Multi-Touch Attribution for Hotels: Crediting the Booking Correctly

Last-click attribution hands the whole booking to whichever channel happened to be standing closest when the guest finally hit "reserve," and that channel is almost always brand search or metasearch. This is the owner's guide to fixing it: which attribution model to run, how to join GA4, the booking engine, the PMS, and the CRM into one journey, how to run an incrementality test that settles the argument, and how to move budget on the result without blowing up the direct channel.

The booking that four channels claim and one gets paid for

A guest in Chicago sees a friend's photos from your property on Instagram in March. In April she reads a travel article that mentions you, opens the OTA app to check prices, and closes it. In May she gets an email from your list because she signed up for the guide two years ago. In June she searches your hotel's name on Google, clicks the metasearch ad because it was the first thing on the page, and books direct for a September stay. Your marketing report credits the booking to Google Hotel Ads. The Instagram spend, the PR retainer, the content program, and the email platform each produced nothing, according to the only report ownership ever sees.

That is last-click attribution, and it is still the default operating model at most independent hotels because it is the default in the booking engine, in the agency's dashboard, and in the mental model of everyone who grew up on it. The problem is not that it is imprecise. The problem is that it is precisely, consistently wrong in the same direction. It overpays the bottom of the funnel, where the guest was already going to book, and it underpays everything that put the hotel in the consideration set in the first place. A hotel that runs its budget on last-click will, year after year, shift money into brand search and metasearch, starve upper-funnel programs, and then wonder why the same campaigns cost more each year to produce the same bookings.

The scale of what last-click ignores is well documented. Expedia Group's Path to Purchase research with Luth Research found that travelers across seven countries view 141 pages of travel content in the 45 days before booking, and US travelers view 277. They spend an average of 303 minutes with that content, and the activity is not evenly spread: roughly 2.5 page views a day early in planning, rising to 25 page views on the day of purchase. The same study found that OTAs (80%), search engines (61%), social media (58%), airline sites (54%), and metasearch (51%) are all used by a majority of travelers in the window. Last-click sees exactly one of those touches. In the 24 hours before booking it usually sees the least interesting one.

The industry knows this and mostly has not acted on it. Marketing LTB's 2026 compilation puts the share of marketers still relying on last-touch at 41%, and the share of organizations relying on it exclusively at 22%. Gartner's 2025 UK survey, cited in the same roundup, found only 24% of B2B organizations using multi-touch at all. Hotels are not better than the average; they are worse, because the booking engine, the PMS, and the CRM were each built as separate systems with no shared identity, and the attribution problem is fundamentally a data join problem before it is a modeling problem.

Last-click does not measure which channel earned the booking. It measures which channel was open in the browser when the guest stopped researching. Those are different questions, and the budget only cares about the first one.

What last-click actually costs

Put a number on it. A 200-key independent resort spending $600,000 a year on digital marketing will typically show, under last-click, that 55% to 70% of attributed direct bookings came from brand search and metasearch. The agency, quite rationally, recommends putting more money there. But two of the most cited experiments in marketing science say a large share of those clicks were never incremental. The eBay field experiments by Blake, Nosko, and Tadelis, summarized by Recast, turned off branded paid search across markets and found that almost all of the "lost" paid clicks and attributed sales were immediately recaptured by organic. The Edmunds replication found the same substitution at a smaller magnitude: less than half the paid traffic came back organically. Google's own analysis of 390 paused-campaign studies, covered by Search Engine Land, found that ad click incrementality is about 50% when the advertiser already ranks first organically, 82% at organic positions two through four, and 96% at position five and below. Your hotel ranks first for its own name. Half of your brand-search bookings would have arrived anyway.

Now consider the other side of the ledger. Hotel marketing practitioners quoted by Atticus Li treat roughly 40% pass-through on metasearch as tolerable overlap with organic and direct, and treat anything past 50% as a signal of overspend. If 40% to 50% of metasearch bookings would have booked direct without the ad, and metasearch is being credited with 100% of them, the reported cost per acquisition is understated by nearly half and every reallocation decision built on it is wrong. Meanwhile the content, email, and social programs that are underreported get cut first in a soft quarter because they show no bookings, which further concentrates spend at the bottom of the funnel, which makes the metasearch numbers look even better. The loop is self-reinforcing and it runs in the wrong direction.

