Marketing & artificial intelligence · Evidence review

How does product information affect agentic commerce readiness and performance?

1. Why the information environment belongs in the analysis

An AI system asked to recommend a product receives an objective and some representation of the available products. Neither is necessarily complete. A request for the best deal may leave quality, delivery and package size unspecified. A retrieved listing may supply a headline price but omit the quantity required to compare unit value. The agent's eventual selection is conditional on this information environment.

Agentic commerce means using AI agents to discover, compare, recommend or purchase products for a shopper. This review asks how the availability, retrieval and presentation of product information affect readiness for those tasks and the performance observed within them. Readiness concerns whether a task can be attempted with sufficiently clear information; performance concerns what the configured agent and subsequent customer journey actually achieve. Its focus is the choice process, rather than the mechanics of payments or a forecast of adoption. A recommendation interface can become commercially important before it routinely completes purchases. Conversely, the existence of a checkout facility does not establish that its recommendations satisfy the user's objective.

The object being studied differs from a synthetic respondent used to predict a person. A shopping agent can be evaluated as software carrying out a delegated task. Its behaviour need not resemble human behaviour to be useful. However, whether the selected product is useful requires a criterion grounded in that delegation. Agreement with a benchmark choice, completion of a transaction and improvement in consumer welfare remain different outcomes.

Seller-controlled marketing statements, including best-seller badges, endorsements and scarcity messages, also require definition. A verified product rating contains information. A scarcity statement can be true or false. A visually prominent badge can alter attention without altering the product. An instruction embedded in a page may attempt to redirect the agent's task. These are not interchangeable interventions. Treating every response to seller content as manipulation would conceal legitimate use of product evidence; treating every such response as rational information use would conceal irrelevant influence.

2. Scope, search and source appraisal

Searches on 23 October 2026 combined AI shopping agents, consumer choice, price-equivalent auditing, product information, marketing cues, visual persuasion and agent benchmarks. Original arXiv versions, proceedings records and official product documentation were prioritised. Reference tracing connected recent choice experiments with established shopping environments and security evaluation. Fourteen sources were selected to distinguish mechanisms, outcomes and deployment claims. This is a selective narrative synthesis, without a systematic-review claim.

The 19 October 2026 price-equivalent manuscript [1] was consulted in full-text HTML, including its estimation description, results, confidence-interval table and limitations. The revised 13 October 2026 Tool-Lab manuscript [2] was examined for its experimental design and interpretation of information acquisition. The ACES manuscript [3] provided an additional model-dependent choice comparison. Other studies are used within their accessible abstract or original-record claims. Official Google material establishes documented product functions and feed fields, rather than independent evidence of effectiveness [11] [12].

No agents were run for this review. No purchase logs, analysis code or underlying trial records were independently reanalysed. The source register distinguishes manuscript access from replication. Several recent findings remain preprints, while some older benchmarks use models superseded by current services. The review does not pool success percentages across environments or infer a single susceptibility ranking from unlike tasks. Source versions and their dates are retained so that a result is not silently combined with a later configuration. A merchant-side readiness preprint was added during revision to distinguish observable information sufficiency from validated commercial performance [14]. Its measurement design and limitations were inspected in full-text HTML; its proposed score is not adopted as a validated indicator.

3. What controlled choice experiments establish

ABxLab varies option attributes and persuasive cues in a web shopping environment, reporting substantial changes in agent choice [4]. ACES documents model-dependent concentration, position effects and reactions to descriptions and endorsements [3]. These findings challenge an assumption of a neutral, stable machine chooser. They do not establish the fraction of today's shoppers who delegate purchasing, nor the market revenue attributable to the tested effects.

Controlled interventions have an advantage over observing a seller's commercial results: the researcher can alter a defined input while holding other features constant. This improves interpretation of the difference within the experiment. The scope remains the tested configuration and environment. A controlled model experiment identifies behaviour of the model under those interventions; it does not randomise human adoption of a shopping assistant.

The unit of observation is particularly important. Thousands of repeated choices can reveal a stable response probability for a fixed prompt and product pair. They do not create thousands of independent product markets. Product diversity, task diversity and variation across model versions determine different forms of generalisation. Treating all calls as interchangeable observations risks overstating how broadly the estimated relationship is known.

