The friend most shoppers now ask first has never used the product. In a September 2025 interview study by Conveo, 51% of 330 shoppers across eight countries named AI recommendations as a source they trust, against 26% who said the same of friends and family. Conveo sells AI research tools, so treat the exact split with care. The direction matches larger samples, though. Capgemini’s January 2025 consumer report found 58% of 12,000 consumers had replaced traditional search engines with generative AI tools for product or service recommendations, up from 25% a year earlier.
That should unsettle anyone who runs a referral program. Referral marketing was built on a bet that people believe people they know. The bet still pays. But a new intermediary now sits between the customer who loves you and the prospect who has never heard of you, reading what your customers say in public and handing it back as a recommendation. The referral marketing statistics in this report cover both halves of that story, the classic economics of referred customers and the new behavior of AI as the referrer, with every figure dated and sourced. At the end you get a scoring tool for whether the machines would refer you the way your best customers do.
Referred customers stay longer, and the data explains why
The most cited research on referral value comes from a German bank. Philipp Schmitt, Bernd Skiera and Wharton’s Christophe Van den Bulte published “Referral Programs and Customer Value” in the Journal of Marketing in January 2011. They tracked 5,181 referred and 4,633 nonreferred customers acquired in 2006, following each one until September 2008. Referrers received a 25 euro voucher for each new customer.
The findings hold up well. Referred customers produced a contribution margin about 25% higher in the early period, after controlling for demographics. Their churn hazard was about 18% lower. Over a six-year horizon, the authors estimated referred customers were worth at least 16% more than nonreferred ones, and about 25% more once the lower acquisition cost was counted. They put the return on the referral voucher at roughly 60%.
Two details matter more than the headline. The margin advantage vanished after about 29 months, by the authors’ estimate. The retention advantage did not fade. Referral brings in a customer who is harder to lose.
A 2018 follow-up in the Journal of Marketing Research by Van den Bulte, Emanuel Bayer, Skiera and Schmitt, using almost 1,800 referred customers at the same bank, explained why. As the American Marketing Association summarized it, better matching drives the margins: referrers bring in people who share traits no application form captures, such as openness to advice. Social enrichment drives the loyalty: a friend inside the relationship builds trust and attachment.
The most telling result: when a referrer left the bank, the person they referred became about as likely to leave as any other customer. The loyalty was to the person who vouched for the bank. Remember that when we get to AI assistants, because they vouch for brands too, and can stop without telling you.
How much buying still runs on word of mouth?
More than most attribution models will ever show. McKinsey’s Jacques Bughin, Jonathan Doogan and Ole Jørgen Vetvik estimated in the McKinsey Quarterly in April 2010 that word of mouth is the primary factor behind 20% to 50% of all purchasing decisions. The effect was strongest when consumers bought a product for the first time or when it was expensive, the moments when people do the most research and feel the most risk.
The same piece made a point referral teams still underrate. A high-impact recommendation, from a trusted person who knows the category and describes the product in specific terms, was up to 50 times more likely to trigger a purchase than a low-impact one.
Trust data points the same way. Nielsen’s 2021 Trust in Advertising study found 89% of consumers trust recommendations from people they know, the most trusted channel in that study. We covered more of that trust picture in our word of mouth marketing statistics, so here we will only draw the line that matters for referral: personal recommendation is the format consumers trust most, and McKinsey’s work says the content of that recommendation decides whether it converts.
That is where AI changes the economics. A human referral carries one person’s trust to a handful of friends. An AI answer carries a compressed version of thousands of recommendations to millions of people. If that version is specific and favorable, it behaves like a high-impact referral at scale.
What does a referral program actually return?
Referral programs earn their budget when aimed at the right customers, who are often not the biggest spenders. V. Kumar, J. Andrew Petersen and Robert Leone studied 9,900 customers of a telecom company and 6,700 customers of a financial services firm, published in the Harvard Business Review in October 2007. Their central finding was that high-volume spenders and good referrers are generally not the same people.
When the firms targeted customers by referral value instead of spending, the telecom campaign cost about $31,500 and produced gains of $486,090, roughly $15.40 per dollar spent, according to Insurance Journal’s February 2008 coverage. The financial services campaign returned about $13.60 per dollar. Both firms had been seeing $4 to $6 per marketing dollar from their usual campaigns. Those figures exclude indirect referrals, so the true return was likely higher.
