Independent Researcher, Nairobi, Kenya

teresiadigital@gmail.com

Abstract

This paper examines whether and how artificial intelligence (AI) is changing the way consumers discover and evaluate businesses. Rather than treating this as a binary question, the paper argues for a more precise claim supported by current evidence: AI is changing the interface and sequence through which consumers encounter, compare, and interpret businesses, without yet demonstrating that it consistently replaces independent research, peer recommendation, or direct brand interaction. Drawing on behavioural data from Pew Research Center, a pre-registered experiment on trust in generative search, and a range of industry consumer surveys, the paper proposes a three-stage framework — discovery, evaluation, and action — for separating visibility, persuasion, and conversion, which are frequently conflated in public discussion of this topic. The review finds that evidence for AI-mediated discovery is strongest for click behaviour on U.S. Google searches and weakest for causal effects on purchase decisions, business fairness, and long-term market concentration. It further identifies a substantial geographic gap: existing evidence is concentrated in the United States and Western Europe, leaving open how AI-mediated discovery operates in markets such as Kenya, where local and service businesses depend on different forms of documentation and reputation. The paper closes with eight original research questions and a four-phase mixed-method design: evidence review, prompt audit, controlled experiment, and qualitative interviews intended to generate the first systematic account of AI-mediated business discovery in an East African market.

Keywords: artificial intelligence; generative search; consumer behaviour; business discovery; source credibility; digital marketing; emerging markets; Kenya

1. Introduction

Generative artificial intelligence has become embedded in how many consumers search for information, compare products, and form judgments about businesses. Search engines now generate synthesized answers rather than only lists of links, and a growing share of consumers report using conversational AI tools for shopping and service research (KPMG LLP, 2024; Bazaarvoice, 2025). This shift has prompted a wave of commentary asserting that the conventional marketing funnel has been disrupted or replaced by AI-driven discovery.

This paper takes a more cautious position. It argues that the strongest available evidence supports a narrower and more defensible claim: AI is altering the interface and sequence through which consumers encounter, compare, and interpret businesses, but it has not yet been shown to consistently replace human judgement, independent research, peer review, or direct interaction with a business. The distinction matters because AI may reduce the number of observable clicks in a consumer journey without necessarily reducing — and possibly increasing — its influence over the final decision.

The paper has four objectives. First, it synthesises the most credible available evidence on AI-mediated discovery and evaluation, distinguishing well-supported findings from claims that rest on thinner or more heterogeneous survey evidence. Second, it proposes a conceptual framework — discovery, evaluation, and action — that prevents a common analytical error in this literature: treating visibility, persuasion, and conversion as equivalent outcomes. Third, it identifies specific gaps in the evidence base, with particular attention to the near-total absence of research on African and, specifically, Kenyan consumer markets. Fourth, it proposes a concrete, mixed-method research agenda capable of producing original empirical evidence rather than restating existing commentary.

2. Conceptual Framework: Separating Discovery, Evaluation, and Action

Traditional accounts of consumer decision-making describe a sequential funnel in which a need leads to search, comparison of alternatives, and eventual purchase (Court et al., 2009). Generative AI does not eliminate this sequence so much as compress and relocate parts of it. A consumer may now describe a need to an AI system and receive, in a single interaction, a shortlist, a comparison, and a recommendation that previously required multiple searches and site visits.

To analyse this shift without conflating distinct outcomes, this paper separates the consumer journey into three stages, summarised in Table 1.

Stage	What is at stake	Illustrative questions
Discovery	Whether the business enters the consumer’s field of consideration at all	Was the business named or shortlisted? Was inclusion driven by the business’s own content, by third-party content, or by platform data?
Evaluation	How the consumer judges the business once it is visible	Which attributes did the AI system emphasise? Was the business compared with competitors? Did citations increase perceived credibility? Did the consumer verify the claims independently?
Action	What the consumer does as a result	Did the consumer click through, search separately, contact the business, visit a physical location, purchase, delay the decision, or choose an alternative?

This framework is useful because most existing evidence speaks to only one of these stages. Behavioural click data speaks mainly to discovery and the early part of evaluation; consumer-trust surveys speak mainly to evaluation; almost no publicly available evidence speaks directly and causally to action. Treating a reduction in clicks as evidence of reduced business value, or treating stated trust in AI as evidence of increased purchasing, risks overstating what the data can support.

