Buyer Behavior Science
Trust in the Machine
Chatbot use jumped from a third of US adults to about half in two years, but the moment a synthesized answer appears on a results page, buyers click and verify less, not more.
Part of Choice Science in the Insights library.
Abstract
AI chatbot use among US adults jumped from 33% to 49% in two years, yet the people using them remain openly skeptical of the accuracy of what they're told. The clearest evidence of what that means for buying decisions isn't in a trust survey at all: once a synthesized answer appears on a search results page, click-through to any underlying source, including the ones the answer itself cites, falls sharply. Trust in AI output looks less like a single number rising or falling and more like a rearrangement of when buyers bother to check at all.
The surface itself is moving fast, and moving unevenly. In the United States, Pew Research Center found that 49% of adults now use AI chatbots, up from 33% just two years earlier. The same acceleration shows up in a narrower, more consequential slice of the buying process: weekly use of AI chatbots for news reached 10% of audiences across the 48 markets the Reuters Institute surveys, up from 7% the year before, concentrated among under-35 audiences at 16%. But usage growth and trust are not moving together. Globally, only 20% of people say they trust the answers AI chatbots give about news, a full 17 points under the 37% who say they trust news overall, and in the UK that figure collapses to just 6%. The pattern holding across both the US and global data is the same: people are adopting the surface faster than they are extending it credit.
What that gap does to actual buyer behavior is more interesting than the stated-trust numbers suggest, and it shows up in what people click rather than what they say in a survey. Pew's browsing-data study, drawn from real Google search sessions rather than self-report, found that when an AI summary appears on a results page, click-through to a traditional search result falls from 15% to 8%, and only 1% of visits to a summary page click through to a source the summary itself cites. Session abandonment rises too, from 16% without a summary to 26% with one. None of this reads as buyers deciding an AI answer is more trustworthy than a linked source underneath it; it reads as buyers deciding, in the moment, that checking isn't worth the click. That lines up with older algorithm aversion and algorithm appreciation research: Dietvorst and colleagues found people lose confidence in an algorithm faster than in an equally fallible human once they see it err, while Logg, Minson, and Moore found the opposite in more routine, objective judgment tasks, where advice framed as coming from an algorithm was weighted more heavily than identical advice from a person. Buyers aren't applying one trust rule. They're applying a different rule depending on the stakes and the framing of the decision in front of them.
For where attention and return concentrate next, the review layer and the AI-answer layer are converging rather than one replacing the other. BrightLocal's 2026 panel found 97% of shoppers still read reviews before choosing a local business and 49% now trust those reviews as much as a personal recommendation, while the same panel found AI chatbot use as a local-discovery source jumped from 6% to 45% year over year, with 42% saying they trust AI platforms as much as traditional reviews. Read together with the click-through collapse, that means the moment of decision is moving inside the answer itself, not onto a separate new channel next to the old ones. Early empirical work on generative engine optimization found that specific content and wording choices can lift a source's visibility inside a generative engine's answer by as much as 40%, with the effective method varying by topic. That is the frontier the Visibility Corpus exists to map: whether a brand is the thing an engine chooses to say.
The data, in one read
The Number That Doesn't Match the Mood
Forty-nine percent of US adults now use AI chatbots, up from 33% just two years ago, according to Pew Research Center's Americans and AI 2026 survey of 5,119 US adults fielded in February. That's the kind of adoption curve that usually accompanies a technology people have decided to trust. This one didn't come with that decision attached.
Among the adults who still don't use AI chatbots, distrust of accuracy is a leading reason: 45% call it a major reason for staying away, 76% call it a major or minor reason combined, and roughly six in ten US adults overall say they aren't confident AI companies will develop and use the technology responsibly, per the same Pew report. Put the two findings side by side and the picture isn't a market that has made up its mind. It's a market where a majority has started using something a majority also says it doesn't fully trust, which is a much less stable foundation for a buying decision than the adoption number alone suggests.
Trust Depends on What You're Asking It to Do
Ask a chatbot for news and the trust gap widens further. The Reuters Institute for the Study of Journalism found that global trust in AI chatbot answers for news sits at just 20%, a 17-point deficit against the 37% of people who say they trust news overall, drawn from a 48-market survey of roughly 2,000 respondents per market. In the UK, trust in chatbot news answers falls to just 6%.