Rewards help, but least of all for strong brands. Gangseog Ryu and Lawrence Feick ran four experiments for a 2007 Journal of Marketing paper on referral reward programs. As summarized by Baylor University’s Keller Center in 2008, offering a reward increased referral likelihood by more than 20 percentage points for a weaker brand and by less than 10 points for a stronger brand. The authors also found that rewards matter more between weak ties, where the relationship runs on reciprocity, than between close friends, who refer without expecting anything back.
One caution before you put referral marketing statistics in a board deck. A figure attributed to Texas Tech University, claiming most satisfied customers are willing to refer but few actually do, appears on dozens of vendor sites. None of the pages we checked name a study, author, year or journal. We could not trace it to a published source, so it does not appear in this report. The idea may be true. A number without a source is still not evidence.
The new friend in the conversation is a chatbot
Capgemini Research Institute’s What Matters to Today’s Consumer report, released January 9, 2025 and based on 12,000 consumers in 12 countries surveyed in October and November 2024, found 58% had replaced traditional search engines with generative AI tools for product or service recommendations. In 2023 the figure was 25%.
Younger buyers have gone furthest. A Mastercard survey, reported by YPulse on September 17, 2026, found 31% of teens trust AI-generated product recommendations more than a friend’s, and 23% trust them more than their parents’. Sixty-two percent of teens use AI to research products every month, compared with 49% of parents. YPulse’s own data adds a counterweight: trust in friends and family still outranks AI in several product categories. The friend has company now, not a replacement.
The Conveo study from the opening shows what happens next. Of its 330 participants, interviewed in September 2025 across eight countries including the US, UK, India and Brazil, 53% said they had tried a new brand because of an AI suggestion and 15% said they had switched brands. Nine in ten said they check AI outputs against other sources, including reviews, forums and people they know. That last figure is the important one for referral teams. Buyers use AI to form the shortlist and humans to confirm it. You need to be present in both.
Business buyers moved as fast. G2 surveyed more than 1,000 B2B software buyers in August 2025 and published the results on October 6, 2025. Half said they now start their buying journey in an AI chatbot rather than Google Search, a 71% jump from a survey four months earlier. AI chat ranked as the top source for building a software shortlist, ahead of review sites and vendor websites. In B2B, where referrals have always carried enormous weight, the first recommendation a buyer hears is increasingly a machine’s.
Why do AI referrals convert like personal ones?
Because they arrive the way a friend’s tip arrives: pre-filtered, specific and attached to a reason. The conversion data reflects that. Adobe Analytics reported on January 12, 2026 that traffic from generative AI tools to U.S. retail sites converted 31% better than other traffic sources during the 2025 holiday season, nearly double the gap of a year earlier. AI-referred shoppers spent 45% more time on site and were 33% less likely to leave immediately.
Adobe’s survey of more than 1,000 U.S. consumers found 68% were less likely to return a product bought with AI help. That is the same signature the bank study found in referred customers: better matching, fewer regrets.
Software shows an even sharper pattern. Ahrefs published its own numbers on June 16, 2025. AI search sent just 0.5% of its website visitors over a 30-day period, yet those visitors produced 12.1% of signups, a conversion rate 23 times higher than traditional organic search. Patrick Stox, who wrote the analysis, noted it covers a single site. Still, anyone who has studied referral cohorts knows the shape: small volume, outsized value.
The parallel is not perfect. A friend has used the product; an assistant has read about people who did. But for the buyer, a trusted voice narrowing a big field to a few names feels the same, and converts the same. Treat AI answers as a referral channel, not only a search channel.
Machines repeat what people say in public
If AI assistants are the new referrers, where do they get their opinions? Largely from the public record of human recommendation. Profound analyzed 680 million citations across ChatGPT, Google AI Overviews and Perplexity between August 2024 and June 2025, published in June 2025 and updated that August. Reddit was the most cited source on Perplexity at 6.6% of all citations, and 46.7% of citations among Perplexity’s top 10 sources. On Google AI Overviews, Reddit led too, at 2.2% of all citations. ChatGPT leaned hardest on Wikipedia, at 7.8%.