3. Evidence on Changing Search and Discovery Behaviour

3.1 The growth of answer-oriented search

Google extended its AI Overviews feature to more than 100 countries and territories in October 2024, a rollout the company said would bring the feature to more than one billion users each month (Google, 2024). This expansion is significant because it means many consumers now routinely encounter a synthesised, AI-generated answer positioned above traditional search results, rather than proceeding directly to a list of links.

3.2 Behavioural evidence on clicking behaviour

The most rigorous behavioural evidence currently available comes from a Pew Research Center analysis of browsing data from 900 U.S. adults, covering 68,879 Google searches conducted in March 2025, of which 12,593 (about 18%) produced an AI-generated summary (Chapekis & Lieb, 2025). The study reported that when an AI summary appeared, users clicked a traditional search result in only 8% of visits, compared with 15% of visits when no summary appeared; clicks on a link cited within the AI summary itself occurred in only about 1% of visits (Chapekis & Lieb, 2025).

This finding should be interpreted carefully. It demonstrates an association between AI summaries and fewer outbound clicks; it does not, by itself, demonstrate that consumers made worse decisions, that businesses received no value from being mentioned in a summary, or that a business’s overall discovery outcomes worsened. A consumer may see a business named in a summary, remember it, search for it separately later, or visit a physical location without generating a measurable referral click. The evidence is therefore best read as showing a shift from click-based discovery toward impression-, mention-, and shortlist-based discovery, rather than a straightforward decline in the value of appearing in search results.

Two additional limitations of this evidence should be noted for a reader assessing its external validity. First, the sample is limited to U.S. adults using Google on tracked devices, which restricts generalisability to other markets, search engines, and devices. Second, the study measures clicking behaviour rather than downstream outcomes such as purchase, contact, or satisfaction, so it cannot on its own establish effects on decision quality.

3.3 Survey evidence on AI use in product and business evaluation

Beyond search-engine behaviour, several industry surveys report growing consumer use of generative AI for shopping and business research. Because these surveys differ in sample composition, wording, and date, their point estimates should be treated as broadly indicative rather than precisely comparable. Table 2 summarises the main figures referenced in this review.

Source	Finding	Sample / date
KPMG LLP (2024)	54% of respondents said they trusted generative AI for personalised recommendations	1,000 U.S. college-educated, self-identified informed consumers, 2024
Bain & Company and Dynata, as reported by eMarketer	51% of U.S. consumers said they trusted generative AI for shopping and product recommendations	U.S. consumers, December 2024
Bazaarvoice (2025)	24% of consumers had used a generative AI tool to search for a product instead of a search engine; the figure rose to 41% among respondents aged 18–34	Over 7,000 consumers across multiple countries, 2025 Shopper Experience Index
Capgemini Research Institute (2025)	68% of surveyed consumers said they were prepared to act on generative-AI shopping recommendations, up from 63% the previous year; the figure was higher in some national markets (e.g., 74% in the UK)	12,000 global consumers, published January 2025

These figures indicate that generative AI has moved beyond novelty use for a meaningful share of consumers in the markets surveyed. However, “using AI” spans a wide range of behaviours — from asking a chatbot for a general idea, to relying on a search summary, to comparing options through a dedicated shopping tool — and these should not be treated as equivalent. None of the cited surveys establishes a causal link between AI use and a specific purchase, since consumers who consult AI may already have had strong purchase intent (a point discussed further in Section 7).

4. Trust as a Mediating Mechanism

A central and often overlooked finding in this literature is that consumers can rely on AI-generated answers even when they do not fully trust them. Li and Aral (2025), in a pre-registered experiment combining approximately 12,000 search queries across seven countries with a randomised, U.S.-representative behavioural study, found that participants trusted generative search less than traditional search on average. However, the presence of citations and reference links increased trust in the AI-generated answer even when some of those citations were incorrect or fabricated, and greater trust was in turn associated with more clicking and less time spent independently evaluating results.

This has a direct implication for business discovery. An AI system does not merely retrieve information about a business; it frames how that information is interpreted, for example by describing a business as “best for beginners,” “a premium alternative,” or “not recommended for this use case.” Because such framing can shape consumer judgement even when the underlying source is never visited, the central question for a business is not only whether it is mentioned by an AI system, but how it is positioned relative to named or implied alternatives.