That's a strikingly different number from what a separate, vendor-run panel finds in commerce. BrightLocal's 2026 US consumer panel found 42% of consumers now say they trust AI platforms as much as traditional reviews for local recommendations, and 49% say they trust online reviews as much as a personal recommendation from a friend or family member. News and a local-business recommendation are not the same trust problem: one carries civic and factual stakes, the other is closer to a low-cost, reversible, largely subjective judgment. The same technology earns very different credit depending on what it's being asked to settle.
The same technology earns dramatically different trust depending on what it's being asked to settle.
What People Click Tells a Different Story Than What People Say
Survey answers about trust are useful, but they aren't behavior, and the clearest evidence in this study didn't come from a survey at all. Pew Research Center built a panel of 900 people and watched actual Google browsing data covering 68,879 unique searches in March 2025. When an AI summary appeared on the results page, which happened in about 18% of the searches studied, click-through to a traditional search result fell from 15% to 8%. Only 1% of visits to a page with an AI summary clicked through to a source the summary itself cited. Session abandonment, meaning the user left without clicking anything, rose from 16% to 26%.
That is not a picture of buyers deciding the AI answer is more trustworthy than the sources underneath it. It's a picture of buyers deciding, in the moment, that the friction of checking isn't worth it once an answer is already sitting in front of them. Verification didn't get replaced by more trust. It got replaced by less checking.
Verification didn't get replaced by more trust in AI. It got replaced by less checking, period.
When Aversion Flips to Appreciation
This split between stated distrust and reduced verification lines up with a body of behavioral research that predates chatbots entirely. In a foundational 2015 study, Dietvorst, Simmons, and Massey found that people who watched a statistical algorithm make forecasts lost confidence in it faster than in an equally error-prone human, and stayed less willing to rely on it even after watching it outperform the person, a pattern the field now calls algorithm aversion.
A later stream of research found the opposite effect under different conditions. Logg, Minson, and Moore, writing in Organizational Behavior and Human Decision Processes in 2019, found that across multiple experiments, people weighted advice more heavily when it was attributed to an algorithm than when identical advice was attributed to a person, a pattern they called algorithm appreciation, strongest for objective, quantifiable judgments.
A tertiary synthesis of that literature, sourced to a crowd-edited reference rather than a peer-reviewed paper, which is why we hold it at emerging rather than established, proposes the moderator that reconciles the two findings: aversion dominates for high-stakes, subjective, or emotionally loaded decisions such as health or financial advice, while appreciation dominates for low-risk, objective, repeatable judgments, and aversion runs consistently higher for autonomous decide-for-you systems than for advisory recommend-but-you-decide ones. It's a directional frame worth holding, not a number to cite on its own.
Reviews Aren't Losing. They're Merging.
None of this means reviews are being displaced. BrightLocal's 2026 US consumer panel of 1,002 shoppers found 97% still read reviews before choosing a local business, and 49% now trust those reviews as much as a personal recommendation from someone they know.
What's changed is that AI chatbots have joined reviews as a discovery layer rather than replacing them: use of AI chatbots as a source for local business recommendations jumped from 6% to 45% year over year in the same panel, and 42% of consumers say they trust AI platforms as much as traditional reviews for that purpose. This is a single vendor panel, not an independently audited academic study, so we treat the specific percentages as directional rather than settled fact. But the direction itself, AI answers sitting alongside reviews rather than instead of them, is consistent with everything else in this study.
Skepticism Without Slowdown
If people were becoming more confident in AI accuracy, that would be one explanation for rising use. The evidence says the opposite: about half of adults who get news from AI chatbots say they at least sometimes come across information there they believe is inaccurate, per Pew's March 2026 synthesis of its own survey work.
And they keep using it anyway. Weekly use of AI chatbots for news rose from 7% to 10% across the Reuters Institute's surveyed markets, concentrated among under-35 audiences at 16%. Awareness of error and growth in use are running side by side, not trading off against each other. That's a harder pattern for a business to plan around than either straightforward trust or straightforward distrust would be, because it means buyers are using an imperfect surface with their eyes open, not because they've been convinced it's reliable.
The FTC Just Raised the Cost of Faking It
As synthesized content spreads across the same surfaces where reviews live, the US Federal Trade Commission has moved to protect one of the signals buyers still lean on heavily: reviews they believe came from real customers. Its final rule on fake reviews and testimonials, effective October 21, 2024, bans reviews fabricated with generative tools, undisclosed insider reviews, suppression of negative reviews through threats or legal intimidation, and bought or bot-driven trust signals, backed by civil penalties of $51,744 per violation.