The commercial stakes are large because direct is the most profitable channel in the building. Kalibri Labs puts direct booking margins 9% to 20% above indirect channels after marketing and technology costs, and D-EDGE's 2025 Direct Distribution Report shows an average direct distribution cost of 3.5% against the 12% to 28% typically charged by OTAs. The 2025 HEDNA, NYU Tisch, and RateGain benchmark found direct matching OTA share at about 21% of bookings for the first time, but for independents specifically OTAs still take roughly 63%. Every dollar of direct marketing that is misallocated because of bad attribution is a dollar not spent moving a booking off a 15% commission. The organizations that fix attribution see it in the margin: Dataslayer's model comparison reports 14% to 36% improvements in cost per acquisition after moving to multi-touch, and the travel-specific case studies are more dramatic still, mostly because travel journeys are so long that last-click was so wrong to begin with.

Choosing the model: what each one overpays and underpays

Every attribution model is a rule for dividing one booking's credit among the touches that preceded it. There is no neutral model. Each one has a bias, and the job is to choose the bias you can live with, understand it, and correct for it with a test. The table below lays out the six models a hotel will encounter in GA4, in Google Ads, and in the booking engine's own reporting, and what each one systematically gets wrong.

Two of these deserve comment. Position based, which GA4 calls "position-based" and older Google Ads called "U-shaped," is the right rules-based default for most hotels because the two touches it rewards, first and last, are the two you can actually verify. The first touch is where consideration began; the last touch is where the guest converted. Everything in between gets a share without any single middle touch being able to dominate. It is transparent enough to explain to an owner in one sentence, which matters more than most analysts admit.

Data-driven attribution is now the default in GA4 and in Google Ads, and it is better than any rules-based model when it has enough data, because it looks at both converting and non-converting paths and asks what each touch actually changed. Google has lowered the entry thresholds over the last two years; the old requirement of 3,000 ad interactions and 300 conversions in 30 days has been relaxed, and the current recommended floor is around 2,000 interactions and 200 conversions in a 30-day window, below which the model quietly reverts toward simpler logic. A 200-key hotel producing 150 to 400 direct bookings a month is right on that line. The bigger issue is not volume. It is that Google's model only sees what Google tracks, weights its own inventory well, and has no view of the OTA search that never clicked through, the phone call to reservations, or the group block that came from a sales call. A data-driven model that is fed a partial journey produces a confident answer about a partial journey. That is why the data join comes before the model choice.

The channel credit shift: what changes when you switch

Abstract model comparisons do not move budgets. A worked example does. Below is a representative 90-day window for a 200-key independent resort with roughly $150,000 of quarterly digital spend and 1,000 direct bookings, run through three models on the same booking data. The numbers are illustrative but the pattern is the one we see almost every time a hotel does this for the first time: brand search and metasearch give up 25% to 40% of their last-click credit, and the upper and middle funnel absorb it.

Read the last column against the marketing budget. In this example brand paid search was consuming about 22% of spend and claiming 31% of bookings under last-click, which made it look like the most efficient channel in the mix. Under data-driven attribution it claims 19%, which makes it roughly average. Paid social was consuming 18% of spend and claiming 5% of bookings, which is why it was on the list to be cut; under data-driven it claims 14%, which puts it close to cost-neutral and, given that it is the channel most responsible for bringing new guests into the consideration set, arguably the one to protect. Email, which costs almost nothing per booking, nearly doubles its credit. None of these numbers are the truth. They are a less wrong version of the truth than last-click, and the point of the next section is to find out how wrong they still are.

The data join: GA4, booking engine, PMS, and CRM

The reason hotels run last-click is not that they prefer it. It is that last-click is the only model that works when the systems do not talk to each other. GA4 knows the sessions and the sources. The booking engine knows the reservation and, if it is instrumented, the source of the final session. The PMS knows the stay, the revenue, the cancellation, and the guest profile. The CRM knows the email history and, sometimes, the loyalty identity. Multi-touch attribution requires all four to agree on who the guest is, and they were not designed to.

The table below is the join map we build with clients. It is less glamorous than the modeling and it is where 80% of the work sits.

Three of these failures deserve a closer look because they are nearly universal. First, the cross-domain break. Most independent hotels run the marketing site on one domain and the booking engine on the vendor's domain. Unless cross-domain measurement is configured correctly, GA4 sees the hand-off to the booking engine as a new session with the referral source "yourhotel.com," and every booking is attributed to your own website. Hotels that discover this usually discover it after a year of reporting that told them nothing. The fix is a configuration change, not a project, and it should be the first thing verified.

