{"id":"752c4536-4daa-4982-a08a-c2a2e4febb59","arxiv_id":"2608.08395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Manual search collapses as preference complexity grows, while agentic search with consumer refinement stays viable, creates an adoption lag, and can lead platforms to assign lower AI fidelity to attention-rich consumers.","lead":"This paper models how AI agents change online shopping: it shows manual search stops working when preferences get complex, while agentic search with consumer refinement keeps generating purchases. It matters because it predicts when consumers will be slow to adopt agents and why platforms may give their most attentive users lower-quality AI.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No-collapse result rests on the dense-catalog limit n→∞: with a finite catalog, as complexity grows the best available product recedes to the boundary, so agentic mismatch tends to 1 and revenue to 0 even with refinement.","rationale":"The paper proves its theorems carefully and transparently; the math inside the model is sound. The reader's concern about Assumption 1 is legitimate: if refinement only reduces noise without contracting the search radius, D_k(σ(T)) still approaches 1 as k→∞ for any feasible finite T, so the no-collapse bound would fail. However, an even more load-bearing fragility is the dense-catalog limit of Section 3.3. The theorem's conclusion that expected mismatch stays below 1 relies on the agent being able to resample products from an arbitrarily small ball B_k(R(T)) with infinite density. In any fixed, finite catalog, the realized set of products is drawn once, and as k grows the nearest product in the entire catalog moves to the boundary (distance →1 in probability). Refinement can only select among existing products; it cannot create new ones near the ideal. Thus even granting Assumption 1 and β<η(0), the agent's optimal refinement cannot keep mismatch bounded away from 1: the 'avoidance of collapse' is an artifact of the n→∞ limit. Since the paper's headline interprets the result as a shift from product inspection to preference articulation, the finite-catalog issue strikes at the economic meaning of the central claim. The paper does flag the dense-catalog benchmark, but the abstract and introduction state the no-collapse finding without that qualification. The reader's conditional verdict is appropriate; our concern strengthens the case for adding a finite-catalog robustness analysis or an explicit scaling condition on catalog size with complexity. Verdict remains CONDITIONAL, so no change from the reader's assessment.","tokens_in":43722,"tokens_out":16911,"duration_ms":188420,"concrete_test":"Analytically recompute Theorem 2(c) with a finite catalog: fix N products drawn once from B_k(1), allow refinement to contract attention to B_k(R(T)) but only among the realized products (fallback to the full catalog if empty), and let the consumer choose T optimally with β<η(0). Show that as k→∞, the expected mismatch of the best available product tends to 1, so Π^AI→0. If the limit is 1, the dense-catalog ceiling is essential; if it is <1 for some N, quantify how N must scale with k for no-collapse to survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 2(c)) is proven under the dense-catalog limit n→∞ (Section 3.3). In that limit the agent after refinement draws fresh products from B_k(R(T)), so expected mismatch is R(T)D_k(σ) and stays below 1. For a real finite catalog of N fixed products, the choice set after refinement is the subset of the N realized products inside B_k(R(T)); its expected size is N R(T)^k, which vanishes as k→∞ for any R(T)<1. The closest of N i.i.d. points in B_k(1) has distance →1 in probability as k→∞. Hence even with zero representation noise and optimal refinement, the best available product's distance tends to 1, expected mismatch →1, and platform revenue →0. The no-collapse therefore depends on the catalog being infinitely dense at every preference point, an assumption that the paper labels a benchmark but that is doing the work: without it, agentic search collapses in high complexity just as manual search does, and the 'shift to preference articulation' narrative loses its quantitative basis. This is a different and more fundamental fragility than the functional form of R(T) in Assumption 1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies AI-mediated product search in a k-dimensional spatial model. Consumers choose between manual search, which inspects a limited number of products at perfect fidelity, and agentic search, which screens a dense catalog through noisy representations and lets consumers choose a refinement depth T. Preference complexity k is the number of hard-to-articulate satisfaction dimensions. The main findings are: (i) manual search collapses at a finite complexity threshold, whereas agentic search displays complexity attenuation without collapse (Theorem 2); (ii) platforms and consumers can rank the regimes differently, producing an adoption lag when manual inspection is cheap (Theorem 3); and (iii) conditional on agentic participation, platforms may assign lower fidelity to attention-rich consumers (Propositions 3 and 4). All formal results are proved in Appendix A, with exact closed forms for the manual-search and dense-catalog agentic-search mismatch.","tokens_in":44026,"tokens_out":6525,"duration_ms":79165,"significance":"If the results hold, the paper makes a useful contribution to the emerging economics of agentic commerce. It endogenizes consumer-controlled preference refinement, derives comparative statics from primitive assumptions rather than