That matters for the trust calculus in this study because it's a regulatory bet on the same finding the behavioral data supports: buyers still extend more credit to a review they believe came from a real customer than to an answer an engine produced, but only as long as the review is actually real. As synthesized answers get easier to produce at scale, the authenticity premium on a verifiably human review, and on any transparently sourced answer, gets more valuable, not less.
Getting Cited Is Now a Discipline of Its Own
The click-through data in this study makes one thing clear: if a source isn't the thing an AI answer chooses to cite, it may not get clicked at all, even if it's ranked correctly underneath. Early academic work on generative engine optimization, first published as a 2023 preprint and accepted to KDD 2024, found that specific content and wording interventions could lift a source's visibility inside a generative engine's answer by as much as 40%, with the effective method varying sharply by topic domain. It's one study, not yet a settled field, which is why we hold it at emerging.
But it points at the shift this whole study is really about. The old fight was for a ranking position a buyer would scroll to and click. The new fight, running alongside it rather than replacing it, is for a place inside the answer itself, in a moment when the buyer may never click through to check who's actually behind it. That's exactly the terrain the Visibility Corpus is built to map: where a market's attention actually lands, and whether it lands on you.
The new fight is for a place inside the answer itself, in a moment when the buyer may never click through to check who's behind it.
The evidence, in numbers
Key findings, dated and sourced
-
49% of US adults now use AI chatbots, up from 33% in 2024.
established Pew Research Center, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact (American Trends Panel, n=5,119 US adults, fielded Feb 17-23 2026) (2026-06-17)
-
Among US adults who don't use AI chatbots, distrust of accuracy is a leading reason: 45% cite it as a major reason (76% major or minor combined), and about six in ten US adults overall say they aren't confident AI companies will develop and use the technology responsibly.
established Pew Research Center, Americans and AI 2026, sub-chapter: Why Don't People Use Chatbots? (2026-06-17)
-
Global trust in AI chatbot answers for news sits at just 20%, a 17-point deficit versus the 37% of people who say they trust news overall; in the UK, trust in chatbot news answers falls to just 6%.
established Reuters Institute for the Study of Journalism, University of Oxford, Digital News Report 2026, Executive Summary (48-market survey, roughly 2,000 respondents per market) (2026-06-16)
-
Weekly use of AI chatbots for news reached 10% of audiences across the surveyed markets in 2026, up from 7% the prior year, with usage concentrated among under-35s (16%).
established Reuters Institute for the Study of Journalism, University of Oxford, Digital News Report 2026, Executive Summary (2026-06-16)
-
The FTC's final rule on fake reviews and testimonials, effective October 21, 2024, bans reviews fabricated with generative tools, undisclosed insider reviews, suppression of negative reviews through threats or legal intimidation, and bought or bot-driven trust signals, backed by civil penalties of $51,744 per violation.
established US Federal Trade Commission, FTC Announces Final Rule Banning Fake Reviews and Testimonials (Federal Register doc 2024-18375) (2024-08-14)
-
In the foundational study establishing algorithm aversion, participants who watched a statistical algorithm make forecasts lost confidence in it faster than in an equally error-prone human forecaster, and stayed less willing to rely on the algorithm going forward, even after directly observing it outperform the human.
established Journal of Experimental Psychology: General (American Psychological Association), Dietvorst, B.J., Simmons, J.P., & Massey, C., 'Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err,' 144(1), 114-126, DOI 10.1037/xge0000033 (2015)
-
Countering algorithm aversion, a later stream of research found algorithm appreciation: across multiple experiments, people weighted advice more heavily when it was attributed to an algorithm than when identical advice was attributed to another person, particularly for objective, quantifiable judgment tasks.
established Organizational Behavior and Human Decision Processes, Logg, J.M., Minson, J.A., & Moore, D.A., 'Algorithm Appreciation: People Prefer Algorithmic to Human Judgment,' 151, 90-103, DOI 10.1016/j.obhdp.2018.12.005 (2019)
-
The direction of algorithmic trust flips on task type: aversion to AI dominates for high-stakes, subjective, or emotionally loaded decisions (health, moral judgment, financial advice), while comfort with AI dominates for low-risk, objective, repeatable judgments; aversion also runs consistently higher for autonomous decide-for-you systems than for advisory recommend-but-you-decide systems.
emerging Wikipedia (tertiary synthesis citing Dietvorst 2015, Logg et al. 2019, and subsequent replications), Algorithm aversion (article synthesizing the experimental literature) (2026-06-22)
-
In a 2026 US consumer panel, 97% of shoppers read reviews before choosing a local business and 49% now say they trust online reviews as much as a personal recommendation from a friend or family member; use of AI chatbots as a discovery source for local business recommendations jumped from 6% to 45% year over year, and 42% of consumers say they trust AI platforms as much as traditional reviews for local recommendations.