The comparison also needs a meaningful alternative. A product with better specifications may be selected because of those specifications. A badge may introduce credible information or merely rephrase a fact already supplied. A manipulation intended to isolate presentation therefore needs to preserve the underlying information. Experiments that change several features at once can evaluate a package of changes, but cannot uniquely identify which feature caused the difference.

4. Price-equivalent auditing and its interpretation

Buecheler, Berbeglia and Guo introduce a parity threshold: the premium at which a cued product is selected half the time in a controlled comparison. Their 19 October preprint uses 113,472 trials across seven models and two product categories. For GPT-4o and headphones, strong social proof has a reported price-equivalent of 12.9%, with a pair-clustered 95% bootstrap interval of 10.9% to 17.1% [1]. The aggregate set retains seven of eight audited language models after a non-response exclusion; cross-model claims further exclude a strongly position-driven model. Each category contains six product pairs. The primary estimator interpolates the parity crossing over eight price gaps, with 24 repetitions per pair and price point, balancing cue target and display order. The six-pair bootstrap can understate uncertainty with so few clusters [1]. This is a simulated choice threshold, not an estimate of human willingness to pay.

The measurement distinction can be explained without assuming that the measured premium is commercially recoverable. At equal price, two cues can both lead to frequent selection. That one observation does not show whether either preference persists when price changes. Varying price supplies a second dimension. A ranking based on equal-price frequency can therefore differ from a ranking based on a parity threshold without either calculation being arithmetically wrong.

Such a threshold is conditional on the agent's price response. A system that scarcely responds to price can appear to tolerate a large premium even when its decision procedure is unsuitable for price comparison. Failure to observe a crossover within the tested range is a limit on identification, rather than evidence of an exact value beyond that range. It is therefore inappropriate to turn censored thresholds into a precise league table.

There is also a distinction between an outcome expressed in money units and an economic welfare estimate. A model's selection can be made less favourable to the user even if the measured cue premium is large. Whether a seller could realise a higher price depends on actual delegation, competing offers, user confirmation and subsequent behaviour. None follows simply because the experimental measurement uses a price scale.

5. Information acquisition can alter what looks like preference

Wadi and Ma's Tool-Lab study makes product attributes accessible through tool calls and varies the cost of obtaining them. The October revision examines eight models and contrasts vague with specific shopping objectives. Under costly retrieval and vague goals, omissions of attributes required for unit-price comparison contribute to inferior selections; performance differs when the relevant objective is explicit [2]. These experimental acquisition costs are not observations of normal retail API charges.

Study 1 Tool-Lab contains 3,200 sessions across eight models, two cost conditions and two goal conditions, with 100 sessions per cell. Under a vague goal, pooled optimal selection falls from 0.89 at zero cost to 0.29 under the imposed charge; under the specific goal, it falls from 0.94 to 0.83. These descriptive proportions concern a fixed coffee comparison and the authors’ unit-value criterion [2]. They are not population purchase rates. The associated mediation analysis is model-based evidence about a pathway; it does not make the intermediate retrieval count a separately randomised treatment.

The interpretive point is that an apparent preference can depend on which facts reached the chooser. If package quantity is missing, selecting the lower headline price need not reveal a stable preference for a smaller pack. It may reveal failure to retrieve the denominator needed for the delegated comparison. The same final choice can arise from incomplete information, a different objective or an arithmetic error. Observing the chosen product alone cannot distinguish them.

Process observations help by recording what was acquired before commitment. They still leave questions about explanation. A generated rationale saying that price mattered is a report by the model, whereas a recorded retrieval identifies an observable action. Neither automatically establishes every internal calculation. The stronger interpretation concerns the information available at the point of choice and the controlled changes to that information.

This matters for transfer from laboratory to storefront. A natural website imposes time, navigation, rendering and extraction costs that differ from explicit experimental charges. Testing one representation cannot establish that all such frictions have the same effect. The study provides a mechanism to investigate, with specific experimental support; its numerical costs cannot be inserted directly into a forecast of how current consumers will shop.

