At risk is a label used in customer relationship management (CRM) for buyers whose history makes them valuable but whose recent silence suggests they are drifting away. In the recency, frequency and monetary value (RFM) model, it marks one corner of the grid: customers who used to order often and spend well, but whose last order was some time ago. It exists because retail customers rarely announce that they are leaving; a business only sees the orders stop. At risk names the window between the last purchase and the point at which the customer is written off, when a reminder, an offer or an advertisement might still bring them back.

How the segment is built

RFM scores each customer on three axes: how recently they bought, how often and how much they spent. Arthur Middleton Hughes, whose 1994 book Strategic Database Marketing popularised the method, sorted a customer file into five equal parts on each axis, producing codes from 555 down to 111 and 125 cells in all. Named segments are groupings of those cells.

Shopify's version shows the mechanics. Each customer receives a score from 1 to 5 per axis, relative to the store: 5 means the top 20%, 1 the bottom 20%, according to Shopify's help centre. Frequency and monetary scores are averaged and rounded down. A customer with a recency score of 2 or less and a combined score above 2 but no higher than 4 is At risk: no recent purchases, "but with a strong history of orders and spend". A shopper scored 2-4-3 averages 3.5, rounds down to 3 and lands in At risk; one scored 2-5-5 is Previously loyal, and one scored 2-1-2 is Dormant. Merchants see only the group, which segment queries reach through the filter rfm_group = 'AT_RISK' and developers through the Admin API since version 2025-04.

Other tools use absolute thresholds. Klaviyo scores each axis from 1 to 3: recency earns 3 for a purchase within 180 days, 2 within 365 days and 1 beyond that, while frequency and spend are split into thirds, according to its help centre. It needs at least 500 customers with orders and 180 days of history. Omnisend's fallback, for stores with fewer than 100 returning customers, treats a last purchase 90 to 180 days ago as at risk, according to its support documentation.

Predictive tools replace cells with probabilities. Google Analytics 4 (GA4) calculates churn probability, the likelihood that a user active in the past seven days will not be active in the next seven, and offers a likely 7-day churners audience, with the cut-off set at the 80th percentile in Google's template. The model needs at least 1,000 returning users who churned and 1,000 who did not over a seven-day period within the last 28 days, according to Google. Braze scores churn risk from 0 to 100 and treats 75 and above as high risk, according to its documentation. Klaviyo's predictive layer attaches a churn risk to every profile.

Where it meets advertising

CRM teams build the segment; paid media increasingly buys against it. Google's customer retention goal, whose win-back modes were announced in April 2025, bids higher for lapsed customers in Performance Max. Google leaves the definition of lapsed to the advertiser, who uploads a Customer Match list and values such customers between new and existing ones, according to Google's help centre. Uploaded lists have also been subject since April 7, 2025 to a 540-day membership cap, which limits how long a quiet customer stays addressable.

Meta has adopted the label itself. Its Marketing API lets advertisers tag a custom audience as AT_RISK, for customers showing signs of disengaging or churning, separately from DISENGAGED, according to Meta's developer documentation. Labels feed value rules for audiences, according to the consultant Jon Loomer. Meta opened value rules to all ad accounts in June 2025, and they can raise bids by up to 1,000% or cut them by 90%. Meta warns that cost per result may rise.

Retail media sells the raw material: Amazon Marketing Cloud sorts a brand's customers into percentile buckets by historical spend.

Origin and evolution

RFM grew out of catalogue and direct mail, where the question was which names justified postage. Practitioner histories credit the direct marketer George Cullinan with the method in the early 1960s, though the earliest dated document cited for him is a 1977 Direct Mail/Marketing Association manual, according to a 2004 paper by Alencar and co-authors. Hughes later wrote that "recency is a more powerful predictor of customer response than frequency". Many online guides credit Jan Roelf Bult and Tom Wansbeek's 1995 Marketing Science paper with introducing RFM, though it postdates Hughes's book and its abstract describes a method for selecting mailing targets.

