A measurement plan starts with what the activity is for, not with what the platforms can count. Hoffman and Fodor (2010) argued that social media return is usually measured the wrong way round: brands count what they spend and what platforms report, when they should start from why customers use social media and measure the investments customers make in the brand (attention, conversation, content, advocacy) before tracing those to sales or cost savings. The point survives: the metrics a dashboard offers by default are rarely the ones an objective needs.

The communication measurement profession has codified the same discipline. The Barcelona Principles, from AMEC, the International Association for the Measurement and Evaluation of Communication, updated as version 4.0 in June 2025, open with setting clear and measurable objectives, and call for reporting outputs, outcomes and impact, using qualitative as well as quantitative analysis across all relevant channels (AMEC, 2025). AMEC's Integrated Evaluation Framework, launched in 2016, turns this into stages from objectives: inputs (audience, plan, budget), activities (what was done), outputs (what audiences were exposed to: posts published, reach), outtakes (how they responded: attention, engagement, clicks), outcomes (what changed in them: understanding, preference, intent, trial) and impact (what changed for the organisation: sales, customers, reputation). Most social reports stop at outputs and outtakes; budget decisions turn on outcomes and impact.

Each stage needs a metric, a data source, a baseline and a reading date. Outputs and outtakes are leading indicators: available within days and useful for managing a campaign in flight. Outcomes and impact are lagging indicators: slower, noisier and harder to attribute, but they are what the objective is about. Each target waits for a baseline, as Module 1 insisted.

Time horizons matter too. Binet and Field (2013) distinguish short-term sales activation, whose effects show within weeks, from long-term brand-building, whose effects accumulate over months and years. A six-week campaign report can see the first clearly and the second barely at all, so a framework should say which effects it will read when, and which it cannot see.

For Moorfield Botanics, Amara returns to the objectives set at her platform audit: new customers whose first recorded touchpoint is TikTok or Instagram; reach among people outside the customer base; and growth in Instagram saves, shares and replies. For the Strong Skin Season launch she maps creator posts live and reach as outputs; saves, shares and link clicks as outtakes; branded search and visits from new visitors as outcomes; and first orders, new customers and their contribution to the FY2027 goal of 60,000 new customers as impact, with repeat purchase read later.

Most social metrics are easy to collect, hard to interpret and narrower than their names suggest.

Reach counts the unique accounts that saw content; impressions count every view, so one person seeing a post three times is one reach and three impressions. Both are platform estimates.

Engagement rate has no single definition. It can be engagements divided by followers, by reach or by impressions, and "engagements" may include likes, comments, shares, saves, clicks or video views depending on the tool, so one post can show three different rates. Rate by followers flatters content shown mainly to followers and misreads content that travels beyond them in recommendation-led feeds; rate by reach measures how the people who saw it responded, usually the more useful question. This module uses engagements (likes, comments, shares and saves) divided by reach; any report should state its definition. Even then, engagement is an outtake, not an outcome: a like costs the user nothing, and saves and shares signal more.

Sentiment analysis classifies mentions as positive, negative or neutral, usually automatically. As Module 2 showed, automated classification varies widely in accuracy and struggles with sarcasm and slang (Hartmann et al., 2023), so trends are more trustworthy than single scores, and samples should be read by hand. Share of voice, a brand's share of category conversation, is useful as a relative measure over time but depends entirely on which competitors and keywords are tracked. Follower growth is a vanity metric when followers are not the objective: in recommendation-led feeds following is a weak predictor of who sees content, and followers can be bought.

Earned media value (EMV) is the most seductive figure of all. Vendors price engagement or impressions at what equivalent advertising might have cost, then add it up. EMV descends from the advertising value equivalent (AVE), long used to value press coverage, and the Barcelona Principles 4.0 (AMEC, 2025) exclude AVEs as an invalid measure. EMV formulas differ by vendor and are rarely disclosed, the rates are assumptions, and the result measures neither sales nor profit nor what would have happened without the activity.

Finally, ask who produced each figure. Platform-reported results come from the company selling the media, under its own attribution rules; agency reports come from the company paid to run the activity. Neither is necessarily wrong, but both are marking their own homework. Independent evidence (the brand's own orders, site analytics, customer surveys and experiments) should carry more weight in any budget decision.

For Moorfield, the agency's launch report credits the influencer programme with £410,000 of EMV against a cost of £100,000. Amara sets it aside: it is not revenue, and it cannot say whether a single extra serum was sold.

Attribution assigns credit for a sale to the marketing touchpoints that preceded it. It is indispensable for running campaigns and dangerous when mistaken for a measure of what caused the sale.

Influencer activity is usually tracked in two ways. Tracked links carry UTM parameters, tags that tell web analytics which source, medium and campaign sent a visitor; unique discount codes tie an order to the creator whose code was used, even when the buyer never clicked. Moorfield set up both per creator in Module 5. Both miss and over-count. They miss buyers who saw a creator's video, later searched for the brand and bought without a code; much influence leaves no click. They over-count because codes leak: shared in group chats, posted on voucher sites and found at checkout by people who were already buying. A leaked code records a discount taken, not a sale created.

