The guest asks where to eat on Thursday. You send the wine bar two streets over, the small one, with the short list and the owner who will talk to them about it. They reply with a thumbs up. Then nothing. Whether they went, whether they liked it, whether they spent thirty euros or three hundred, you have no idea. Neither does the wine bar.
Six months later you are sitting with that owner talking about a referral arrangement, and the conversation stalls on the only question that matters. How many people have you actually sent him? He thinks a few. You think rather more than a few. Neither of you can produce a number, so whatever you agree gets priced on goodwill and mutual optimism.
That is the whole problem in one scene. There is no join key between a message on a phone and a person walking through a door. Everything below is an attempt to manufacture one, and every attempt leaks.
This article covers the four mechanisms that produce a usable signal, in ascending order of reliability and friction; what each one genuinely measures and where it undercounts; and why conversion rate is the wrong headline number even once you have one.
Why This Attribution Problem Is Genuinely Hard
Online, a click and a purchase happen in the same browser and something stitches them together. Here the recommendation lands on a phone and the transaction is a person walking into a building and paying with a card you will never see. Nothing in that chain is instrumented, and most of it never will be.
Three specific breaks do the damage. The gap between suggestion and action can be days, so any window you define is arbitrary. The guest often acts on the name rather than the link: they read "Taberna do Mar", type it into a maps app, and walk, which is a successful recommendation that leaves no trace in your system. And the party is usually a group: one person reads, one person decides, four people eat.
There is also a deeper problem that no counting method touches. Even a perfect count of arrivals would not tell you whether you caused them. That one comes back later, and it is the reason conversion rate is a worse headline number than it looks.
Mechanism One: Ask the Business
The cheapest method is a conversation. "Have you had many of my guests in this month?"
What it measures is the owner's memory, filtered through their interest in the relationship continuing. That is not nothing, but it is not a count. It undercounts for ordinary reasons (staff turnover, the busy Saturday when nobody had time to notice, the guest who never said where they heard of the place) and it overcounts for equally ordinary ones, because anyone arriving with a suitcase gets credited to you and because an owner who wants the arrangement to continue rounds up.
Its real value is one binary signal, and that signal is reliable in only one direction. If the owner has no recollection of your guests at all, believe it. Absence of memory is much better evidence than any number the same conversation produces. A "handful, I think" tells you almost nothing; a blank look tells you the channel is not working.
It also does the thing a spreadsheet cannot, which is keep the relationship warm. If the destination of all this is a formal arrangement, the conversation is doing double duty: the shape those agreements take is covered in structuring a local business partnership agreement.
Mechanism Two: A Mention or a Code
The next step up is asking the guest to identify themselves: "Tell them you are staying at Casa Verde and they will bring you a glass of the house white."
What this measures is guests who remembered, were willing to say it, and were served by someone who logged it. Three filters in series, each lossy. It undercounts heavily: the guest who forgets in the moment, the one who feels awkward claiming a perk, the party of four that mentions it once, the waiter mid-service who means to write it down and does not.
It also changes what you are measuring. The free glass drives some of the visits it counts, so the number describes "your recommendation plus an incentive", which is a different product from your recommendation. Drop the perk and use a plain greeting instead ("say Ana sent you") and the distortion goes away along with most of the count, because now there is no reason for anyone to mention it. Pick which error you would rather have; you cannot avoid both.
The operational reality is that this method needs someone on the business side to own it. A notebook by the till, or a button on the till system, and a person who cares. Without an owner it decays within weeks, and you are back to mechanism one without noticing. A count you collect in person once a month will outlast an elaborate system nobody runs.
Mechanism Three: A Link You Control
Send every recommendation with a link that passes through something you own: a page per business in your guidebook, or a short link you can see the traffic on.
This is the first mechanism that produces a number you hold yourself, can break down by property and by moment in the stay, and can compare over time. What it measures is intent: somebody tapped. It undercounts the guest who reads the name and searches for it directly, which is a large group precisely because a good recommendation is memorable. It misses the guest who shows the phone to a partner who then books, and the group where one tap represents six people. It overcounts browse taps, accidental taps, and the guest who looks and goes elsewhere.
The value is not accuracy. The value is that it is wrong in a consistent direction, which makes movement meaningful even when the level is not. If taps on the same suggestion go from six in a quarter to nineteen after you start sending it on arrival evening instead of waiting to be asked, that is real information about timing, and the timing question is worth more than the level, which is the argument in sending proactive recommendations at the right moment.
One friction cost worth naming: a bare tracked URL in a message reads commercial, and guests can feel it. The link has to lead somewhere genuinely useful (the address, the hours, what to order, why you like it) or you have traded a small amount of trust for a tap you did not need.
Mechanism Four: Ask the Guest Afterwards
The last mechanism is the only one that can see a non-conversion and tell you why. "Did you make it to any of the places we suggested?" followed by "if not, where did you end up?"
The second question carries most of the information. The answer is usually one of a few things: they went to the place with two thousand reviews instead, they ate at the flat, they were too tired, or they never saw the message. Each of those points at a different fix, and none of them shows up in a tap count.
