Engagement metrics that predict a sale six months out

A private client director asked in January which clients will buy by June cannot answer from last quarter's revenue. She answers from behaviour: who asked for an appointment without being prompted, who replies to their advisor inside a day, who came to the November viewing and stayed ninety minutes. Weighted and scored, those signals are the nearest thing this sector has to a forecastable pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.View full definition →, and not one of them appears in a web analytics dashboard.

This lesson takes the leading indicators one at a time, turns them into a six-month purchase probability, and then covers the ways that probability lies to you.

Why web traffic misleads in luxury

Sessions, click-through rateclick-through rateClick-Through Rate (CTR) is the percentage of people who click a link, ad, or call to action out of those who viewed it.View full definition → and on-site conversion describe an audience that is mostly aspirational. A client weighing a €35,000 watch browses anonymously for months, then acts through channels no pixel sees: a boutique appointment, a dinner, a call from an advisor who already knows her collection.

Net-a-Porter is the useful counter-example, since it sells luxury goods online and does close the transaction on site. Even there the interesting number is not sessions. Its top customer tier, the Extremely Important People, has long been reported as a low single-digit share of the customer base producing something close to 40% of sales, and those clients are handled by named personal shoppers over messaging rather than by merchandising. The channel changed; the predictive signal did not. Burberry's much-discussed work with Salesforce in the early 2010s pushed the same way: the payoff was less the traffic data than the client record an associate could open on the shop floor, the persistent profile the one-to-one lesson describes.

The practical implication: build the dashboard around identified behaviour, and keep traffic where it belongs, as a brand-reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → measure with no place in a six-month forecast.

The core leading indicators

1. Appointment requests (inboundinboundA strategy that attracts prospects organically via valuable content (blog, SEO, social) rather than interrupting them.View full definition → intent)

An unsolicited request for a private appointment, especially outside a gifting season or a launch, is the strongest single signal available. It costs the client effort and it identifies them.

• Request volume by tier, split between existing top clients, rising clients and prospects. Mixing the three produces a score that tells you only who is already rich in your database.
• Request-to-purchase conversion, measured over three and six months, since the two windows diverge sharply for high jewellery and bespoke orders.
• Time-to-appointment. Advisor-reported data across the sector points the same direction: slipping past one to two weeks costs conversion, because the client has already walked into someone else's salon. This is the one metric on the list your operations team, not your marketing team, controls.

2. Advisor message replies

Reply behaviour is the highest-frequency signal you have, which makes it the most abusable. Track the reply rate over the last six months, the median latency, and the share of exchanges the client starts. A client who initiates one conversation in ten is in a different state from one who only ever answers.

Two failure modes sit here. Score advisors on reply rate and they quietly stop messaging anyone who might not answer, protecting the ratio and starving the pipeline. Score them on message volume and clients receive identical broadcasts, after which the reply rate collapses for everyone and the signal is worthless for a year. Cap outboundoutboundProactive outreach that pushes your message to targeted audiences through advertising, email, or direct prospecting, initiated by the seller rather than the buyer.View full definition → contacts per client per quarter, measure both numbers, and treat a falling reply rate on stable volume as the alarm it is.

The second-order problem is ownership. Replies arrive on WhatsApp, on a phone the advisor carries. Annual turnover in store teams runs well into double digits at most groups, and when an advisor leaves, both the relationship and the behavioural history can walk out with her unless the exchange is logged against the client record.

3. Private-event attendance

Previews, trunk shows and atelier visits sit at the early stages the funnelfunnelThe customer journey from awareness to purchase, typically Awareness, Interest, Consideration, Decision, Action, with prospects narrowing at each stage.View full definition →-mapping lesson sets out, and they generate two measurable ratios.

• RSVP-to-attendance: of clients invited, who actually arrives. Figures cited by CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → vendors for high-net-worth (HNW) events (estimate, 2024-2025) sit around 35-50%, against single digits for mass retail promotions.
• Attendance-to-appointment: of attendees, who books a private follow-up within 60 days. Advisors watch this one hardest, because it separates consideration from a pleasant evening.

Worked example:

A maison invites 200 top-tier clients to a high jewellery exhibition.

• 90 attend (RSVP-to-attendance = 45%)
• 27 book a private appointment within 60 days (attendance-to-appointment = 30%)
• Historically, 40% of clients who book that follow-up buy within six months

Expected purchasers ≈ 27 × 0.40 = ~11 clients. At an €80,000 average order value, expected revenue ≈ €880,000, which is how event budgets running into six or seven figures get defended.