Second, cancellations. Attribution runs on the booking event, and the PMS is the only system that knows whether the guest ever arrived. Channels differ sharply in cancellation behavior: flexible-rate metasearch bookings in a competitive market can cancel at 30% to 40%, while a loyalty email offer to past guests may cancel at 8%. If the attribution model never learns about the cancellation, it credits metasearch with bookings that produced no revenue and undercredits the email that produced real stays. Feed realized revenue from the PMS back into the attribution table, ideally at checkout, and rerun the credit on stayed revenue rather than booked revenue. This single change reorders the channel ranking at most hotels.

Third, the privacy window. Safari's Intelligent Tracking Prevention limits JavaScript-set first-party cookies to seven days, and to 24 hours when the landing URL carries a click identifier like gclid. For a hotel with a 45-day research window and a 30-plus day booking window, that means a Safari user who clicks an ad, researches for two weeks, and returns to book is counted as a brand-new direct visitor. Given that Safari carries roughly half of US mobile traffic and a disproportionate share of affluent leisure travelers, this alone erases a large share of upper-funnel credit before any model runs. The mitigations are server-side tagging so cookies are set by your own server rather than JavaScript, first-party identity via the email captured at booking, and enhanced conversions that match hashed emails back to ad platforms. They are worth doing. They are also the reason a click-based model, however good, will always undercount, and why the incrementality test is not optional.

The attribution model is a map drawn from the roads your tracking can see. Incrementality testing is driving out to check whether the roads exist. Owners should fund the second before they trust the first.

Incrementality: the test that ends the argument

Every attribution model, including data-driven, is correlational. It tells you which touches appeared in converting paths more than in non-converting ones. It cannot tell you what would have happened if a touch had not been there. Incrementality testing does exactly that, by withholding a channel from a comparable group and measuring the difference in bookings. It is the only method that answers the question ownership actually asks, which is: if we cut this budget, what do we lose?

Supermetrics' guide lays out the general methods. The table below adapts them to what a single hotel or a small portfolio can practically run. The pattern that matters most is the geo holdout, because hotels have a natural geographic structure in their feeder markets that most e-commerce businesses lack.

A few practical rules. Run tests in shoulder season, not in the six weeks that pay for the year, because the cost of a wrong test is low then and the noise from compression nights is lower. Match test and control markets on historical booking volume, lead time, and rate, not just on population. Measure the outcome in the PMS, on stayed revenue, not in the ad platform. And measure the right outcome: a metasearch test should track total direct bookings from the test markets plus OTA bookings from those markets, because the cost of turning off metasearch is not only lost direct bookings, it is bookings that shift to the OTA at a 15% commission. A test that shows metasearch is 55% incremental to direct but that OTA bookings rose to absorb the rest is a test that says keep metasearch and negotiate the bid.

Metasearch vendors have started to run this analysis themselves, which is worth taking with appropriate skepticism but not dismissing. Tripadvisor's published test-and-control studies for hotel brands are structured the right way, with matched groups and lift in clicks and bookings as the outcome. Ask your metasearch integrator whether they will run a true holdout on your account. The ones that will are confident in the channel. The ones that redirect you to the ROAS dashboard are telling you something.

The broader industry is moving toward blending these methods rather than choosing between them. eMarketer data cited by TapClicks now shows 27.6% of US marketers rating marketing mix modeling as their most reliable measurement method, ahead of multi-touch at 19.4%, a reversal from three years ago that tracks directly with the loss of cookie-based identity. Ruler Analytics makes the practical case that the two answer different questions: multi-touch tells you which tactic to optimize this week, mix modeling tells you which channel to fund this year. For a single hotel, a full mix model is usually more than the data can support until there are two or three years of clean weekly history. For a portfolio of eight or more properties it is the right tool and the one that finally captures the OTA billboard effect, which no click-based model will ever see. Recent academic work on cannibalization-corrected attribution is converging on the same conclusion: attributed and incremental are different quantities, and the gap between them is largest exactly where hotels spend the most.

Reallocating on the result

The point of all of this is a budget decision, and the budget decision has to be made carefully because the direct channel is fragile in a way that the attribution report does not show. Cutting brand search to zero because a test showed it was 50% incremental is a mistake if the 50% that was incremental was being intercepted by an OTA bidding on your name. The reallocation logic below is the one we use, and it is deliberately incremental: move 15% to 25% of a channel's budget per quarter, measure, then move again.