emerging BrightLocal, Local Consumer Review Survey (US consumer panel, n=1,002) (2026)
-
Measured (not self-reported) Google browsing data shows that when an AI summary appears on a results page, only 8% of users click a traditional search-result link versus 15% when no summary appears; only 1% of visits to AI-summary pages click a link cited inside the summary itself; and 26% of users abandon the session entirely after an AI summary versus 16% without one. AI summaries appeared on about 18% of all searches studied.
established Pew Research Center, Google Users Are Less Likely to Click on Links When an AI Summary Appears in the Results (KnowledgePanel browsing-data study, n=900 panelists, 68,879 unique searches, March 2025) (2025-07-22)
-
About half of adults who get news from AI chatbots say they at least sometimes come across information there they believe is inaccurate, meaning skepticism coexists with continued, growing use rather than suppressing it.
established Pew Research Center, What the Data Says About Americans' Views of Artificial Intelligence (citing an August 2025 Pew survey on AI chatbot news use) (2026-03-12)
-
The first published empirical model for optimizing content visibility inside AI chatbot (generative engine) answers found that specific content and wording interventions can lift a source's visibility in generative-engine responses by up to 40%, with the effective strategy varying sharply by topic domain.
emerging arXiv preprint, accepted to KDD 2024, Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A., 'GEO: Generative Engine Optimization' (2023-11-16)
Learning outcomes
What this study teaches
- Don't read stated distrust of AI as a reason to ignore it: usage and skepticism are rising together, and the buyers you're trying to reach are already using imperfect answers with their eyes open.
- Treat the click-through collapse as the real headline: once an AI summary sits on the page, most buyers won't check further, so being the thing the answer says matters more than being the link underneath it.
- Keep investing in real reviews. Shoppers still read them overwhelmingly (97%), and they're converging with AI answers as a trust source rather than being replaced by them, which raises the cost of any fabricated review under the FTC's rule.
- Match your trust strategy to the stakes of the decision: buyers extend AI more credit for low-risk, objective choices and hold it to a much higher bar for anything high-stakes, subjective, or irreversible.
Honest limits
What this does not yet settle
- There is no independent, standardized census yet of share-of-answer or AI-citation visibility across product and service categories. This remains a genuine measurement frontier, not a solved one, and category-specific answers require direct capture.
- The foundational algorithm aversion and algorithm appreciation studies (Dietvorst 2015, Logg et al. 2019) predate mainstream conversational AI. Whether the same dynamics hold identically for chatbot-style conversational recommendations, versus the single-forecast algorithms those studies tested, is not yet established in peer-reviewed research.
- No controlled purchase-simulation study was found that directly measures conversion or basket-size differences between an AI-recommendation path and a review-research path for the same purchase decision. The available evidence is lab-based judgment tasks or observational search behavior, not end-to-end purchase data.
- The BrightLocal figures on trusting AI as much as reviews come from a single vendor panel, not an independently audited academic or government study, and are held at emerging tier until corroborated elsewhere.
- Two widely circulated claims, a specific trust-threshold figure for AI purchase recommendations over $25 and an Edelman Trust Barometer citation on generative AI's role in the broader trust crisis, could not be independently verified against their primary sources and were excluded rather than included on secondhand paraphrase.
- Cultural and cross-country differences in AI trust, such as individualism versus collectivism, are documented qualitatively in the academic literature but not yet quantified at scale for a US buyer base or an India-based buyer base specifically.
This is a synthesis of dated, attributed evidence, not a census. The AI-answer layer in particular has no independent, Nielsen-grade measurement yet, so readings of it are directional and named as a frontier, never presented as settled.
Straight answers
Frequently asked questions
Is trust in AI chatbots actually growing, or is it just usage that's growing?
It's usage, not trust. Pew Research Center found that 49% of US adults now use AI chatbots, up from 33% just two years earlier, but among the adults who still avoid them, distrust of accuracy is a leading reason, and roughly six in ten US adults overall say they aren't confident AI companies will develop and use the technology responsibly. Globally, the Reuters Institute found only 20% of people trust the answers AI chatbots give about news, 17 points under the 37% who trust news overall. Adoption and trust are moving on separate tracks.
Once an AI summary shows up on a search results page, do buyers check the sources underneath it?