Now the edge case that ruins this arithmetic. With 27 clients, a 40% rate carries a standard error of about 9 points, so the true rate could plausibly be 22% or 58%: your €880,000 forecast is really a range from roughly €480,000 to €1.3m. Report it as a range, and pool several events before you change anything. Second, strip out guests invited for social reasons, press, stylists and friends of the house. They inflate attendance and never convert, and if they stay in the denominator every event looks like a disappointment.

4. Composite score with decay

Combine the signals into one number, with recency built in and declines counted against:

engagement_score = (
    4 * appointment_requested_last_90d +
    3 * appointment_completed_last_180d +
    2 * event_attended_last_180d +
    2 * advisor_reply_rate_last_180d +   # 0 to 1, scaled
    1 * event_attended_181_to_365d -
    3 * invitations_declined_last_180d -
    2 * no_show_last_180d
)

Illustrative logic, not a disclosed formula. Salesforce, which sells the CRM software a large share of these programmes run on, ships scoring models out of the box; the shipped weights mean nothing until they are refitted on your own conversion history, ideally a logistic regression over the last 24 months of outcomes per tier. An appointment eleven months old and one from three weeks ago are not the same evidence, and any score without decay drifts into a loyalty ranking.

Set the action threshold by capacity, not by probability. If your advisors can hold forty substantial conversations a quarter, the cutoff is the fortieth client on the list, whatever their score.

Checking the score against cost and against reality

Two tests decide whether the score earns its keep.

The first is lift over the base rate. If 12% of your top tier buys in any given six months regardless, and your top decile converts at 18%, you have built an expensive reordering of a list you already had. Recalibrate by decile every two quarters: predicted probability against realised purchases, decile by decile. Miscalibration usually shows up at the top, where a handful of very large clients drag the model.

The second is the self-fulfilling prophecy, and almost nobody runs it. Advisors see the score, spend their hours on the high scorers, and those clients buy. That proves the advisors followed instructions. Hold back a random 10% of high scorers from prioritised treatment for two cycles, or compare stores where the score is displayed against stores where it is not. It costs some sales. It is the only way to know whether you are predicting demand or manufacturing it.

Cost sits underneath both. An event seat runs from a few hundred to a few thousand euros per attendee once space, security and pieces in transit are counted, and that number has to clear the payback threshold the acquisition-cost lesson sets, measured against the lifetime valuelifetime valueLifetime Value: the total revenue (or profit) a customer generates throughout their entire relationship with your business.View full definition → figure the twice-a-decade lesson computes. For the general mechanics behind those two, see HBR on customer lifetime value.

Negative signals, and the client who hates events

Churn shows up in engagement data long before it shows up in transactions, which is the practical argument for scoring at all. Reply latency stretching from hours to days, two consecutive declined invitations, a quiet unsubscribe from the viewing list: each precedes the missing purchase by months. Because purchases are rare and messages are frequent, the negative signal is almost always the faster one.

The trap is treating every decline as disengagement. A meaningful slice of the highest-spending clients in hard luxury will not attend anything, will not be photographed, and buy twice a year through one advisor by appointment. Penalise them for declining and the model will rank your best clients as churn risks, after which the advisor stops calling and the prophecy completes itself. Carry an explicit contact-preference flag on the record and exclude those clients from event-based scoring entirely; score them on appointment cadence and reply behaviour alone.

Bain's long-running work (Bain-Altagamma Luxury Goods Worldwide Market Study) keeps showing a small group of repeat clients producing a share of revenue far out of proportion to their number. That concentration is exactly why a false negative in the score costs more than a false positive: an unnecessary invitation wastes a seat, a wrongly deprioritised top client can cost a year of purchases.

🎬 [VIDEO: "How Luxury Brands Build Customer Loyalty" - youtube.com - search for recent Bain, McKinsey, or business school channel explainers on luxury CRM and clienteling strategy]

Key Takeaways

• Three signals carry the six-month forecast: unprompted appointment requests, advisor reply behaviour, and private-event attendance converted into follow-up appointments. Traffic carries none of it.
• Decay and negative terms are what separate a predictive score from a loyalty ranking. Weight recent evidence higher, subtract declines and no-shows, and refit the weights on your own outcomes.
• Small numbers make confident forecasts dishonest: 40% conversion on 27 clients has a standard error near 9 points, so publish the range and pool events before acting.
• Validate with a holdout. Advisors working the high scorers will always make the score look right; a randomly withheld 10% is the only way to measure real lift over the base rate.
• Set the action threshold by advisor capacity, and exempt the contact-averse high spenders from event-based scoring, or the model will quietly demote your most valuable clients.