show all for welfare","Platform revenue maximized by top-k product display"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on assuming that the equilibrium revenue curve is quasi-convex (no interior valley) over the whole feasible range of outside-option demand, including the case where one seller commands at least half the market; the paper supports this case in Appendix H with graphical intersection and inequality checks rather than a fully algebraic derivation, and if the curve can dip then replacing a lower-quality displayed product with a higher-quality hidden one might not increase revenue.","fun_headline_variants_meta":{"raw":{"variants":["Top-quality listings maximize revenue, not welfare","Revenue peaks when platforms show only top sellers","Quality threshold for revenue, full display for welfare","Optimal search: filter quality for revenue, show all for welfare","Platform revenue maximized by top-k product display"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1593,"prompt_tokens":1005,"completion_tokens":588,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":515}},"tokens_in":621,"tokens_out":588,"duration_ms":6632,"temperature":1.0,"reasoning_tokens":515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:18:17.648926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a Bertrand-MNL instance with $n=4$ products where the top product's equilibrium share satisfies $\\bar q_1\\ge 0.5$ (e.g., choose the highest quality $\\theta_1$ large), compute $re(\\bar q_0)$ from Theorem 2 over the feasible interval $[\\bar q_0^{\\min},\\bar q_0^{\\max}]$, and check whether any interior stationary point has negative second derivative; if so, quasi-convexity fails and the replacement argument in Theorem 4 no longer goes through. Equivalently, enumerate all $2^n$ subsets for such quality vectors and look for a case where the best revenue set is not a quality-ordered prefix.","supporting_citations":[{"cited_title":"McFadden","cited_arxiv_id":null,"evidence_quote":"Supplies the multinomial logit model that yields the softmax demand probabilities from which seller payoffs and equilibrium conditions are derived."},{"cited_title":"Gallego, W","cited_arxiv_id":null,"evidence_quote":"Establishes existence and uniqueness of the Bertrand-Nash equilibrium and the first-order-condition characterization used as the equilibrium constraint."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Lambert W function used in the closed-form equilibrium expressions for prices and demands in both Bertrand and Cournot settings."},{"cited_title":"Talluri and G","cited_arxiv_id":null,"evidence_quote":"Solves assortment optimization under MNL with fixed prices, the closest benchmark that this paper extends by making prices endogenous and ranking by quality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the MNL choice probabilities under i.i.d. Gumbel utility shocks, grounding the demand model."}],"review_version":1}