5. Competing Interpretations of AI’s Role in Discovery

The evidence reviewed above is compatible with more than one interpretation of how consequential AI-mediated discovery actually is. Four interpretations recur in the literature and in industry commentary; they are not mutually exclusive, and the balance among them likely varies by product category, market, and consumer segment.

5.1 AI as a new discovery layer

Under this interpretation, the operative question for a business shifts from “can consumers find my website” to “when a consumer describes a need to an AI system, does my business enter its shortlist, and how is it described.” This view is potentially favourable to smaller or less-established businesses, but only to the extent that AI systems have access to reliable, current, and sufficiently detailed information about them.

5.2 AI as a convenience layer

A more cautious interpretation holds that consumers use AI mainly for early-stage research while continuing to rely on established signals — reviews, personal recommendations, price, direct interaction, and credentials — before committing to a decision. Under this view, AI changes the information-gathering stage without eliminating the underlying economics of trust; it may produce a shortlist, but the business still has to prove its value directly.

5.3 AI as a concentrator of attention

A third interpretation cautions that AI-generated recommendations may increase efficiency without increasing diversity of consideration. If systems disproportionately draw on a small number of highly visible sources, they may reproduce existing prominence rather than surface less-established but high-quality businesses. Consistent with this concern, Chapekis and Lieb (2025) found that Wikipedia, YouTube, and Reddit were among the most frequently cited sources in both AI summaries and standard search results, which raises the possibility of a feedback loop in which existing visibility increases the likelihood of AI inclusion, which in turn increases perceived legitimacy and further visibility.

5.4 AI as a form of persuasion

A fourth interpretation treats AI-generated comparisons as a persuasive intermediary rather than a neutral information source. Because such comparisons are typically phrased conversationally rather than as advertising, consumers may not experience them as promotional even though they still shape attention, perceived differences, and exclusion from consideration. This raises unresolved questions of disclosure and commercial influence that are addressed only partially in the current literature.

6. What the Current Evidence Does and Does Not Establish

Table 3 summarises the confidence that can reasonably be placed in several claims commonly made about AI and consumer discovery, based on the evidence reviewed in Sections 3 and 4.

Claim	Current evidentiary support	Confidence
AI-generated summaries reduce outbound search clicks	Supported by Pew’s 2025 U.S. browsing-data study (Chapekis & Lieb, 2025)	Moderate–high for U.S. Google users in that sample
Consumers are increasingly using AI for product and business research	Supported by multiple independent surveys (KPMG, 2024; Bazaarvoice, 2025; Capgemini, 2025)	Moderate; samples and definitions of “use” vary
AI influences which businesses consumers consider	Plausible and consistent with recommendation-use and trust evidence	Moderate; direct causal evidence is limited
AI increases conversion or revenue for recommended businesses	Largely asserted in commercial/industry surveys; independent causal evidence is scarce	Low–moderate
AI improves consumer decision quality	Not established; convenience may improve while errors and unwarranted confidence remain risks	Low
AI reduces the importance of brand, reviews, and personal recommendation	Not established	Low
AI creates fairer discovery conditions for smaller businesses	Possible in principle; not demonstrated consistently	Uncertain

The recurring limitation across this evidence base is that most available data measures self-reported use or observed clicks rather than the full causal chain from AI exposure to business outcome.

7. Gaps in the Existing Evidence

7.1 Geographic concentration of research

The strongest behavioural study identified in this review is limited to U.S. Google users (Chapekis & Lieb, 2025), and the survey evidence cited in Section 3.3 is drawn predominantly from the United States, the United Kingdom, and other high-income markets. This leaves substantial gaps for markets such as Kenya and other African economies, including basic descriptive questions: how frequently consumers use AI for local business discovery; which tools they use; whether AI recommendations are made in English, Kiswahili, Sheng, or a mix of languages; how AI-mediated discovery interacts with WhatsApp, TikTok, Google Maps, and Facebook, which play a comparatively larger role in local commerce in this region; and whether consumers trust AI more or less than referrals from people they know.

7.2 Local and service businesses

Much existing evidence concerns product categories with structured, standardised data, such as electronics, fashion, and travel. Local and service businesses — restaurants, clinics, schools, contractors, and financial services — depend on harder-to-standardise signals such as location, hours, licensing, informal reputation, and photographic evidence. It is not yet established whether AI systems perform well or poorly when this kind of information is incomplete, outdated, or contradictory, which creates an important distinction between the underlying quality of a business and its machine-readable presence.