Academic work then recast silence as probability. The Pareto/NBD model of David Schmittlein, Donald Morrison and Richard Colombo, published in Management Science in 1987, computed the chance that a customer is still active. In 2005 Peter Fader, Bruce Hardie and Ka Lok Lee found the logic of the at-risk cell in the data. "For people with low recency, higher frequency seems to be a bad thing," they wrote in the Journal of Marketing Research: a regular buyer who stops has probably gone.

The named table came from practitioners. An 11-segment version, with At Risk described as having spent big and bought often but long ago, beside Can't Lose Them and Hibernating, was published by the analytics vendor Putler on April 14, 2017. Who devised it is unclear. Software followed: GA4's predictive metrics in July 2020, Braze churn prediction in October 2020, Klaviyo's RFM groups in late 2023 and Shopify's RFM report in its Winter '25 Edition, released in December 2024.

Why marketers care

The commercial case rests on old arithmetic. Frederick Reichheld and W. Earl Sasser reported in the Harvard Business Review in 1990 that cutting defections by 5% raised profits by 25% to 85% in the service businesses they studied. By 2014 an HBR retelling had stretched the range to 95%. Timing matters because departures rarely reverse: RevenueCat found that 95% of annual app subscribers who cancel never return, against 18% to 24% of monthly subscribers who do.

One widely shared funnel framework lists predictive AI to flag churn risk as a loyalty-stage task, and the IAB Australia Retail Media Council's December 2025 blueprint assigned churn-risk segments to retailers while calling RFM static and backward-looking. For advertisers, the segment decides who is excluded from acquisition campaigns, who receives win-back bids, and which spend is reported as retention rather than growth.

Limitations and disputes

The label is not standard. Shopify reserves At risk for customers with strong histories. Klaviyo applies it to customers who have not bought recently "and/or tend to spend less overall", and calls the valuable lapsing group Needs attention. Bloomreach places At risk beside "Cannot lose them" but with more recent purchases. A segment exported from one tool to another can change meaning on the way.

Relative scoring also guarantees supply. When scores are quintiles, a fifth of every customer base holds the lowest recency score whether or not behaviour has changed. Fixed windows ignore purchase cycles: 120 days without an order is normal for a mattress buyer and alarming for a coffee subscriber. Fader and Hardie noted in 2009 that in non-contractual settings the moment a customer leaves "is unobserved by the firm", and showed that Amazon's count of active customers shrinks if its 12-month cut-off becomes nine months.

The deeper dispute is whether the riskiest customers are the right targets. Eva Ascarza's 2018 paper Retention Futilityfound that ranking customers by their response to an intervention, rather than by churn risk, would have cut churn by a further 4.1 and 8.7 percentage points. In one study, contacting the 40% at highest risk would have raised churn by 4.4 points. Contact can backfire outright. In a 2016 study of 64,147 customers of a South American wireless carrier, Ascarza, Raghuram Iyengar and Martin Schleicher found churn of 10.0% among customers encouraged to switch plans, against 6.4% in the control group.

Profiling has legal weight too. Under the General Data Protection Regulation (GDPR), Article 21(2) gives people an unconditional right to object to direct marketing, including profiling related to it, as guidelines endorsed by the European Data Protection Board state.

Not the same as

Churn is the departure itself, or the rate of departures. Churn is counted after the fact; at risk is a forecast made before it.

Lapsed and dormant customers have already crossed a line the business drew. Google's re-engagement mode works on lapsed lists, and Meta's DISENGAGED label covers customers with no recent purchases.

Can't Lose Them, which Shopify calls Previously loyal, is the neighbouring cell for the most valuable lapsing customers.

Budget at risk is a delivery metric, not a customer segment. Amazon DSP introduced it on September 28, 2023 as budget likely to go unspent if current pacing continues, according to Amazon Ads, and its overview page highlights campaigns most at risk of underdelivering. Google AdSense's "Earnings at risk" alert warns publishers about ads.txt problems.

