Outcome Harvesting: a practical guide

Outcome harvesting is a participatory evaluation approach for identifying, formulating, verifying and making sense of outcomes — defined narrowly and specifically as observable changes in the behaviour, relationships, actions, policies or practices of a social actor that the intervention plausibly influenced.

Three commitments distinguish it from most evaluation designs. First, it does not begin from predefined objectives or indicators; outcomes are harvested from evidence, whatever they turn out to be, including negative and unintended ones. Second, every harvested outcome is written as a discrete, verifiable statement of who changed what, when and where — not as an aggregate or a rating. Third, contribution is claimed cautiously: the outcome statement is separated from the contribution description, so a reader can accept the change happened while still disputing the programme's role in it.

The unit of analysis is the outcome, not the project. A harvest of an advocacy programme might produce eighty outcome statements, of which the programme was the primary influence on twelve, one contributor among many on fifty, and a marginal presence on the rest. Conventional reporting would flatten that into a percentage. Harvesting keeps it legible.

The approach was developed by Ricardo Wilson-Grau with colleagues from the mid-2000s, out of frustration with evaluating international development and advocacy programmes whose results were real but did not resemble the logframe. It was formalised in Outcome Harvesting: Principles, Steps, and Evaluation Applications (Wilson-Grau, 2018), which remains the reference text.

It draws directly on outcome mapping, developed at Canada's International Development Research Centre, and shares its central move: shifting the object of measurement from a programme's outputs to changes in the boundary partners it works through. Where outcome mapping is a planning and monitoring system used prospectively, harvesting is retrospective — it can be applied to a programme that never used outcome mapping at all, which is much of its practical appeal.

Outcome harvesting earns its cost in a specific set of conditions, and is a poor choice outside them.

• Use it when outcomes are not knowable in advance. Advocacy, systems change, policy influence, coalition-building — anywhere the programme's job is to create conditions rather than deliver a countable service.
• Use it when the theory of change has moved. Harvesting is indifferent to whether the original design predicted the outcome, which makes it honest about adaptive programmes rather than punishing them.
• Use it when you need defensible evidence from a dispersed programme. Multi-country, multi-partner work where no single monitoring system saw everything.
• Do not use it to measure service delivery. If the question is how many people were treated, trained or fed, a results framework answers it faster and cheaper.
• Do not use it when you need attribution. Harvesting establishes plausible contribution. If a funder needs a counterfactual — what would have happened anyway — this is the wrong instrument and no amount of substantiation will make it the right one.
• Do not use it with no budget for verification. An unsubstantiated harvest is a list of claims. The verification step is not optional trimming; it is where the method's credibility lives.

A regional programme spent four years supporting civil-society organisations to influence national water-governance policy across six countries. Its logframe promised three policy changes. At close, none of the three had happened in the form described.

The harvest read four years of partner reports, board minutes and press coverage, and drafted 94 candidate outcomes. Engagement with partners cut that to 61 — several drafts described the same change twice, and a dozen were activities dressed as outcomes. Substantiation went to 22 of the 61, chosen to cover each country and each type of claimed change: ministry officials, journalists, and staff at organisations that had competed with the grantees.

What came back reframed the programme. Two of the promised policy changes had in fact occurred, in different ministries and under different names than the logframe anticipated. A third had been abandoned after a change of government — and the harvest documented that the coalition built to pursue it had been redeployed to a water-tariff consultation nobody had planned for, producing a change the programme had never claimed credit for because it appeared in no results framework.

The substantiators also disputed two contribution claims outright, judging that a national NGO outside the programme had driven the change. Those two stayed in the harvest, marked as disputed. That is the method working correctly, not failing.

• Harvesting activities instead of outcomes. "Held three workshops with the ministry" is not an outcome. "The ministry's planning unit adopted the consultation protocol in March 2025" is. This is the single most common failure and it is usually caught at the engagement step, expensively.
• Writing outcomes that cannot be verified. If no independent person could confirm or deny the statement, it will not survive substantiation and should not have been drafted.
• Merging outcome and contribution. Once the two are written as one sentence, a reader who doubts the contribution claim has to reject the whole outcome. Keeping them separate is what lets a sceptical funder accept the change while arguing about the credit.
• Choosing friendly substantiators. Substantiation by people who like the programme produces a document that reads well and proves nothing.
• Treating the harvest as a census. A harvest is purposive. Claiming it captured everything that changed invites a challenge it cannot survive.
• Skipping negative and unintended outcomes. Their absence from a harvest is a signal to reviewers that the harvest was not independent.

Harvesting cannot answer counterfactual questions. It establishes that a change occurred and that the programme plausibly contributed; it does not establish what would have happened otherwise, and presenting it as though it does is the most common misuse.

It is also vulnerable to the memory and availability of its sources. Outcomes from year one of a five-year programme are systematically under-harvested because the people who witnessed them have moved on and the documents are thinner. Harvests conducted annually rather than once at close mitigate this, at higher cost.

Finally, it is labour-intensive in a way that scales badly. The document-review step grows with programme size and duration, and the substantiation step grows with the number of outcomes retained. Harvests are routinely under-budgeted on both.

Wilson-Grau, R. (2018). Outcome Harvesting: Principles, Steps, and Evaluation Applications. Information Age Publishing. https://doi.org/10.1108/978-1-64113-394-4 — the reference text, covering the six steps and a range of applied cases.

Mayne, J. (2012). Contribution analysis: Coming of age? Evaluation, 18(3), 270-280. https://doi.org/10.1177/1356389012451663 — the companion argument for making contribution claims rigorous without a counterfactual.

Befani, B., & Mayne, J. (2014). Process tracing and contribution analysis: A combined approach to generative causal inference for impact evaluation. IDS Bulletin, 45(6), 17-36. https://doi.org/10.1111/1759-5436.12110 — how to strengthen the substantiation step with formal evidentiary tests.

Outcome harvesting sits alongside several approaches that share its causal logic and differ in what they demand of you.

• Outcome mapping — the prospective sibling. Use it if you are designing monitoring now rather than evaluating retrospectively.
• Contribution analysis — a more structured route to the same kind of claim, built around iteratively testing a postulated theory of change.
• Contribution tracing — adds Bayesian evidentiary tests to contribution claims, useful when a funder wants stated confidence levels.
• Most significant change — also story-based and participatory, but selects a small number of accounts through structured deliberation rather than harvesting many.
• Process tracing — the underlying causal-inference machinery several of the above borrow from.

The bottleneck in a harvest is step 2: reading years of reports, minutes and correspondence to draft candidate outcomes. That is document analysis at a volume where consistency degrades — the outcomes drafted in week four rarely match the standard of week one.

AI-assisted analysis is genuinely useful here, for surfacing candidate outcome statements across a large corpus and for applying a consistent classification once outcomes are agreed. Evidano is used this way in published evaluation work; JET Education Services applied it across twelve years of reporting in an outcome-harvesting exercise, and UN evaluation offices have used it for comparable document synthesis. See published case studies for how those teams handled review and verification.

What software cannot do is steps 3 and 4. Engagement with change agents and independent substantiation are the reason a harvest is credible, and they are irreducibly human. A harvest that automates the reading and skips the verifying has produced a faster list of unverified claims, which is worse than a slower one.