fitted parameters, and produces testable predictions: refinement depth increases with complexity, agentic revenue remains positive in complex categories, adoption lags arise for cheap manual inspection, and fidelity allocation is inverted by attention budget. The appendix contains complete, self-contained proofs, including exact closed forms in Lemma 2 and Theorem EC.1, which is a strength. The main caveat is that the headline no-collapse result is proven in a dense-catalog limit and depends on a specific radial-contraction assumption, so the scope of the central claim needs to be made precise.","major_comments":[{"comment":"The no-collapse result is an artifact of the dense-catalog limit. The model defines agentic mismatch in the limit n→∞; for any fixed finite catalog of N products, the expected number of products in the contracted ball B_k(R(T)) is N R(T)^k, which vanishes as k→∞ for any R(T)<1, and the closest of N points in the unit ball has distance tending to 1 in probability as k→∞. Consequently, with a finite catalog, expected mismatch under agentic search tends to 1 and platform revenue tends to 0 even with optimal refinement. This directly contradicts the statement in Theorem 2(c) and the abstract that agentic search 'never' returns to the no-search outcome. The paper does label the dense-catalog case a benchmark, but the central narrative in Sections 1 and 8 presents the absence of collapse as a property of agentic search. I ask the authors to either extend the model to finite N (for example, by characterizing the minimal catalog size needed to preserve attenuation) or to explicitly restate Theorems 1 and 2 and the headline claims as dense-catalog limit results with a discussion of the finite-N regime.","section":"§3.3, Theorem 2(c)"},{"comment":"The no-collapse conclusion is tied to the multiplicative radial-contraction form of Assumption 1. By Theorem EC.1, D_k(σ)→1 as k→∞ for any fixed σ>0, so the only force keeping d_AI below 1 in the large-k limit is the radius factor R(T)<1. If refinement instead reduces representation noise σ while keeping R(T)=1, then for any finite refinement depth T, D_k(σ(T))→1 as k→∞, and the attenuation result fails unless σ(T) shrinks sufficiently fast as k grows. The paper acknowledges in footnote 4 that the multiplicative form is a benchmark, but the abstract and conclusion present complexity attenuation as a general qualitative result. I request a robustness analysis of the alternative noise-reduction refinement technology, or at least a precise statement in the theorems of which conclusions rely on radial rescaling versus log-convexity.","section":"§3.4, Theorem EC.1"}],"minor_comments":[{"comment":"The notation for the agentic mismatch switches between d_AI_k(R;σ), d_AI_k(T;σ), and d_AI_k(B,β;k,σ); please define each object once at first use and keep the notation consistent throughout.","section":"§3.3–3.4"},{"comment":"Panel (a) labels the regimes 'A1 A2 A3' but the curve is not otherwise explained; please add a caption line stating which curve corresponds to which regime and how σ_min and σ_max are marked.","section":"Figure 6"},{"comment":"In the proof of Theorem EC.1, the notation ~Z is used before it is formally introduced; please define the limiting random variable explicitly.","section":"Appendix A.1"},{"comment":"The statement that Σ is a compact subset of (0,∞) is fine, but the paper should note explicitly that compactness in (0,∞) means the feasible noise levels are bounded away from zero, so perfect fidelity is never feasible.","section":"§7.1"},{"comment":"Several references are listed as 'arXiv preprint' or 'Working paper' with 2025/2026 dates; please verify that all entries are publicly available and provide stable links or identifiers where applicable.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The finite-catalog critique is the main substantive obstacle. The mathematics under the dense-catalog limit is internally consistent, but the paper's framing overstates the robustness of its central result. If the authors can show that the attenuation result holds approximately when N grows sufficiently fast relative to k, or if they substantially rewrite the claims to make the dense-catalog scope explicit, the paper would be a solid contribution. The self-citations to the authors' own related work are acceptable but should not crowd out independent validation of the agentic-commerce claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nQuick take: this is a careful, well-written theory paper with complete proofs, but the marquee result—agentic search avoids complexity collapse—turns on the dense-catalog limit n→∞. If the catalog is finite, that result does not survive.\n\nWhat's new: the paper makes refinement depth a consumer choice and puts preference elicitation on the consumer side, not the platform. That yields three genuinely new findings: complexity attenuation without collapse, an adoption lag between what platform revenue wants and what consumers voluntarily choose, and an inverted fidelity allocation where attention-rich consumers get lower fidelity. The math is clean. Lemma 2 gives closed-form manual mismatch; Theorem EC.1 gives a truncated-Gaussian representation of D_k(σ). The main theorems have full proofs in the appendix, and the comparative statics are derived from primitives, not fitted to data. Credit where due: this is a solid piece of economic theory.