The benchmark for what "good" looks like at the bottom of the funnel is well established. RoomMaster's analysis puts Google Hotel Ads at 8% to 14% cost per acquisition against OTA commissions of 15% to 25%, and Foundry CRO's 2026 benchmarks put a well-run independent hotel's metasearch ROAS at 8 to 1 or better. Those are attributed figures, not incremental ones, so apply your own pass-through rate to them before comparing to the OTA. A metasearch program at 10% attributed CPA and 55% incrementality is really running at about 18% incremental CPA, which is roughly an OTA commission with more work. That does not mean stop; it means the channel is a parity defense, not a growth engine, and should be budgeted as one.

A 90-day implementation plan

This does not require a new platform. It requires the systems you already own to be configured correctly, one honest test, and a report ownership can read. Here is how we sequence it.

Weeks 1 to 3: fix the plumbing. Verify cross-domain measurement between the website and the booking engine. Confirm the booking engine's purchase event carries the GA client ID, the confirmation number, and the booking value. Standardize UTM conventions across the agency, the email platform, the metasearch integrator, and any PR or content partners; the most common finding at this stage is that email traffic has been reporting as "referral" for years. Turn on enhanced conversions in Google Ads and the equivalent hashed-email matching in Meta. If the site runs on Safari-heavy traffic, scope server-side tagging now; it can be deployed inside the 90 days but it needs to be scoped in week one.

Weeks 3 to 6: build the join and the baseline. Create a single attribution table, in a warehouse or even a well-structured spreadsheet at first, with one row per booking: confirmation number, booking date, arrival date, channel path from GA4, booking value, and, once the stay passes, realized revenue and cancellation status from the PMS. Run the last 90 days through last-click, position-based, and GA4's data-driven model and put the three columns side by side, as in the credit shift table above. This is the baseline. Nobody reallocates yet.

Weeks 6 to 12: run one test. Pick the channel where the credit shift was largest and the spend is meaningful, which is nearly always brand search or metasearch. Design a geo holdout with matched control markets, run it for four to six weeks outside peak, and measure stayed direct revenue plus OTA revenue from the test markets against control. Document the incrementality rate. This number, applied to the attributed CPA, is the first true cost per acquisition the hotel has ever had for that channel.

Week 12: the reallocation memo. One page. The three-model comparison, the test result, the true CPA by channel, and a proposed 15% to 25% shift for the next quarter with the guardrail metric that will trigger a reversal. Present it as the first quarter of a rolling process, not as a verdict, because the second test will refine the first and the model will get better as the join gets cleaner. Hotels that are beginning this work usually find the hardest part is not the analytics but agreeing on what the scorecard should track; a structured review of the reporting stack, the tracking configuration, and the channel benchmarks is where most of them start - explore our AI & Technology Scorecard and Reporting service →.

Where AI actually helps

Most of the work above is configuration and discipline, and it would be dishonest to dress it up as artificial intelligence. There are three places where machine learning changes the outcome. The first is the data-driven model itself, which is a genuine improvement over rules-based models once the journey data is complete enough to feed it. The second is identity resolution: probabilistic matching of the guest across the GA client ID, the booking email, the PMS profile, and the loyalty record is a well-solved machine learning problem and it is the piece that turns four partial journeys into one. The third is the mix model, which has moved from a six-figure consulting engagement to open-source tooling in the last three years and which a portfolio's revenue analyst can now run on a laptop with the right guidance.

What AI does not do is substitute for the incrementality test. A model, however sophisticated, that has never seen a world where the channel was turned off cannot tell you what turning it off would do. The hotels getting this right run a light rules-based model for weekly optimization, a data-driven model for monthly reporting, one holdout test per quarter for truth, and, once the history supports it, a mix model annually for the budget. Each layer checks the one below it. The attribution report ownership sees at the end is the same one-page scorecard every quarter, with the credit shift, the incrementality rates, and the true CPA by channel, and it gets more trustworthy each time it is produced.

Last-click will not disappear from the booking engine dashboard, and it does not need to. It is a fine answer to the question of which channel closed the booking. It is a wrong answer to the question of which channels earned it, and for twenty years hotels have been funding the first answer while asking the second question. The fix is not expensive. It is a data join, a test, and the willingness to move budget on what the test says.

Frequently Asked Questions