Rarely. Pew's browsing-data study, built from real Google search sessions rather than self-reported survey answers, found that when an AI summary appears on a results page, click-through to a traditional search result falls from 15% to 8%, and only 1% of visits to a summary page click through to a source the summary itself cites. Session abandonment also rises, from 16% without a summary to 26% with one. That's not buyers deciding the AI answer is more trustworthy; it's buyers deciding, in the moment, that checking isn't worth the click.
Does AI trust work the same way for every kind of decision?
No. The study finds trust depends heavily on the stakes and framing of what's being asked. Older behavioral research backs this up: Dietvorst, Simmons, and Massey found people lose confidence in an algorithm faster than in an equally fallible human once they see it err (algorithm aversion), while Logg, Minson, and Moore found the opposite for routine, objective judgment tasks, where advice framed as coming from an algorithm was weighted more heavily than identical advice from a person (algorithm appreciation). A tertiary synthesis proposes that aversion dominates for high-stakes or subjective decisions and appreciation dominates for low-risk, objective ones, though the study holds that reconciling frame at emerging tier since it's sourced to a crowd-edited reference rather than a peer-reviewed paper.
Are AI answers replacing customer reviews, or do reviews still matter?
Reviews still matter and aren't being displaced. BrightLocal's 2026 panel found 97% of shoppers still read reviews before choosing a local business and 49% now trust those reviews as much as a personal recommendation. The same panel found AI chatbot use as a local-discovery source jumped from 6% to 45% year over year, with 42% saying they trust AI platforms as much as traditional reviews, meaning AI answers are joining reviews as a discovery layer rather than replacing them. The study flags this BrightLocal data as a single vendor panel, not an independently audited academic study, so it treats the specific percentages as directional rather than settled.
If click-throughs are dropping, what should a small business actually do about it?
Being the thing an AI answer says now matters more than being the link ranked underneath it, since most buyers won't click further once a summary is on the page. Early academic work on generative engine optimization found that specific content and wording choices can lift a source's visibility inside a generative engine's answer by as much as 40%, though this comes from one study accepted to KDD 2024 and is held at emerging tier, not yet a settled field. The study also stresses keeping real reviews strong, since the FTC's rule against fake reviews (effective October 21, 2024, with penalties of $51,744 per violation) raises the value of a review buyers can verify is genuinely human.
Provenance
References
- Pew Research Center, 'Americans and AI 2026: Chatbots, Smart Devices and Views on Impact,' June 17, 2026. https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/
- Pew Research Center, 'Why Don't People Use Chatbots?' (sub-chapter, Americans and AI 2026), June 17, 2026. https://www.pewresearch.org/internet/2026/06/17/why-dont-people-use-chatbots/
- Reuters Institute for the Study of Journalism, University of Oxford, 'Digital News Report 2026: Executive Summary,' June 16, 2026. https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/dnr-executive-summary
- US Federal Trade Commission, 'FTC Announces Final Rule Banning Fake Reviews and Testimonials,' August 14, 2024 (Federal Register doc 2024-18375). https://www.ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commission-announces-final-rule-banning-fake-reviews-testimonials
- Dietvorst, B.J., Simmons, J.P., & Massey, C., 'Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err,' Journal of Experimental Psychology: General, 144(1), 2015. https://pubmed.ncbi.nlm.nih.gov/25401381/
- Logg, J.M., Minson, J.A., & Moore, D.A., 'Algorithm Appreciation: People Prefer Algorithmic to Human Judgment,' Organizational Behavior and Human Decision Processes, 151, 2019. https://doi.org/10.1016/j.obhdp.2018.12.005
- Wikipedia, 'Algorithm Aversion' (tertiary synthesis of the experimental literature), accessed June 22, 2026. https://en.wikipedia.org/wiki/Algorithm_aversion
- BrightLocal, 'Local Consumer Review Survey 2026.' https://www.brightlocal.com/research/local-consumer-review-survey/
- Pew Research Center, 'Google Users Are Less Likely to Click on Links When an AI Summary Appears in the Results,' July 22, 2025. 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/
- Pew Research Center, 'What the Data Says About Americans' Views of Artificial Intelligence,' March 12, 2026. https://www.pewresearch.org/short-reads/2026/03/12/key-findings-about-how-americans-view-artificial-intelligence/
- Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A., 'GEO: Generative Engine Optimization,' arXiv preprint, November 16, 2023 (accepted KDD 2024). https://arxiv.org/abs/2311.09735
Every measured figure is dated to its capture and tagged with an evidence tier. Every cited work is real and locatable. Where an engine could not be captured this round, it is named as uncaptured, not estimated. Small-sample readings are labelled as directional.