Recent developments

Google Ads API version 25, effective July 22, 2026, removed the legacy lifecycle goal resources and added optimisation for retaining loyalty programme members, months after Google extended loyalty programme ads to 14 countries. Google Ads conversion goals now include a customer lifecycle section, according to Search Engine Roundtable on August 26, 2026.

Meta rebuilt the Audiences section of Ads Manager around labels in July 2026, with At risk among the customer categories and one label allowed per audience. Loomer doubted many advertisers would bother. "It feels too complicated," he wrote. Sam's Club said in September 2026 that it could reach lapsed or staggered buyers, members who stopped or slowed buying a brand.

Timeline

  • Early 1960s: Later histories credit direct marketer George Cullinan with devising recency, frequency and monetary selection for mailing lists
  • 1977: The Direct Mail/Marketing Association publishes Cullinan's RFM manual
  • 1987: Schmittlein, Morrison and Colombo publish the Pareto/NBD model, estimating whether a customer is still active
  • September 1990: Reichheld and Sasser publish Zero Defections in the Harvard Business Review
  • 1994: Arthur Middleton Hughes publishes Strategic Database Marketing, with quintile scoring into 125 RFM cells
  • 1995: Bult and Wansbeek publish Optimal Selection for Direct Mail in Marketing Science
  • November 2005: Fader, Hardie and Lee link RFM to customer lifetime value in the Journal of Marketing Research
  • 2009: Fader and Hardie set out probability models for non-contractual customer bases
  • February 2016: Ascarza, Iyengar and Schleicher report a retention experiment that raised churn
  • April 14, 2017: Putler publishes an 11-segment RFM table including At Risk
  • February 2018: Ascarza publishes Retention Futility in the Journal of Marketing Research
  • July 10, 2020: Google introduces predictive metrics, including churn probability, in Google Analytics
  • October 20, 2020: Braze launches churn prediction
  • September 28, 2023: Amazon DSP adds a budget at risk metric for order delivery
  • November 6, 2023: Klaviyo describes the research behind its RFM groups
  • December 2024: Shopify adds an RFM analysis report with an At risk group
  • April 1, 2025: Shopify exposes RFM groups in its Admin API
  • April 7, 2025: Google's 540-day Customer Match membership cap takes effect
  • April 8, 2025: Google announces win-back modes for lapsed customers in Performance Max
  • June 2025: Meta opens value rules to all ad accounts
  • December 10, 2025: IAB Australia's Retail Media Council publishes its audience targeting blueprint
  • April 2026: Meta value rules for audiences, using labels including At risk, reach some advertisers
  • July 2026: Meta reorganises Ads Manager audiences around labels
  • July 22, 2026: Google Ads API version 25 removes legacy lifecycle goal resources
  • August 26, 2026: Search Engine Roundtable reports a customer lifecycle section inside Google Ads conversion goals
  • September 2026: Sam's Club introduces predictive targeting of lapsed and staggered buyers

Summary

Who: Brands' CRM, email and loyalty teams build at-risk segments in tools such as Shopify, Klaviyo, Braze, Omnisend and GA4. Advertising platforms including Google, Meta and Amazon accept the segments as uploaded lists or audience labels, and retailers such as Sam's Club sell predictive versions. Academics including Peter Fader, Bruce Hardie and Eva Ascarza have tested how well the idea works.

What: A customer segment for buyers with a strong record of purchases and spend who have not bought recently. In RFM it means a low recency score combined with middling to high frequency and monetary scores; in predictive tools it is a churn probability above a threshold.

When: RFM selection is traced to direct mail in the 1960s and was codified by Hughes in 1994. The named At Risk segment spread through practitioner guides from 2017, and was built into analytics and CRM software between 2020 and 2024 and into Google and Meta advertising tools in 2025 and 2026.

Where: In e-commerce platforms, email and CRM software, and analytics products, and from there in Customer Match lists, Meta custom audiences and retail media networks.

Why: Customers in non-contractual businesses leave silently, and those who cancel rarely return. The segment gives marketers a point at which to intervene, though research shows the customers at highest risk are not always the ones an intervention helps, and the label means different things in different tools.