How do I measure influencer marketing ROI?

Attribution and incrementality answer different questions

Attribution tells you which sales touched a creator. Incrementality tells you which sales the creator caused. The gap between those two numbers is not small, and in the evidence below it runs in the same direction every time: the observational estimate comes in larger than the experimental one.

The cleanest demonstration comes from eBay's economists. Blake, Nosko and Tadelis ran a large field experiment on paid search and reported that standard methods produced "a ROI of over 4,100% without time and geographic controls, and a ROI of over 1,400% with such controls," while their experimental method on the same spend found "a ROI of -63%, with a 95% confidence interval of [-124%, -3%], rejecting the hypothesis that the channel yields positive returns at all." The spend and the underlying data were the same in both cases. Only the measurement method changed.

The same pattern shows up at scale. Gordon, Moakler and Zettelmeyer analyzed 663 large-scale experiments at Facebook and found median true lift of 5% for lower-funnel outcomes, against 24% and 64% from the two standard observational methods applied to the same campaigns. Gordon and colleagues had more than 5,000 user-level features available and still concluded: "despite having access to large-scale experiments and rich user-level data, we are unable to reliably estimate an ad campaign's causal effect."

Influencer content is particularly exposed to this bias, because creators reach audiences already inclined toward the category. A fitness creator's followers were going to buy protein powder from somebody.

What each measurement method can and cannot tell you

No single instrument answers the whole question. The useful move is knowing which direction each one is wrong in, so you can read the numbers with the bias built in rather than treating any one of them as truth.

Codes and links are attribution instruments. Holdouts, geo tests and mix models are incrementality instruments. Reporting an attribution number as ROI is the most common error in this category.

Instrument before launch, because most of this cannot be added later

The measurement decisions that matter are made before any content goes live. Meta's lift study documentation states the rule plainly: "Study start time must be in the future." No system lets you build a control group after a campaign has run, so a campaign that launches without one is stuck with correlational tools for its whole life.

Six things to have in place before the first post:

• A defined holdout or control geography, agreed and locked before launch. Meta's documentation also notes that once a study starts you cannot update its start time or treatment percentage.
• A clean pre-period baseline covering at least one full purchase cycle, captured before any creator content publishes.
• The post-purchase survey question live on your confirmation page before launch, so you have baseline answers to compare against. A survey that starts on launch day has nothing to compare to.
• One unique discount code per creator, at a value that exists nowhere else on your site or in email.
• An Amazon Attribution tag per creator if any traffic lands on Amazon. Amazon Ads describes it as "a free measurement solution" for professional sellers in Brand Registry, vendors, KDP authors, and agencies with clients selling on Amazon, and its reports cover a 14-day attribution window with new to brand among the conversion metrics.
• One named primary metric and a written measurement window, agreed before results start arriving rather than after.

How to run an incrementality test without buying one

Meta notes that "Conversion Lift Measurement is currently limited" and directs advertisers to contact a Meta representative for access, so most brands cannot simply switch it on. The practical alternative is a geographic test, and the tooling for it is free and open source.

Meta publishes GeoLift under an MIT license, described as an "end-to-end geo-experimental methodology based on Synthetic Control Methods used to measure the true incremental effect (Lift) of ad campaign." Google publishes CausalImpact under Apache 2.0 as "an R package for causal inference in time series." Both repositories were updated in 2026, and neither requires a vendor contract or user-level tracking.

A workable geo test needs enough markets to build a synthetic control, a pre-period long enough for the model to learn the baseline relationship between those markets, and treatment markets that do not spill heavily into control markets. Suppressing creator content geographically is the hard part, which is why paid amplification of creator posts makes clean holdouts easier than organic-only programs. Whitelisted content runs through ad delivery, and ad delivery can be geo-fenced.

Why your platform numbers will never match your store

Platform reporting and store data are built to disagree. Google Analytics 4 uses data-driven attribution by default, with a session conversion window where, per Google's documentation, "by default, it's 90 days." Meta applies different click and view-through windows, and a different credit model, to the inventory it sells. Two systems using different windows and different credit rules will never produce the same number from the same sales.

