Start from requirements
A requirement is a sentence that says what the tool must do for a written use case, in a form that can be checked. "Send a message within fifteen minutes of a basket being abandoned, using the product data from the event" is a requirement. "Good automation features" is not. The use cases come from the stack design, which the designing a marketing stack guide covers; the evaluation turns the use cases a layer needs into the requirements a tool must meet.
Write them in four groups. Functional: what the tool does for each use case, with the specific capability named. Integration: what it must receive, from where, how quickly, and what it must send back, which the data flow guide explains how to specify. Data and governance: which identifiers it keys on, how it handles consent, where data is processed, what it retains and how it is deleted. Operational: who will run it, what skills it assumes, how changes are made and approved, what the vendor's support covers. Mark each requirement as essential, important or useful, and resist marking everything essential; a list where everything is essential cannot rank anything.
Involve the people who will use and govern the tool in writing the list: the channel owner who will run campaigns, the analyst who will report on it, the engineer who will integrate it, the privacy owner who will review its data handling. Each will add requirements the others would miss, and a requirement added after the demo is a requirement the demo was designed to hide. Keep the list with the evaluation and later with the contract; it is the document the renewal conversation starts from.
When the tool being chosen is the website platform itself, the guide to choosing a CMS for SEO supplies the search requirements for the list, from URL and redirect control to rendering. When it is the email platform, the guide to choosing an email platform supplies the email-specific rows, from signing mail with the business's own domain to a suppression list that survives every import.
Scoring criteria
A scoring table turns requirements into a comparison. Each criterion is scored for each candidate, weighted by importance, and the weighted scores are summed. The point is less the total than the discipline: every candidate is judged on the same questions, and the questions were written before anyone saw a product.
| Criterion | What is scored | Evidence | Weight |
|---|---|---|---|
| Use-case fit | Does it meet each essential and important functional requirement? | Proof of concept, not the demo | High |
| Integration | Native connectors or documented interfaces for what it must receive and send; latency it supports | Documentation, a test integration in the proof of concept | High |
| Identity and consent | Keys on identifiers the stack holds; reads and respects consent state | Documentation, test with consent combinations | High |
| Data handling | Where data is processed and stored; retention, deletion, sub-processors; export on exit | Vendor security and privacy documentation, contract terms | High |
| Usability for the owner | Can the people who will run it do so without the vendor? | Hands-on time in the proof of concept | Medium |
| Reporting and measurement | Emits the events the measurement layer needs; its own reports match the stack's definitions | Reconcile its numbers to analytics in the proof of concept | Medium |
| Governance | Roles, permissions, change history, approval flows, environments | Documentation and a test of the permission model | Medium |
| Vendor | Stability, roadmap transparency, support quality, reference customers of similar shape | References, support interactions during the trial | Medium |
| Total cost | Licence, implementation, integration, training, owner time, data costs | Costed plan for the contract term plus a renewal | High |
| Exit | Data export in a usable form, configuration portability, notice terms | Contract and a test export | Medium |
Weights are set from the stack design before scoring, and the design usually makes integration, identity and data handling the heaviest, because they are the criteria that cannot be fixed after purchase. Score with evidence rather than impressions: a criterion scored from a demo is a guess, and the table should say where the evidence came from. Where two candidates are close on the weighted total, the essential requirements decide; a tool that fails one essential requirement does not win on the sum of the others. The same gate-before-sum rule, together with a check on how far the ranking depends on the weights, is set out in the weighted scoring matrix method for choosing between options.
The table is also where the incumbent is scored, if there is one. Evaluations that score only the challengers end up comparing a known tool's faults against unknown tools' promises.
Trials and proofs of concept
A demo shows what the vendor wants to show; a proof of concept shows whether the tool does the use case with the business's own data. The second is the only evidence that counts for the high-weight criteria.
Design the proof of concept around two or three use cases chosen for difficulty rather than ease: the one with the hardest integration, the one that needs identity resolved across sources, the one the owner will run most often. Give the candidates the same data (a sample of events, a segment, a conversion feed) and the same success definition, written in advance: the message sent within the latency required, the segment matching the warehouse's count, the conversion arriving in analytics with the right parameters. Have the people who will own the tool do the work, with the vendor available to help but not driving. And set a fixed period, because a proof of concept that runs until it succeeds has no result.
Record what happened against each requirement, including what needed vendor intervention, what needed a workaround and what did not work. The record is the evidence column of the scoring table. Where a candidate cannot support a proof of concept (no trial, no sandbox, no test data), that is a finding about the vendor, and it should be scored. For AI features, the proof of concept needs a quality test as well: a set of real tasks, a scoring rubric and a pass bar agreed in advance, as testing prompts and models sets out, because AI output varies from run to run and one good result in a demo proves little.
A trial of a tool already in the stack, for a new use case, follows the same method with less ceremony. The question is still whether the tool does the job with real data, and the answer still goes in a record. When the new use case is an AI feature, decide first whether it belongs in the tool's built-in AI or a separate assistant, as the comparison of standalone AI assistants and built-in AI features explains.
Integration and data checks
Integration and data handling are the criteria most often checked last and regretted first. They are discoverable from documentation and a short test before any commitment, and the checks are the same for every category.
