Are AI Rewards Creating Repeat Customers—or Just More Discounts?

A busy redemption dashboard can hide an expensive habit: paying customers to make visits they would have made anyway. The better test is additional contribution, sustained return visits and trust.

Why this matters

A restaurant can send more personalised offers than ever and still fail to build a more valuable customer relationship. AI can identify likely visitors, choose a reward and time the message. None of that answers the harder question: did the offer create a visit the customer would not otherwise have made, or did it merely discount a visit that was already likely?

After reading this article, the reader will understand why a higher redemption count can still mean lower contribution, and be able to compare an offer group with a no-new-offer group on contribution per assigned customer, including the period after the offer ends.

Key points

A redemption is a record that a reward was used. It is not, by itself, evidence of an additional visit or of additional profit.
A useful comparison holds out customers who could have received the offer and measures contribution per assigned customer, not only among people who redeemed.
The same number of return visits can produce a gain, a loss or a larger gain, depending on the cost of the reward and whether the visit was already likely.
A reported wish for personalised rewards, and reported discomfort with the data used to create them, are different questions. Neither is a profit result, and neither describes Qatar or Gulf diners.

The reward problem

A QAR 20 reward may feel inexpensive beside a QAR 100 bill. Its effect depends on the customer’s alternatives. One customer might have returned at full price. Another might have visited only because the reward made the trip worthwhile. A third might have stopped visiting because of poor service; another coupon may do little until the underlying problem is addressed. The same redemption can therefore represent retained revenue, bought traffic or a temporary patch over an experience problem.

That distinction is the core of incremental measurement. The relevant comparison is not “reward visitors versus everyone else” after the fact. Customers who receive and redeem rewards often differ from customers who do not. A stronger test holds out a comparable group that could have received the offer but did not, then asks whether the offer group produced enough additional contribution to cover the reward and the campaign.

What the evidence supports—and what it does not

Restaurant technology provider Toast surveyed 500 UK diners and 400 restaurant managers through Pollfish on 29–31 July 2026. In its 21 September 2026 report, 23% of the surveyed diners selected personalised discounts and loyalty rewards when asked which AI-powered improvement they would most like to see more of. The two Toast pages do not rank that answer the same way against faster ordering, so this article does not treat 23% as the leading result. On a different question put to the same 500 diners, 29% said they were uncomfortable with restaurants using order history and visit frequency to personalise offers. [1] [2]

These are different questions. They should not be presented as proof that the same diners simultaneously demand personalisation and reject it. Toast’s 24 September reading of the same survey says the overlap is limited: among the 114 diners who most wanted personalised discounts and loyalty rewards, 62% said they were comfortable with that data use and 17% said they were uncomfortable. That cross-tab is the vendor’s interpretation of its own sample, not a separate study, and it still measures reported preference rather than additional profit.

The survey is also limited to reported UK preferences. It does not establish what diners in Qatar or the wider Gulf would accept, and it does not show that personalised rewards increase restaurant profit. The managers’ responses are likewise self-reported. They cannot tell a reader whether an AI offer recovered its cost.

A separate academic study examines a non-tiered loyalty programme at a men’s hair-salon chain. It is not a restaurant study and it is not evidence about AI. Its authors use observational data and a model-based approach, and they explicitly address the problem that customers who join loyalty programmes may already have been different from those who do not. Their estimated value came mainly through lower attrition, with little change in spending per visit. The result is useful as a warning against equating loyalty with a larger bill each visit, but its size should not be borrowed as a restaurant forecast. [3]

Three offers, one return count, three outcomes

The comparison below is a hypothetical illustration, not a forecast or a report of any restaurant’s results.

Assume a QAR 100 bill and QAR 40 of variable fulfilment cost. Before any reward, the visit contributes QAR 60. Across 100 eligible customers who receive no new offer, 20 return. Their total contribution is:

20 × QAR 60 = QAR 1,200

Now compare three offer designs sent to separate groups of 100 customers. In every case, 30 customers return and all 30 redeem. The campaign also costs QAR 100 to execute. The reward is subtracted once, inside the contribution of each redeemed visit. It is not subtracted again as a separate campaign cost.