Browser behavior widens the gap. WebKit's engineering blog states that Safari's tracking prevention has capped the expiry of client-side cookies at seven days since February 2019, and that a matching seven-day limit was later extended to LocalStorage, IndexedDB and Service Worker registrations under the heading "7-Day Cap on All Script-Writeable Storage," clearing that data after seven days of Safari use without user interaction on the site. If your average purchase cycle runs longer than a week, link-based attribution built on JavaScript-set cookies is losing Safari buyers before they convert, and no amount of tag auditing recovers them.

Last-click reporting cuts against creator work specifically. Amazon's measurement researchers describe the problem this way: last-touch attribution "credits the full value of a conversion to one ad, usually the last one on the path to purchase, overlooking the impact of earlier touchpoints like awareness campaigns that build interest before shoppers are ready to buy." Influencer content sits mostly upper and mid funnel, so last-click undercounts it while discount codes over-count it.

Reconcile platform reports against store data once, early, to catch broken tags. Then pick one system as the record and stop relitigating the difference. Google's documentation notes that "more than 95% of key events get attributed within the first 14 days," which is a reasonable default window when the purchase cycle is short.

Engagement is not a proxy for sales

Engagement gets used as an ROI stand-in because it is easy to collect. The peer-reviewed evidence says it should not be.

Leung and colleagues, writing in the Journal of Marketing in 2022, estimated that "a 1% increase in influencer marketing spend increases engagement by .457%," and that the firms in their data set "could increase consumer engagement by 16.6% if they allocated their budgets proportional to these elasticities and base engagement levels, rather than their current allocations." The outcome variable in that study is engagement, not sales, so it does not tell you what influencer spend does to revenue.

A meta-analysis in the Journal of the Academy of Marketing Science, synthesizing 1,531 effect sizes from 251 papers, splits results into non-transactional outcomes such as attitude and purchase intention and transactional outcomes such as purchase behavior and sales. Different things drive each. For non-transactional outcomes it reports that "follower characteristics (social identity) have the strongest effects on consumer attitudes and behavioral engagement." For transactional outcomes it reports that "influencer characteristics (influencer communication) have the strongest effects on purchase behavior." A creator who maximizes one is not automatically maximizing the other. That is why an engagement number cannot stand in for a sales number.

The honest ceiling on what any measurement can resolve

Below a certain spend and sample size, single-campaign ROI is not measurable at useful precision, regardless of method or vendor.

The Gordon, Moakler and Zettelmeyer work is the clearest evidence of that ceiling. They had 663 randomized experiments and more than 5,000 user-level features to work with, which is more measurement infrastructure than most brands will ever assemble, and their conclusion was that they were "unable to reliably estimate an ad campaign's causal effect." A single creator flight measured with codes and links has a tiny fraction of that data behind it, so it will not resolve a precise return figure no matter how the report is formatted.

The practical response is to change the unit of measurement. Judge a portfolio of creator activity across quarters instead of demanding a verdict on each post. Codes and links are then operational tools, useful for deciding which creators to renew and which content to amplify. Incrementality tests belong to the bigger question of whether the channel earns its budget, which is worth answering a few times a year rather than continuously. Anyone promising a precise ROI figure on a small single-flight campaign is reporting attribution and calling it something else.

Numbers you will see quoted that do not hold up

Two figures dominate this topic, an "11x return" and "$6.50 for every $1 spent," and neither arrives with the things that would make it checkable. Neither is published with a sample size, a stated method, or a control group, and neither traces to a peer-reviewed study. They circulate as industry constants anyway.

These figures persist because the demand for them is real. The CMO Survey's 35th edition, fielded January 7 to 29, 2026 with 308 responding marketing leaders at U.S. for-profit companies, found that 58.8% of those answering the question reported increased CEO pressure to prove the value of marketing, and 55.5% reported the same from the CFO. A single memorable multiple travels well under that kind of pressure.

Ask three questions of any return multiple a vendor quotes you: what was the control group, what was the sample size, and who paid for the measurement. A number that cannot answer all three is a marketing asset rather than a result.

Frequently asked questions