6. Visual presentation and the interface

Cherep and colleagues investigate vision-language models using controlled choices between edited images. Their accepted ICML 2026 manuscript reports that visual modifications can shift head-to-head selection probabilities [5]. The result concerns model choice under the study's comparison. It does not establish that the same edit attracts people, improves the physical product or increases sales.

VisualWebArena evaluates agents on tasks requiring image-text interpretation and website actions [8]. WebShop provides a simulated shopping environment grounded in real product information and written requirements [7]. Both clarify why successful operation is demanding: the system must interpret an instruction, locate relevant information and act within an interface. Their task-completion results address competence under the benchmark's criteria, rather than immunity to seller influence.

A useful distinction is between availability and use. Information can exist in a product feed but fail to appear in a rendered page. It can appear visually but be absent from a text extraction. It can be extracted and then ignored when the agent ranks alternatives. These are separate points of failure. A high-quality catalogue does not by itself prove that every assistant receives or uses the relevant fields.

Presentation effects also complicate comparisons between providers. Different interfaces may supply different subsets or orders of products before the language model acts. An observed difference between assistants can consequently reflect retrieval, presentation, the model or the policy that combines them. Without controlling those layers, describing the difference as a model preference attributes more than the evidence identifies.

7. An agent chooser is not a human demand forecast

Lu and colleagues compare generated action sequences with records from 31,865 real shopping sessions. Their April 2026 revision reports substantial limitations in prompt-only simulation and improvements from training on relevant behaviour [6]. The study concerns prediction of human actions, a different target from asking software to satisfy an explicit shopping instruction. It prevents an easy substitution between the two claims.

A recommendation experiment may estimate which product an assistant selects. A consumer forecast asks how people behave, including whether they use the assistant, trust it, inspect alternatives or abandon the purchase. If the system offers a recommendation rather than executing a binding instruction, the user's response is part of the commercial process. A forced-choice simulation omits that response unless it is explicitly modelled and validated.

The distinction becomes clearer with denominators. Selection share among presented products differs from selection share among the entire catalogue. Purchase share among assistant users differs from sales among all customers. Revenue attributed to assistant referrals differs from revenue that would not have occurred through another route. None of these quantities can be inferred from another without additional observations or assumptions.

Model updates create a further forecasting limit. Evidence that choices depend on a model version means that a historical distribution of simulated selections is not a stable demand curve by default [3]. Repeating a test can detect drift in the configured system. It does not establish whether the drift improves consumer benefit. That judgement requires a reference objective and evidence about the outcomes that follow.

8. Accurate information, persuasion and instruction attacks

GEO research studies methods for changing visibility in generated answers [10]. Its outcome is a representation of source visibility in a generative engine. A source being cited more often is distinct from its product being chosen, and both are distinct from a completed purchase. Applying the same label to all three can obscure which part of a commercial journey has actually changed.

AgentDojo studies prompt injection in environments where agents use tools over untrusted material [9]. It establishes a separate class of problem: external content can attempt to redirect the agent from its authorised task. Ordinary product claims and malicious instructions require different analysis even if both influence an output. A statement about a product's battery life supplies candidate evidence; an instruction to ignore the user's budget attempts to change the task.

Truthfulness is another independent dimension. A credible-looking endorsement may affect selection while remaining unverified. An experiment can stipulate a statement as true to isolate response to that statement. Deployment needs evidence that the statement actually holds. A finding about responsiveness cannot serve as substantiation of a product claim or permission to invent an endorsement.

The resulting evaluation is layered. One question asks whether the information is accurate. Another asks whether the system treats it as data rather than authority to change its objective. A third asks whether the final choice satisfies the shopper. A system can pass one and fail another. Combining them into a single trust score makes the failure difficult to locate and the claimed improvement difficult to interpret.

9. What the evidence supports for product information

Google's product-data specification documents fields such as price, availability and product identifiers [11]. It establishes what a particular merchant interface requests, not a general causal effect of every field on assistant selection. Its operational relevance is that structured, consistent records make the intended information explicit. Their contribution to any particular agent still depends on retrieval and use.

An illustrative unit-price comparison shows why the content of a record matters. A £9 pack containing 300 grams costs £3 per 100 grams. A £12 pack containing 500 grams costs £2.40 per 100 grams. These are invented arithmetic examples, not observed offers or experimental results. Headline price and unit price rank the two differently. Which comparison is relevant depends on the shopper's quantity, budget and other constraints.

