When AI Sets the Expectation

An AI recommendation can shape a customer's expectations before the first visit. What happens when those expectations meet the company's own information?

By Stefano Zoia · · 6 min

What the evidence shows

AI can influence purchase decisions before a customer reaches a retailer. Separate research documents pricing and inventory inconsistencies across channels, but does not establish how often these affect AI answers.

What remains unknown

None of the studies reviewed here directly measures the mismatch between an assistant’s claims and a retailer’s verified information, or its effect on sales.

What businesses can do

Audit a small set of consequential claims — prices, availability, delivery and returns — against authoritative information, recording errors, missing conditions and changes over time.

Imagine a customer arriving on a retailer’s website after an AI assistant has recommended a product and described its returns as free. The product is right. The price is right. But free returns apply only to loyalty members.

The customer now has to resolve a contradiction before deciding whether to buy.

This is a hypothetical example, but it captures a practical problem. A product page may be read as confirmation or contradiction of a claim the company never made.

The company controls what it publishes. It has less control over how an assistant selects, combines and qualifies that information. Yet the customer experiences both as parts of the same purchase decision.

Being recommended is valuable. What the recommendation leads someone to expect deserves attention too.

A recommendation does not settle the decision

A Contentsquare survey of 2,000 consumers in France and the United States, fielded by Pollfish, found that 47% said AI had changed a final purchase decision, while 57% reported completing a purchase guided by it. [1]

These are self-reported experiences in two markets, not a measure of AI’s share of all purchases. They suggest that, for some customers, an assistant already helps establish the terms on which a retailer will be assessed.

The same survey asked what respondents would do if a website disappointed them after an AI recommendation. Only 3% said they would complete payment; 57% would continue researching, 23% would turn to a competitor, and 16% would abandon.

Those answers describe hypothetical intentions. “Disappointment” is also broader than a contradiction: a slow site, confusing navigation or unexpected delivery charge could produce the same response.

The useful observation is narrower. Most respondents said they would continue evaluating their options, either through further research or by turning to a competitor. The recommendation had not closed the decision.

Observed behaviour provides another reason to examine this part of the journey. Adobe reported that, in March 2026, visits from AI sources converted 42% better than non-AI traffic across the US retail sites it analysed. A year earlier, they had converted 38% worse. [2]

That reversal cautions against treating either figure as a permanent property of the channel. The higher conversion rate is consistent with visitors arriving further along in their decision, although it does not establish why they convert better or how they respond to contradictory information.

It does suggest that this is traffic worth understanding beyond the referral count.

Agreement is not enough

An assistant’s answer and a company’s page can diverge for several reasons.

The assistant may be wrong. The page may be out of date. Both may be correct for different countries, product variants or customer groups. A price may have changed between the recommendation and the visit.

Two sources can also agree and both be wrong.

The objective cannot simply be to make the website match the answer. Both need to be checked against information the business can stand behind, with the relevant conditions attached.

A returns policy is more than “free” or “paid”. It may depend on membership, location, product category and time since delivery. Availability needs a location and a timestamp. A certification needs a scope and a verifiable source.

These details can change a purchase decision. They are also easy to lose when a complex policy becomes a short, confident answer.

The research measures adjacent problems

None of the studies reviewed here directly measures how often an AI assistant’s description of a retailer conflicts with verified information, then follows the effect on purchasing. The scale and commercial impact of that gap remain open questions.

There is evidence of inconsistencies within businesses themselves. Salesforce’s fourth State of Commerce study surveyed 3,450 commerce professionals across 20 countries. Among multichannel organisations, 41% reported inconsistent pricing and promotions across channels, and 40% reported inventory that was not synchronised in real time. [3]

These findings do not establish that an assistant will give an incorrect answer. They show that some businesses already struggle to maintain consistent information across the systems from which an answer might be assembled.

Salesforce sells products addressing these problems. That commercial interest is relevant context for interpreting the research, rather than a reason to dismiss or privilege it.

A separate Empirank audit examined 1,257 verifiable factual claims about 182 businesses recommended by a search-enabled OpenAI configuration. It found at least one contradicted claim for 11% of the businesses; 38% of claims could neither be confirmed nor refuted. [4]

The audit concerns local businesses and one configuration, not retail product recommendations generally. Its useful contribution here is the distinction between a claim that is contradicted and one that cannot be verified.

An unsupported claim is not necessarily false. But if a customer cannot check it, it leaves the business with a different problem from a demonstrably incorrect price or policy.

These studies use different methods, populations and settings. Together, they provide reasons to investigate. They do not supply a single estimate of the risk.

Start with a small set of consequential claims

A company can investigate its own exposure without waiting for a comprehensive market study.

Start with questions customers ask before purchasing. Choose a manageable set where a misleading answer could materially change the decision: total price, availability, delivery, returns, eligibility or certification.

For each answer, keep the prompt, assistant, date, market and cited sources. Repeat the questions over time. One response is an observation, not a reliable account of what every customer sees.

Compare the answer with an agreed authoritative source inside the business, then with the public page a customer would encounter. If those two sources already disagree, that is a finding in its own right.

Record separately:

The distinction helps determine what to fix. An outdated page needs updating. Conflicting channel data needs reconciliation. An omitted restriction may need clearer public wording. An unverifiable certification claim needs supporting evidence or correction.

A model error may persist even when the company’s information is accurate and accessible. Improving the source does not guarantee control over the answer.

The audit should therefore record both the discrepancy and what the business can do about it. Repeating it after a correction can show whether the sampled answers change.

This would measure the accuracy and completeness of answers under recorded conditions. It would not, on its own, reveal how many customers saw them or how many sales were affected. Those questions require further evidence.

Make the expectation checkable

Visibility tells a business whether it appeared in a sampled answer. It does not tell the business whether that answer created an expectation it can fulfil.

A useful next question is specific: when an assistant describes our prices, policies or capabilities, which claims can we verify, which conditions disappear, and where does the customer encounter a different account?

The evidence reviewed here does not establish the scale of that gap. It offers enough reason to start checking it.

For a customer told that returns are free, the distinction becomes practical at the moment they discover the conditions. That is where the recommendation meets the business — and where the business needs a clear, accurate answer of its own.

— Stefano

Sources

  1. Jérôme Pouponnot, “Parcours d’achat : comment l’IA bouscule les décisions des clients”, RelationClientMag.fr, 4 September 2026, reporting a Contentsquare study fielded by Pollfish.
  2. Vivek Pandya, “U.S. retailers see surge in AI traffic, but many websites are not entirely readable by machines”, Adobe Digital Insights, 16 April 2026.
  3. Salesforce, “Shopping’s New First Step: Agentic Search Grows 200%”, State of Commerce, fourth edition, 28 July 2026.
  4. Luís Rijo, “1 in 9 businesses OpenAI recommended carried a wrong claim, audit finds”, PPC Land, 24 August 2026, reporting Empirank audit AIV-009.