Swipe to compare. These figures are hypothetical, not forecasts.

Offer	What the customer receives	Contribution per redeemed visit	Contribution from 30 returns	Compared with QAR 1,200	After QAR 100 campaign cost
A	QAR 10 discount	QAR 50	QAR 1,500	+QAR 300	+QAR 200
B	QAR 20 discount	QAR 40	QAR 1,200	QAR 0	−QAR 100
C	Relevant QAR 4 add-on, offered without a bill discount	QAR 56	QAR 1,680	+QAR 480	+QAR 380

All three campaigns produce the same headline result: 30 return visits and a 30% redemption rate. Their incremental contribution is nevertheless +QAR 200, −QAR 100 and +QAR 380. Offer B looks no worse than the others on a redemption dashboard, but it does not cover the comparison group once campaign execution is included.

Offer C is not automatically “better AI.” It wins in this illustration because the added item carries QAR 4 of variable cost and the visit still clears more than the no-offer case. A premium add-on with a high ingredient or labour cost could reverse that result. The design lesson is to price the reward against variable contribution, not to assume that a smaller-looking offer is a cheaper one.

Why equal redemptions can mislead

The illustration uses the same 30-customer response for every offer so the reward cost is easy to see. A real test would not assume that a QAR 4 add-on and a QAR 20 discount persuade the same people. The richer discount might attract more visits, or it might train regulars to wait for the next offer. The smaller offer might attract fewer visits but protect more contribution on each one. Only the held-out comparison can sort those effects.

Two further checks belong beside the table.

The first is timing. A campaign can pull forward visits that would have happened later in the month. The useful window is therefore longer than the redemption period. If treated customers visit more during the offer and less immediately afterward, the apparent lift may be borrowed demand rather than a new habit.

The second is mix. Suppose 20 of the 30 redeemers would have visited without the offer. The offer did not create those 20 visits; it reduced what they contributed. The other 10 are the incremental visitors. Offer A then gives up QAR 10 on each of the 20 visits already likely to happen, or QAR 200 in total, while the 10 additional visits contribute QAR 50 each. That is QAR 500 before campaign execution. The net result can still be positive, but it is no longer the QAR 300 gap suggested by comparing 30 redeemed visits with the original 20. The aggregate table remains a useful screening tool; a split between likely visitors and additional visitors explains where the contribution came from.

The same arithmetic can be reversed. If nearly all 30 redeemers were already going to come, Offer B gives away QAR 20 on visits the restaurant would otherwise have kept at QAR 60 of contribution. The dashboard still shows a successful redemption rate.

Sample size matters as well. A group of 100 customers is large enough to explain the arithmetic and too small to treat a ten-customer difference as a stable result. A restaurant test needs a pre-set comparison, a period long enough to include the aftermath, and a difference large enough to clear ordinary week-to-week variation.

What a useful rewards system should reveal

AI can make the targeting decision faster. The decision itself still needs a business definition of success. A useful system should be able to show:

who was eligible and who was held out;
the contribution per customer in each group, not merely redemptions;
the variable cost of the reward actually redeemed;
whether any lift survived after the offer expired;
which customers became less responsive unless a discount was present;
and which customers opted out, complained or reduced their visits after receiving personalised messages.

The last item is not a soft metric. The UK diner survey shows that a meaningful minority is uncomfortable with this kind of data use. A restaurant that increases redemptions while increasing opt-outs may be buying short-term visits with long-term permission.

What to watch next

Watch whether those six items move together. A rise in redemptions, on its own, does not answer the question.

The practical question is therefore not whether AI can predict a return visit. It is whether the restaurant has reserved an honest comparison group and is willing to stop an offer that buys visits below the contribution it needs. Personalisation may earn its place when it finds customers who value a relevant offer and would not otherwise have come. It becomes an expensive habit when it trains the customers most likely to return to wait for a discount.

This educational analysis does not recommend a particular offer, software product or customer treatment. The figures above are illustrations for comparing contribution, not a forecast of any restaurant’s profit.