7.3 Self-reported versus observed behaviour

Reporting that one “trusts” AI recommendations is a different behaviour from clicking on one, contacting a business because of one, purchasing from it, returning to it, or verifying it first. Existing survey evidence rarely distinguishes between these behaviours, and future work should combine surveys with behavioural, experimental, and transaction-level data.

7.4 Causal identification

A consumer who asks an AI system for a business recommendation may already have strong purchase intent; if that consumer subsequently buys from the recommended business, the AI system may not have caused the purchase but merely assisted a decision that was already likely. Establishing the relevant counterfactual — whether the same consumer would have chosen the same business without AI assistance — requires controlled experiments or carefully designed comparisons rather than simple correlations between AI use and purchase.

7.5 The content and accuracy of AI-generated descriptions

Most existing research asks whether consumers use AI, but comparatively little research examines the content of what AI systems actually say about specific businesses: the accuracy of claims, the tone used, the order in which options are presented, which businesses are included or excluded, whether outdated information is repeated, and whether small or unfamiliar businesses are treated differently from established ones. A business could be disadvantaged not because it is absent from an AI system’s output, but because it is described inaccurately or evaluated against inappropriate criteria.

7.6 Long-term and dynamic effects

There is very little evidence on whether AI-mediated discovery changes brand loyalty, price sensitivity, market concentration, or customer acquisition costs over time, or whether it changes businesses’ incentives to produce public information about themselves.

8. Proposed Research Questions

The following questions are intended to generate original, empirically grounded findings rather than to restate existing commentary.

1. Accuracy of AI recommendations for local businesses in Nairobi. How accurately do major AI systems identify and describe businesses in categories such as restaurants, clinics, education, beauty, professional services, and retail? Relevant measures include inclusion rate, factual accuracy, use of outdated information, ranking consistency, price and location accuracy, and citation quality.
2. Breadth versus concentration of consideration. Do AI recommendations help consumers discover less-established businesses, or do they repeatedly recommend already prominent businesses? This can be examined by comparing recommendation rates across highly reviewed firms, new firms, small independent firms, and businesses with strong social-media presence but limited website documentation.
3. AI recommendations versus personal referrals. Under what conditions do consumers trust an AI-generated recommendation more than a recommendation from someone they know, and how does this vary with product risk, price, urgency, and the visibility of citations?
4. The persuasive effect of explanation. Does an AI-generated recommendation become more persuasive when it explains why a business is suitable, cites sources, or discloses uncertainty, compared with a bare name-only recommendation?
5. AI exposure and brand perception. How does encountering a business through an AI-generated comparison, rather than through the business’s own advertising, a customer review, or a direct referral, affect perceived expertise, trustworthiness, and distinctiveness?
6. Discovery economics. Does AI-mediated discovery reduce the importance of website clicks while increasing the importance of being shortlisted, and how should businesses and researchers measure value that does not appear as a referral click (for example, direct calls, branded searches, or store visits)?
7. Language and local context. How does the language of a query — English, Kiswahili, Sheng, or mixed usage — affect which businesses are surfaced and how they are described?
8. Machine legibility versus customer value. Do businesses that improve their structured online information become genuinely easier for consumers to evaluate, or merely easier for AI systems to describe, and does improving one improve the other?

9. Proposed Research Design

A credible first study addressing these questions would use a four-phase mixed-method design.

Phase 1: Evidence review

A systematic review distinguishing peer-reviewed research on information search, consumer trust, recommender systems, source credibility, and choice architecture from industry surveys and platform announcements, with explicit attention to methodological quality and sample composition.

Phase 2: Prompt audit

A structured audit of 50 to 100 businesses across several Nairobi categories, using standardised prompts (for example, requests for the best businesses in a category for a stated need, direct comparisons of named businesses, and requests for what a consumer should verify before choosing). Each audit would record which businesses appear, their ordering, the descriptions used, any citations or supporting evidence, contradictions, missing information, and stability across repeated queries over time.

Phase 3: Controlled consumer experiment

Random assignment of participants to different information conditions — traditional search results, an AI summary with citations, an AI summary without citations, a human-written review, and an AI summary with explicit uncertainty disclosure — measuring perceived trust, perceived usefulness, number of alternatives considered and recalled, choice intention, willingness to verify, and confidence in the final decision.