Read that as a referral statistic. Reddit threads are written word of mouth, people asking for recommendations and others answering. When an assistant cites them, it relays referrals at scale. The dinner-table referral leaves no trace a model can read. A detailed forum answer, a podcast mention or a trade quote becomes retrieval material. We go deeper on this in our piece on Reddit as an AEO channel.
Trust in the messenger still follows the old rules. Edelman’s Trust Barometer Flash Poll, released November 18, 2025 and covering Brazil, China, Germany, the UK and the U.S., found that “someone like me” was about twice as trusted as a CEO or a government leader to tell the truth about AI. In the U.S., 49% said they reject the growing use of AI and only 17% embrace it. The buyer asking ChatGPT for a shortlist may distrust AI in general. What they trust is the sense that the answer reflects people like them. The assistant borrows credibility from the humans it cites.
That gives referral programs a second job: making sure the reasons customers recommend you exist in public, in words specific enough for a model to repeat.
Score yourself with the Referral Proxy Scorecard
The Referral Proxy Scorecard is a 10-point check on whether an AI assistant would recommend you the way your best customers already do. Your customers are your first proxies, people who speak on your behalf; the assistants are your second. The scorecard tests whether the second set has enough material to repeat the first. Score each of the five criteria zero, one or two points.
Stated reason. Ask 10 recent customers who came in through a referral one question: what did the person who referred you say about us? Score two if most answers name the same specific reason (fastest onboarding in the category, the only firm that handles a certain problem). Score one if the answers are positive but scattered. Score zero if they are generic, such as “they’re great.” If your own customers cannot say why, a model cannot either.
Public trace. Search Reddit, relevant forums, review platforms and podcast transcripts for your brand name over the last 12 months. Score two if you find at least 10 unprompted, detailed recommendations from people who are not your staff or paid creators. Score one for three to nine. Score zero for fewer than three. This is the written word of mouth that Profound’s citation data shows assistants actually read.
Independent voice. Count articles, trade-press quotes and podcast episodes from independent outlets that named you in the last 12 months. Score two for six or more, one for one to five, zero for none. Buyers check AI answers against exactly these sources. Our earned media statistics cover why that window matters.
AI echo. Write 10 questions a referred customer might have asked before buying, in their words. Run each in ChatGPT, Gemini, Perplexity and Google’s AI Mode. Score two if you are named in more than half of the 40 answers and the stated reason matches what your customers say. Score one if you appear but with the wrong reason or a vague one. Score zero if you rarely appear. A wrong reason sends you poorly matched prospects, undoing the matching advantage the bank research found.
Referrer reach. Score two if your program invites the customers with the highest referral value (often not your top spenders, per Kumar) and gives them a specific way to share in public as well as in private. Score one if you ask everyone the same way. Score zero if you have no structured ask at all.
Read the total in three bands. Eight to 10 means your human referrals and your AI referrals tell the same story, and the main work is maintenance. Four to seven means your customers like you but the public record is thin or off-message; fix the lowest criterion first. Zero to three means the assistants are building shortlists in your category with little or no input from the people who love you.
Where should referral budgets go in 2027?
The referral marketing statistics above point in one direction. Keep paying for the classic program, because the bank, telecom and reward studies all show it works, especially when you choose referrers by referral value. But stop treating the private introduction as the whole channel.
Put most of your referral effort into asking the right customers for something more useful than a forwarded link. Ask whether they would answer a category question on a forum they already use, join a case study a trade reporter can quote, or describe on a podcast the problem you solved. Each creates a public trace an assistant can cite long after the bonus is paid.
Put a smaller, steady share into earned media. A trade-press quote that names the specific reason customers choose you persuades the human reader and feeds the next AI answer.
Then run the AI echo check every quarter, alongside your referral reporting. If referred customers keep citing one reason and the assistants keep citing another, that gap is your next campaign. If you want a baseline before building the scorecard yourself, Instant Press runs a free AEO rating that shows how AI tools describe your brand when buyers ask about your category.
The bank study’s quietest finding may be its most useful one: referred customers stayed because the person who vouched for the brand stayed. Your customers will keep vouching for you over dinner. The open question is who vouches for you in the chat window, so when a buyer asks an assistant for a recommendation in your category tomorrow morning, whose words do you think it will repeat?