Attribution models decide how credit is split when a buyer has several recorded touches. Last-click gives it all to the final touch, favouring channels closest to the sale (brand search, email, retargeting) and starving those that created the demand. Rules-based multi-touch models (first-click, linear, time decay, position-based) spread credit by fixed rules; data-driven models assign it statistically. Google Analytics 4 retired its rules-based multi-touch models during 2023, leaving data-driven and last-click, a reminder that the available models change. Every model shares credit among recorded touches; none says whether the sale would have happened without them. Blake, Nosko and Tadelis (2015) showed the gap at eBay: when brand-keyword search ads were switched off, most of the traffic simply arrived through unpaid search results instead, so much of the attributed return was illusory.

Platforms add their own layer. Each reports the conversions it claims, including view-through conversions from people who saw an ad but did not click, under its own attribution window, so Meta, TikTok and Google can each claim the same order. Since App Tracking Transparency, as Module 6 explained, more of those conversions are modelled rather than observed.

Marketing mix modelling (MMM) works from the other end: regression on aggregate time series of sales, channel spend, price, seasonality and other factors estimates each channel's contribution without tracking individuals. It needs a long history and enough variation in spend, and its estimates rest on the model's assumptions, but it is unaffected by tracking loss. Open-source tools have lowered the cost: Meta's Robyn, and Google's Meridian, made available to all advertisers in early 2025. MMM is strongest when calibrated with experiments, the subject of Lesson 7.4. A cheap complement is a post-purchase survey asking "How did you first hear about us?".

The honest position is a chosen method with its limits stated.

Incrementality is the part of an outcome that would not have happened without the marketing activity. It cannot be observed directly, because no one can see what the same customers would have done without the campaign: the counterfactual. Experiments estimate it by comparing an exposed group with a comparable unexposed one.

Comparing people who saw ads with people who did not misleads, because exposure is not random. Delivery systems show ads to people predicted to buy, and creators' followers are already interested in the category, so exposed people would have bought more anyway. Gordon et al. (2019) tested this with 15 large US advertising experiments at Facebook: common observational methods, even with rich data on users, often failed to reproduce the experimental results and tended to overstate the effect of advertising. Randomisation is the reliable way to break the link between who is exposed and who would have bought.

Three designs are common. A user-level randomised holdout withholds the activity from a random share of the target audience; ghost-ad methods (Johnson, Lewis and Nubbemeyer, 2017) record which held-out users would have been shown the ad, sharpening the comparison. A geo holdout test runs activity in some regions and withholds it from matched regions; it needs no user tracking, suits activity that cannot be randomised person by person, and works with regional sales data (Vaver and Koehler, 2011). Platform lift studies, offered at the time of writing (2026) by platforms including Meta and TikTok, run randomised holdouts inside the platform to measure brand lift (survey-measured awareness or consideration) or conversion lift, subject to eligibility and spend thresholds that change.

Each has limits. Platform lift studies measure only that platform's ads, on its own data. Geo tests need regions with enough sales and similar trends, and spillover (people travel; creators' audiences are national) shrinks the measured effect. Organic creator posts cannot be withheld from a region at all; only paid amplification can. And sales are noisy: Lewis and Rao (2015), examining 25 large field experiments, found the median confidence interval on return on investment was more than 100 percentage points wide. An underpowered test wastes money.

A small brand can still test well: test one decision at a time; choose activity that can be switched off (amplification, a channel, a creator tier); run it long enough to see past weekly noise; fix the decision rule in advance; and report a range, not a point.

Amara's geo holdout withholds creator amplification from matched UK regions during the launch, and she checks how far creator codes have leaked to voucher sites. Together they suggest that 60% of attributed orders were incremental.

Return on investment for influencer marketing is a calculation anyone can do and few do honestly. Four choices decide the answer.

First, cost. The full cost includes creator fees, paid amplification of creator content, product seeding and shipping, and ideally agency, tool and staff time. Leaving out amplification because it sits in a paid social budget is the commonest way to flatter influencer returns.

Second, which orders count. Attributed orders carry a creator's code or link; incremental orders are the share an experiment suggests would not otherwise have happened. Report both, because the gap between them is the finding.

Third, revenue or profit. Return on ad spend (ROAS, revenue ÷ cost) ignores the cost of the goods sold. Return on spend should use gross profit: (revenue × gross margin − cost) ÷ cost. As Module 6 showed, at a 68% margin the break-even ROAS is 1 ÷ 0.68, about 1.47x, so a ROAS above 1 can still lose money.

Fourth, time horizon. A first-order calculation ignores repeat purchases, which may be much of a new customer's value in a replenished product such as skincare, and the longer-term brand effects that Binet and Field (2013) argue build over time. Lifetime value claims are forecasts, not results: state the assumption and measure it rather than invoking it to rescue a weak first-order result. Halo effects, such as extra sales through retailers or of other products, are real but should be estimated, not assumed.

Reporting is where honesty is tested. A one-page evaluation for senior stakeholders states the objective, the full cost, attributed and incremental results side by side, how the incremental estimate was made and how uncertain it is, what is not yet known, and a recommendation with the decision it asks for. Lead with the result, not the activity list. Do not claim precision the data cannot support: a range with its assumptions is more honest than one exact percentage. And keep EMV out of the return calculation entirely.

Senior readers also need the comparison that makes a number meaningful: against break-even, against the plan, and against the next best use of the money, such as the same sum in paid social or email.

For Moorfield, the influencer launch cost £100,000. On attributed orders it returned +17.1% on spend; on the 60% of orders estimated to be incremental, −29.7%. The same campaign is a success or a loss depending on which question is asked, and Amara's job is to tell Tom Hargreaves which question matters.