For a marketer, factual completeness and persuasive strength are therefore related but distinct work. Accurate quantity, compatibility and delivery information help specify the offer. A campaign headline may attract attention while omitting details needed for an automated comparison. The selected literature warrants examining that interface, rather than assuming that a stronger slogan solves a retrieval problem.

The trade-off is that maximal detail is not automatically maximal usefulness. Contradictory fields, stale records and repeated unqualified claims can make the choice harder to audit. The research reviewed here does not identify a universally optimal amount of product text. A useful evaluation tests whether the relevant assistant obtains the facts needed for a defined decision and whether those facts remain consistent across the representations it encounters.

10. Commercial interpretation and counterarguments

The strongest practical case for acting is that the information environment is observable and testable. A business can compare its public product records across relevant interfaces, record what an assistant retrieves and inspect whether the recommendation respects explicit constraints. This is a narrower undertaking than attempting to reverse-engineer a universal machine-consumer psychology.

The strongest counterargument is that a simulated audit may have limited relevance to actual customers. If an audience rarely delegates the relevant task, an elaborate agent experiment can displace more consequential work. Even among assistant users, the dominant commercial issue may be availability, brand preference or delivery rather than sensitivity to a badge. The selected research supplies no general budget-allocation rule across these priorities.

A second counterargument concerns useful persuasion. A system that responds to a verified, relevant endorsement may be incorporating evidence the user would value. The empirical question is whether the response is proportionate to that information and the delegated objective. A finding of sensitivity alone does not establish harm. Conversely, a neutral-sounding response does not establish that the choice was well founded.

These limits favour proportionate testing. A low-cost inspection may reveal an omitted quantity or conflicting price. A broader experiment is justified only when the decision requires it and the benchmark is relevant. Economic claims then need the cost of the intervention and outcomes after deployment, with a credible comparison. A laboratory premium is useful evidence about a configured choice process; commercial return remains a further question.

11. Agentic commerce readiness and performance indicators

A readiness assessment needs an explicit shopping task, a defined product set and an observable agent configuration. An accessible product page can still omit a fact required by that task. Complete information can still be misunderstood. A technically completed order can still supply the wrong product. An audit that records only the presence of structured fields cannot establish the later outcomes.

A recent merchant-side preprint separates access, whether the published terms settle a buyer's requirements, and the evidence retained about those terms [14]. That separation is useful, but its proposed weighted score remains unvalidated. The authors report a single-rater design for the full assessments, substantial measurement limitations in the larger automated exercise and no calibration against transaction outcomes. It is methodological evidence about what can be recorded, rather than proof that a high readiness score predicts sales.

The evidence reviewed here supports keeping diagnostic measures separate from commercial outcomes. For a predetermined set of relevant comparisons, an audit can record whether the required facts were available, whether the agent retrieved them and whether its answer respected the stated constraints. Each rate needs its denominator, exclusions and unresolved cases. A fact that could not be inspected is unknown; it is not evidence that the merchant omitted it. A test set containing only easy cases cannot establish performance on the wider catalogue.

Commercial performance requires further observations: successful authorised handoffs, completed purchases, fulfilment, returns and relevant costs. Referral attribution alone does not establish incrementality. A readiness improvement can be useful before there is enough traffic to estimate a sales effect, provided the claim remains about the corrected information or tested task. No universal weighting, threshold or financial forecast follows from the selected studies. Authentication, payment authorisation and the full security architecture remain outside this review's detailed evidential scope.

12. Limitations and conclusion

Recent experimental evidence shows that agent choices can change with seller content and information access. Its scope remains restricted by product sets, prompts, models and interfaces. The newest price-equivalent findings are provisional, and many repeated trials share a small set of underlying product pairs. Precision within a test does not establish broad market representativeness. No reviewed source supplies a single current susceptibility rate for all shopping assistants.

Official product announcements establish functions and availability claims, rather than independent measures of consumer benefit. Google's 15 October 2026 Universal Cart announcement describes a US launch and shopping functions [12]. It makes the choice-quality question timely, but does not validate any of the laboratory effects for that product. Model Cards provide a discipline for recording intended use and evaluation boundaries [13]; documentation itself remains distinct from a successful evaluation.