The costs are real. Post-stay response is low and skews to the very happy and the very annoyed, so the sample is not the guest body. Recall at day six is poor, and a guest genuinely may not remember which suggestion came from you and which from a maps app. And every question you ask at checkout competes with the review request, which is worth more to your business than your attribution data, so fold this into the feedback you are already collecting rather than adding a separate ask, as in collecting guest feedback before checkout.
The better version is in-stay and immediate. "How was Taberna last night?" the next morning is a sharper question than a survey on day seven, and it reads as hospitality rather than research.
The Four Side by Side
| Mechanism | What it really measures | Main error | Friction | Best used for |
|---|---|---|---|---|
| Ask the business | The owner's memory and goodwill | Wrong in both directions, unbounded | None | A reality check, not a count |
| Mention or code | Guests who remembered and asked | Undercounts badly; the perk distorts | Someone on their side must log it | Establishing a floor |
| A link you control | Intent to look | Misses everyone who types the name | A link in every message | Trends and comparisons |
| Ask the guest | Self-reported behaviour and reasons | Low response, skewed sample, weak recall | Competes with the review ask | Understanding non-conversion |
The stack worth running is two, not four: a link you control for the trend line, and a conversation with the business every couple of months as a sanity check on it. Adding the other two buys precision you cannot use and friction the guest can feel.
Why Conversion Rate Is the Wrong Headline Number
Take an illustrative quarter at one property: forty stays, twenty-five guests sent a dinner suggestion, nine taps on the link, and an owner who remembers "a handful". The headline writes itself as thirty-six percent. Those figures are invented to show the shape of the arithmetic, not measured from anything.
The arithmetic does not survive contact. Nine taps is not nine dinners: some tapped and did not go, some went without tapping. The true figure could plausibly sit anywhere from five to fifteen, and nothing in the method distinguishes those. A single percentage with that much play around it is a decoration on a report, not a measurement.
Two problems run deeper than accuracy, and they would remain even with a perfect count.
A recommendation nobody acts on may still have improved the stay. The guest who read three sentences about the Thursday market and never went still learned that their host knows the neighbourhood. That lands in the review, not at the till. A metric that scores it zero is measuring the wrong thing.
A high-converting recommendation may be converting guests who were going anyway. The most famous restaurant on the square will show an excellent tap rate and you contributed nothing: the guest would have found it in ninety seconds. What you actually want is the incremental recommendation, the place they would never have found, and those convert worse by definition while being worth considerably more. That is the same distinction as a static list versus a recommendation shaped to the moment, covered in static guidebooks versus dynamic guest recommendations.
Rank your library by conversion rate for a year and it will quietly become a list of the obvious and the discounted, which is precisely the guidebook the guest could have assembled from their phone for free.
Four things are worth watching instead, and three of them you fully control. Coverage: what share of stays received any recommendation at all, which is where most of the lost value actually sits: a mediocre library that fires on thirty stays out of forty beats an excellent one that fires on twelve. Timing: how many suggestions arrived before the guest asked. Library health: how many entries you would still personally vouch for, and how many describe a place that has closed or changed hands. And the owner's enthusiasm at month six, which is softer than any number and a better predictor of whether the arrangement survives.
Honest Limits
Attribution here is directional at best. Every mechanism above carries an error term you cannot size, and the errors do not cancel. You are not building a measurement system; you are building a signal with a known bias, which is a smaller and more honest thing.
Treat every number as a trend line, not an audit. The question a link can answer is "more than last quarter, and after which change?" It cannot answer "how many people ate there because of me", and no amount of instrumentation on your side will make it.
Be suspicious of precise conversion figures from anyone, including any tool that offers them. A dashboard reporting thirty-four percent conversion on offline visits is reporting taps or claims with a more confident label on top. Two questions settle it: what is the numerator event, and what is the denominator. If the numerator is a click, then the number is a click-through rate, whatever the column header says.
Do not renegotiate money on a number you cannot defend. If you cannot walk the business owner through how the count was produced, and show them where it under- and overcounts, do not build a payment on it. The wider commercial picture, including how these arrangements are structured and disclosed, sits in earning referral income from local recommendations.
Measurement has a floor cost and a low ceiling of usefulness at small scale. For a host with four properties, the honest answer may be to skip all of this, keep a genuinely good library, and judge it by whether guests name specific places in their reviews.
And about Welco specifically: it is a delivery layer, not a measurement product. It can put your recommendation in front of a verified guest in their own language at the moment it is relevant, and it can carry whatever link you choose to put in it. It does not watch the guest walk through the restaurant door, it does not connect to anybody's till, and it will not hand you a verified conversion figure. Nobody can, for an offline visit, and a vendor who says otherwise is describing something the mechanism does not support.
Where Welco Fits
Welco is a WhatsApp AI assistant for vacation rental hosts. Its job in this picture is the delivery half: answering a verified guest from the recommendations you curated, in any of 30+ languages, and generating the guidebook as a web link and a PDF so the suggestion has somewhere useful to land. The measurement stays yours, and it stays approximate.
Access today is by demo request. If you are trying to work out whether your local knowledge is reaching guests at all, that coverage question is the one worth answering first, and it is answerable.