Methodology · updated 2026-09-08
How we measure our numbers
Every number on this site is listed below with the source behind it, the sample it was measured on, the period it covers and how it was calculated. Where a figure comes from PEKO's own venue data we say so plainly. Where it comes from someone else, we link them. Where we could not substantiate a claim, we removed it from the site and recorded the removal here.
If a figure appears on a PEKO page and not on this page, treat it as an error and tell us — we will either source it or remove it.
By Peko Research Team · first published 2026-09-08
PEKO's own data
60–75% of first-time guests never return
- Source
- PEKO's own venue data. We previously published this as a single figure of 73%; that precision was not defensible across markets and venue types, so the site now states the measured range instead.
- Sample
- 1,200 F&B outlets on PEKO in Vietnam — independent venues and small groups of 1–10 outlets. Venues open under 6 months, and venues with fewer than 200 identified transactions in the window, are excluded.
- Date range
- January 2024 – June 2026, on a rolling 12-month basis per venue
- How measured
- For each venue we take guests whose first identified transaction falls in the window, then check whether the same resolved member identity (phone number or messaging ID) transacts again at that venue within the following 90 days. Guests who do not are counted as lost. Anonymous walk-ins are excluded from both sides of the calculation, because a receipt with no identity cannot be matched to a return visit.
- Limits
- The spread between venues is wide — venues that capture almost no member identities sit above the range, and venues with strong enrolment sit below it — which is why we publish a range and never a single average.
PEKO's own data
8–15 percentage point repeat-rate lift within 90 days
- Source
- PEKO's own venue data. This is the only platform-wide retention claim we make; individual merchant results shown in testimonials are attributed to that venue and are not platform averages.
- Sample
- Venues that ran PEKO continuously for at least 90 days with member enrolment and automated win-back both switched on, drawn from the same base of 1,200 Vietnamese F&B outlets.
- Date range
- Each venue's own first 90 days on PEKO, between January 2024 and June 2026
- How measured
- Repeat rate is the share of identified guests in a period who transact two or more times at the same venue. We compare a venue's repeat rate in the 90 days before PEKO to the 90 days after, and report the change in percentage points — not as a percentage increase, which would make the same movement look several times larger.
- Limits
- Venues that never switch on enrolment see little or no movement, because there is nobody to win back — which is why the claim is scoped to venues running both enrolment and win-back rather than to every account.
PEKO's own data
Measured across 1,200+ venues
- Source
- PEKO's own measurement base. It is deliberately smaller than our onboarding figure of 3,000 venues: measurement uses only venues with enough connected transaction history to compute a rate.
- Sample
- 1,200+ Vietnamese F&B outlets with connected transaction and membership data. All venues are in Vietnam, so any figure applied to Malaysia, Singapore or the Philippines is stated as modelled from Vietnamese data, never as locally measured.
- Date range
- January 2024 – June 2026
- How measured
- Anonymised, aggregated venue-level data. No guest-level record and no individual venue is ever published; splits are reported by market and venue type only.
- Limits
- A Vietnam-only base is a real limit on how far these numbers travel — the PEKO Southeast Asia F&B Retention Benchmark closes its first multi-market collection window on 31 December 2026 and publishes per-market figures, with the method fixed in advance.
PEKO's own data
The churn model uses 40 signals
- Source
- PEKO product documentation — the live feature set of the lapse-prediction model.
- Sample
- Model performance is measured on the same 1,200+ outlet base: 68–74% precision and 60–66% recall on flagging a guest who then fails to return.
- Date range
- Model version in production from March 2026; performance measured over the following 90 days
- How measured
- 40 is the count of input features the model scores per member per day. They fall into six groups: recency, visit frequency and personal visit cycle, spend and basket composition, favourite items, payment and ordering channel, and messaging engagement. Precision is the share of flagged members who did in fact lapse; recall is the share of members who lapsed that the model had flagged. A member needs at least 14 days of history before the model scores them.
- Limits
- Feature counts are a description of the model, not evidence that it works — the precision and recall figures above are the part worth checking, and they are published with the base they were measured on.
External source
Delivery platforms take 28–30% per order
- Source
- The delivery platforms' own published merchant commission rates for standard delivery packages in Southeast Asia. We cite the platforms directly rather than a secondary report.
- Sample
- Publicly listed merchant rates for full-service delivery packages in Vietnam, Malaysia, Singapore and the Philippines, read from the platforms' merchant pages.
- Date range
- Rates as listed in August 2026
- How measured
- We quote the commission band for standard delivery packages before value-added tax and before any promotional co-funding the venue agrees to. Self-pickup and lower-tier packages are cheaper and are not covered by this figure.
- Limits
- Rates are negotiated per merchant and change without notice, so treat 28–30% as the published band for standard delivery rather than the rate on your own contract — check your latest statement, and tell us if the band has moved.
Retired claim
40% direct orders within 3 months — RETIRED, removed from the site
- Source
- None we could stand behind. The figure circulated in our own marketing copy without a cohort, a window or a definition of what counted as a direct order.
- Sample
- No qualifying cohort existed.
- Date range
- Claim removed September 2026
- How measured
- Not measurable as written. Channel mix depends on how much delivery volume a venue starts with, and we had no like-for-like base to compare against.
- Limits
- Rather than restate it more vaguely, we deleted the claim from every page — it will only return with a named cohort, a window and a stated definition of a direct order.
Questions about these figures
- Are these numbers independently audited?
- No. They are our own measurements, published with the sample, window and method so they can be challenged — and where a figure is not ours, it links to the party that published it.
- Why do you publish ranges instead of single figures?
- Because the spread between venues is wide enough that a single average would be the least useful number we could give you. A range tells you what to expect; an average hides which end you are likely to land on.
- Can I cite these figures?
- Yes, with attribution to PEKO and a link to this page. If you need the split behind a figure, ask and we will send it.
- What happens when a figure changes?
- We update it here first, with a new dateModified, then on the pages that quote it. If a figure loses its basis entirely, it is removed from the site and recorded on this page as retired.