\n\nThe soft spots. First, the no-collapse result depends on the agent being able to draw fresh products from the shrunken ball B_k(R(T)) after each refinement. That is an infinite-catalog assumption. With a finite catalog of N products, refinement only narrows the set of already-existing products; the expected number inside the shrunken ball is N R(T)^k, which vanishes as k grows. Even with zero representation noise, the best product available tends to distance 1 and revenue to 0. So the claimed attenuation is not a property of agentic search per se, but of the dense-catalog limit. The paper labels this a benchmark and is transparent about it, but the headline result is stated without that caveat, and the qualitative narrative is weaker if the effect depends on an infinitely dense catalog.\n\nSecond, Assumption 1's radial rescaling with log-convex R(T) drives the monotone comparative statics. The reader flagged this; I agree it's a real limitation, but it's the kind of reduced form that's standard in a first theory piece. The dense-catalog issue is more structural.\n\nA finite-catalog robustness check would still give complexity attenuation for moderate k if N is large, but it would vanish at some complexity threshold. That seems like an important thing for the authors to state, and it changes the policy implications.\n\nWho this is for: people working on the economics of AI agents and search. It's a legitimate contribution to that emerging literature, and it deserves a serious referee. I'd recommend sending it to review, but with a request that the authors add an explicit discussion of finite-catalog robustness and soften the \"no collapse\" claim.","headline":"Clean, honest theory with a marquee no-collapse result that depends on an infinite-catalog limit and needs a serious caveat before the policy narrative is used.","tokens_in":44486,"tokens_out":4256,"would_cite":true,"duration_ms":45905,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B42","91B26"],"pacs":[],"model":"deepseek-v4-flash","headline":"Manual search stops at a finite complexity cutoff; agentic search with consumer refinement does not, keeping mismatch below the no-search level and platform revenue positive.","keywords":["agentic commerce","agentic search","preference complexity","preference articulation","AI fidelity","consumer search","complexity collapse","adoption lag"],"falsifier":"Observe agentic-search mismatch in categories across the complexity spectrum: under the paper's central claim, when interaction is cheap enough expected mismatch must stay strictly below the random-pick benchmark at every finite complexity level even as consumers refine, so finding any finite complexity at which mismatch reaches the no-search level while consumers are still actively refining would falsify the no-collapse result.","tokens_in":43532,"feed_emoji":"🤖","tokens_out":6574,"duration_ms":70714,"temperature":0.7,"pith_summary":"This paper asks what happens to product search when consumers delegate discovery to AI agents that represent preferences noisily. It claims that manual browsing fails catastrophically as the number of hard-to-articulate satisfaction dimensions grows: at a finite complexity cutoff, consumers stop inspecting, expected mismatch returns to the no-search benchmark, and platform revenue is zero. Agentic search with consumer-chosen refinement depth avoids this finite collapse, provided interaction friction is low enough: complexity makes refinement more valuable, so mismatch stays below the no-search level and revenue stays positive at every finite complexity. The paper also finds an adoption lag, in which agentic search is revenue-superior for the platform before consumers voluntarily choose it, and an inverted fidelity allocation in which attention-rich consumers receive lower AI fidelity because they can compensate with extra refinement. The economic bottleneck shifts from product inspection to preference articulation.","feed_headline":"Agentic search avoids the complexity collapse that ends manual search","feed_subtitle":"As preference complexity rises, manual browsing stops while AI refinement keeps mismatch below no-search and revenue positive.","key_machinery":"The load-bearing mechanism is the assumption that refinement acts as a radial rescaling, so interaction depth shrinks the agent's effective search space from the unit ball to a ball of smaller radius while representation noise scales down proportionally, paired with the linear scale invariance that expected mismatch equals the radius times an intrinsic mismatch term. That intrinsic term is the dense-catalog expected true distance when the agent ranks noisy readings of preferences and products, given by a ratio of truncated-Gaussian integrals and strictly increasing in both complexity and noise. The factorization turns the consumer's problem into a one-dimensional stopping rule: refine until the marginal contraction times the intrinsic mismatch equals interaction friction, or the attention budget binds. This structure is what makes greater complexity raise the value of refinement and prevents a finite collapse.","core_discovery":"On the paper's own terms, the central discovery is a contrast between two search technologies. Under manual search, sampled products are evaluated perfectly but the chance of a close draw falls exponentially with dimensionality, so optimal inspection ends at a finite complexity cutoff; beyond it the consumer searches not at all and outcomes equal no-search. Under agentic search, the consumer chooses interaction depth to shrink the effective search radius, and by linear scale invariance expected mismatch is the radius multiplied by the intrinsic dense-catalog mismatch. Because that intrinsic mismatch rises with complexity, so does the marginal value of refinement; once refinement is worthwhile it stays worthwhile, refinement depth rises