The synthesis supports treating AI shopping as a configured decision process. What the agent is asked to achieve, which facts it obtains, how alternatives are presented and how its choices are measured all affect interpretation. Selection probability, a price-equivalent threshold, transaction completion and realised benefit answer different questions. Marketing claims become defensible when they specify which of those outcomes was observed and retain the conditions under which it was measured. The evidence does not establish that machine-mediated commerce removes the need for accurate product information, nor that a simulated preference can be converted directly into a pricing strategy.

References

1. Buecheler, F., Berbeglia, G., & Guo, M. (2026). A Cue’s Hidden Price: Auditing Seller Influence on AI Shoppers. arXiv:2610.25843v1, 19 October. Preprint. Original source ↗Material consulted: Full-text HTML: definition, controls, results, Appendix C and limitations.
2. Wadi, D., & Ma, Y. (2026). Shopping by Algorithm: How Agentic AI Deploys Human Heuristics as a Surrogate Consumer. arXiv:2609.28372v2, 13 October. Preprint. Original source ↗Material consulted: Full-text HTML: experimental design and interpretation of acquisition costs.
3. Allouah, A., Besbes, O., Figueroa, J. D., Kanoria, Y., & Kumar, A. (2025). What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce. arXiv:2508.02630v3, 17 December. Preprint. Original source ↗Material consulted: Author abstract and version record; manuscript comparison description.
4. Cherep, M., Ma, C., Xu, A., Shaked, M., Maes, P., & Singh, N. (2026). A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments. ICLR. Author manuscript arXiv:2509.25609v2, 24 February. Original source ↗Material consulted: Original abstract and version/conference record.
5. Cherep, M., Pranav M R, Maes, P., & Singh, N. (2026). Visual Persuasion: What Influences Decisions of Vision-Language Models? Accepted to ICML 2026. arXiv:2602.15278v2, 31 May. Original source ↗Material consulted: Original abstract and version/conference record.
6. Lu, Y., et al. (2026 revision). Can LLM Agents Simulate Multi-Turn Human Behavior? Evidence from Real Online Customer Behavior Data. arXiv:2503.20749v8, 28 April. Preprint. Original source ↗Material consulted: Original abstract and version history.
7. Yao, S., Chen, H., Yang, J., & Narasimhan, K. (2022). WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents. NeurIPS. Corrected author manuscript arXiv:2207.01206v4, 8 February 2023. Original source ↗Material consulted: Original abstract and corrected-version record.
8. Koh, J. Y., et al. (2024). VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks. ACL. arXiv:2401.13649v2, 6 June. Original source ↗Material consulted: Original abstract and conference/version record.
9. Debenedetti, E., Zhang, J., Balunović, M., Beurer-Kellner, L., Fischer, M., & Tramèr, F. (2024). AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. arXiv:2406.13352v3, 24 November. Original source ↗Material consulted: Original abstract and version record.
10. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD, 5–16. Author manuscript arXiv:2311.09735v3, 28 June. Original source ↗Material consulted: Original abstract and conference/version record.
11. Google Merchant Center. Product Data Specification. Live official documentation, accessed 23 October 2026. Original source ↗Material consulted: Official product specification.
12. Ganapathy, S. (2026). An Intelligent New Shopping Cart. Google, 15 October. Original source ↗Material consulted: Official announcement and dated launch description.
13. Mitchell, M., et al. (2019). Model Cards for Model Reporting. FAT*. DOI: 10.1145/3287560.3287596. Author manuscript arXiv:1810.03993v2, 14 January. Original source ↗Material consulted: Author abstract and version record.
14. de Valois-Franklin, A., Bogdan, A., Ibrahim, A., Shaer, I., & Andguladze, T. (2026). Get In, Decide, Prove: A Merchant-Side Readiness Benchmark for Agentic Commerce. arXiv:2610.17378v1, 12 October. Preprint. Original source ↗Material consulted: Full-text HTML: assessment unit, metrics, missing-evidence rules and limitations (sections 3 and 6).