Phase 4: Qualitative interviews

Interviews with consumers on how they actually use AI for business decisions: what prompts they write, when they ask follow-up questions, what makes them suspicious of an answer, what they verify independently, whether they distinguish paid from organic recommendations, and when they return to conventional search, maps, social platforms, or personal contacts. Interviews are particularly important because consumers often describe AI use as “research” even when it substantially shapes their eventual decision.

10. Discussion and Implications

The evidence reviewed here supports neither of the two extreme positions sometimes advanced in public discussion of this topic — that AI has replaced search and rendered conventional marketing obsolete, or that nothing fundamental has changed. The more defensible position is that AI is becoming an intermediary in both the discovery and evaluation stages of consumer decision-making, that its influence is growing, and that its effects on decision quality, market fairness, and business outcomes remain uneven and, in most respects, under-measured.

For businesses, this suggests that competing to be found is necessary but no longer sufficient; a business must also be described accurately, and favourably, by systems that explain it to someone else on its behalf. Where an AI system cannot identify a business’s clear customer, proposition, price, location, or evidence of value, the underlying difficulty may not be a visibility problem but a clarity problem in the business’s own public information.

For researchers, the priority is to move from self-reported measures of AI use toward behavioural and experimental designs capable of isolating causal effects, and to extend the evidence base beyond the United States and Western Europe into markets, such as Kenya, where the mix of information channels, business documentation, and consumer trust is likely to differ substantially.

11. Limitations of this Review

This paper is a narrative synthesis of secondary evidence rather than a systematic review or original empirical study. It relies substantially on industry-commissioned surveys whose methodologies vary in transparency and rigour, and on a single peer-reviewed-standard behavioural study (Chapekis & Lieb, 2025) and one pre-registered experimental preprint (Li & Aral, 2025) that, at the time of writing, had not yet completed formal peer review. Figures drawn from commercial surveys should be read as broadly indicative of directional trends rather than as precise or directly comparable estimates. The research agenda proposed in Sections 8 and 9 is intended to address these limitations through original, transparent, and pre-registered data collection.

12. Conclusion

Artificial intelligence is changing where and how consumers first encounter and compare businesses, but the available evidence does not yet show that it has replaced independent research, peer recommendation, or direct interaction with a business. The distinction between visibility, persuasion, and conversion — formalised here as discovery, evaluation, and action — offers a more precise way to describe what is changing and what remains uncertain. The most significant gap in the current evidence base is geographic: almost nothing is known about how AI-mediated discovery operates in markets such as Kenya, where local and service businesses depend on different forms of documentation and trust than the structured, product-centred markets that dominate existing research. The research agenda proposed in this paper is intended to begin closing that gap.

References

Bazaarvoice. (2025, September 30). Bazaarvoice Shopper Experience Index 2025: As AI search grows in popularity, ratings and reviews feed LLMs [Press release]. https://www.bazaarvoice.com/press/bazaarvoice-shopper-experience-index-2025-as-ai-search-grows-in-popularity-ratings-and-reviews-feed-llms/

Capgemini Research Institute. (2025, January 9). 71% of consumers want generative AI integrated into their shopping experiences [Press release]. Capgemini. https://www.capgemini.com/no-no/news/press-releases/71-of-consumers-want-generative-ai-integrated-into-their-shopping-experiences

Chapekis, A., & Lieb, A. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

Court, D., Elzinga, D., Mulder, S., & Vetvik, O. J. (2009). The consumer decision journey. McKinsey Quarterly. https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/the-consumer-decision-journey

eMarketer. (2025, July). Shoppers trust genAI more than brand websites for purchase advice. https://www.emarketer.com/content/shoppers-trust-genai-more-than-brand-websites-purchase-advice

Google. (2024, October 28). AI Overviews in Google Search expanding to more than 100 countries [Blog post]. The Keyword. https://blog.google/products/search/ai-overviews-search-october-2024/

KPMG LLP. (2024). 2024 KPMG generative AI consumer trust survey. https://www.kpmg.com/us/en/media/news/generative-ai-consumer-trust-survey.html

Li, H., & Aral, S. (2025). Human trust in AI search: A large-scale experiment (arXiv:2504.06435) [Preprint]. arXiv. https://arxiv.org/abs/2504.06435