weakly with complexity, and in the low-friction case mismatch converges to an interior level below one rather than to the no-search benchmark. Three consequences follow: complexity attenuation without collapse, an adoption lag between the platform's revenue threshold and the consumer's adoption threshold when manual inspection is cheap, and a weakly inverted fidelity allocation across attention segments that becomes strict under intermediate accuracy cost.","pith_inferences":["A testable extension: across real product categories, agentic-search adoption should rise with preference complexity while manual inspection effort should eventually cease, so adoption should concentrate in complex categories with costly per-product evaluation.","The inverted fidelity result suggests an unexplored incentive design: charging attention-rich consumers for higher fidelity could exploit the substitution between consumer effort and platform accuracy, a screening problem the paper leaves open.","If real dialogue mainly reduces representation noise rather than shrinking the product space, the scale-invariance mechanism breaks; measuring whether consideration-set contraction or noise reduction dominates in chat logs would locate where the theory applies.","Hybrid search, in which an agent narrows a shortlist and the consumer inspects it, should combine both advantages; the paper studies pure regimes as benchmarks, so its results bound what hybrid designs build on."],"forward_implications":["In high-complexity categories with costly manual evaluation, agentic search can sustain transactions after manual browsing has ceased, so platform revenue remains positive where manual search would be zero.","Platforms cannot infer voluntary adoption from conversion-based revenue superiority: over the adoption lag, consumers rationally continue manual search even though agentic search would raise platform revenue.","When interaction friction is high, the agentic advantage erodes: consumers never refine, and as complexity grows mismatch approaches the no-search level and revenue approaches zero.","Conditional on agentic participation, attention-rich consumers may receive lower fidelity than attention-constrained ones while remaining weakly more profitable to serve, so accuracy should be invested where consumer effort cannot substitute for it.","Consumer refinement effort weakly increases with complexity but never fully offsets the burden: mismatch weakly rises and revenue weakly falls with complexity even under agentic search."],"supporting_citations":[{"why":"Supplies the noisy high-dimensional spatial search geometry and the dense-catalog truncated-Gaussian formula for intrinsic AI mismatch.","marker":"Liang (2026)"},{"why":"Benchmark conversational recommender in which the platform chooses elicitation intensity, which the paper contrasts by making refinement a consumer choice.","marker":"Kumar et al. (2026)"},{"why":"Benchmark that jointly analyzes solicitation depth and assortment breadth in agentic purchasing, contrasted with consumer-controlled refinement.","marker":"Cao and Hu (2026)"},{"why":"Provides the optimal-stopping benchmark that shapes the manual inspection stopping rule.","marker":"Weitzman (1979)"},{"why":"Quality-versioning baseline whose direction the inverted fidelity allocation reverses.","marker":"Mussa and Rosen (1978)"}],"fun_headline_variants":["Agentic search avoids the complexity collapse that halts manual search","Manual search hits a complexity wall; agentic search keeps going","Agentic commerce: no search collapse, but adoption lag and inverted fidelity","Complexity kills manual search, not agentic: new economics","Agentic search thrives where manual search collapses at high complexity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each unit of interaction contracts the effective search space by a smooth multiplicative factor around the consumer's ideal, with representation noise contracting in proportion; if actual dialogue instead only cleans up noise without narrowing the product set, the complexity-attenuation and fidelity results need not survive.","fun_headline_variants_meta":{"raw":{"variants":["Agentic search avoids the complexity collapse that halts manual search","Manual search hits a complexity wall; agentic search keeps going","Agentic commerce: no search collapse, but adoption lag and inverted fidelity","Complexity kills manual search, not agentic: new economics","Agentic search thrives where manual search collapses at high complexity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000726,"raw_usage":{"total_tokens":3280,"prompt_tokens":997,"completion_tokens":2283,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2196}},"tokens_in":613,"tokens_out":2283,"duration_ms":16799,"temperature":1.0,"reasoning_tokens":2196,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:37:15.482892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Observe agentic-search mismatch in categories across the complexity spectrum: under the paper's central claim, when interaction is cheap enough expected mismatch must stay strictly below the random-pick benchmark at every finite complexity level even as consumers refine, so finding any finite complexity at which mismatch reaches the no-search level while consumers are still actively refining would falsify the no-collapse result.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the optimal-stopping benchmark that shapes the manual inspection stopping rule."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Quality-versioning baseline whose direction the inverted fidelity allocation reverses."}],"review_version":1}