What it receives. For each inbound requirement, does the tool have a native connector to the source, a documented programming interface, a file import, or nothing? What latency does each route support, and does it match the use case? Does it accept events in the stack's schema, or will every event need mapping? A tool whose only inbound route is a file upload will create a manual process that eventually stops.
What it sends. For each outbound requirement, can the tool emit its results (sends, exposures, conversions, segment membership) to the data layer and to analytics, in a form that can be joined back to the person who produced them? A tool that keeps results inside its own reporting breaks the stack's loop.
Identity. Which identifier does the tool key on, and does the stack hold it on the events the tool will receive? If the tool keys on email address and the events carry only a cookie, matches will fail and the tool will create duplicates. Where identity is resolved in a customer data platform (CDPUnified first-party customer profiles.) or warehouse, the tool should consume that resolution rather than build its own.
Consent. Can the tool read consent state and act on it, per purpose? Does its tag, if it has one, work with the tag manager's (GTMTag management system.) consent gating? The tag governance guide covers the vendor review for the tag side, and it should be done during the evaluation rather than after.
Data handling. Where is data processed and stored? Which sub-processors does the vendor use? What is retained and for how long, and how is deletion requested and confirmed? Does the vendor use the data for its own purposes? What is exported on exit, in what form? These answers come from the vendor's documentation and contract, and the privacy owner should read them during the evaluation.
Each check produces a yes, a no or a workaround, and workarounds are costs: a mapping to maintain, a manual step to run, a sync to monitor. Put them in the total cost and in the decision record, so that the person who inherits the tool knows why the nightly export exists.
A software company evaluates a lifecycle messaging tool
A subscription software company with 42,000 customers needs a lifecycle messaging tool for onboarding, upgrade and renewal sequences. It writes 27 requirements (11 essential), weights integration, identity, data handling, use-case fit and total cost at 3, and the rest at 1 or 2, and shortlists three candidates plus the incumbent email platform. The proof of concept runs for four weeks on two use cases: a day-three onboarding message driven by product events from the warehouse, and a renewal sequence keyed on the billing system's subscription date. Candidate A meets 26 requirements and syncs the warehouse segment within 15 minutes; candidate B meets 24 but keys only on email address and creates 1,900 duplicate profiles from the test data; candidate C meets 25 but offers no export of sequence configuration on exit. The incumbent meets 18. Weighted scores are 171, 142, 158 and 119. Two-year total cost, including 60 engineering hours of integration for A, comes to £58,000 against £71,000 for C. A is chosen, with the decision record noting the one unmet requirement (a multilingual template feature) and a review date in 12 months.
Total cost and contracts
The licence is the visible part of a tool's cost and rarely the largest. Total cost over a realistic period, usually the contract term plus a renewal, includes implementation (the vendor's fees and the team's time), integration (engineering hours, any middleware, ongoing maintenance of mappings), training and the learning period, the owner's time to run it, data costs it adds (events ingested, contacts held, usage beyond a tier) and the cost of the workarounds the integration checks found. Cost it for each candidate on the same basis, and include the incumbent, whose implementation cost is already paid.
Contracts decide the cost of leaving as well as of staying, and the terms worth negotiating are the ones that matter later. Price at renewal, not only at signing, and the cap on increases. Notice periods and what happens at the end of the term by default. Data export on exit, in a documented form, and deletion on termination with confirmation; when a tool is later retired, these are the terms that set the order of the exit, as the tutorial on retiring a marketing tool shows. Ownership of the configuration and content the team builds in the tool. Where a vendor offers an AI model "trained on your brandTraining an existing language model further on a set of examples so that a style, format or task pattern is built into the model itself.", ask which mechanism that means and what leaving would involve; the retrieval vs fine-tuning comparison lists the questions. Usage tiers and what happens when they are exceeded. Service levels for support and for the tool's availability, where the use case depends on it. And the security and privacy terms the privacy owner needs, which should be agreed before the proof of concept starts rather than negotiated after the team has built on the tool. For AI tools specifically, the guide to evaluating AI tools for a marketing team adds the questions on training and retention terms, admin and audit controls, and model changes under you.
Procurement and legal should see the requirements and the scoring, not only the chosen vendor's proposal, because the terms that matter are the ones that protect the requirements. The KPIA metric tied directly to an objective. the tool is expected to move belongs in the business caseThe document that lets decision-makers judge one investment: what is asked for, the alternatives, every cost, a range of upside, the risks and the staging. alongside the cost, so that the renewal decision has something to judge against. The tutorial on writing a business case for a marketing investment builds that case step by step, from the full cost and a range of outcomes to a staged commitment with a decision point.
The decision record
The decision record is a short document that says what was chosen, against which alternatives, on which evidence, with which known gaps and what will be reviewed when. It is the artefact most evaluations skip and the one the organisation most needs a year later, when the person who ran the evaluation has moved on and the renewal is due.
It contains: the use cases and requirements; the candidates considered and why any were excluded early; the scoring table with the evidence for each score; the proof of concept results; the total cost comparison; the integration and data findings, including workarounds; the contract terms agreed; the owner; the success measures and when they will be checked; the known gaps and how they will be handled; and the date the decision will be reviewed. A page or two is enough for most tools; a data-layer platform may need more. The marketing tool evaluation template lays out these contents as a workbook to fill during the selection, from requirements and weights through scores, trial findings and total cost to the decision itself.
Keep the record with the inventory the tool categories hub describes, and read it at renewal, at any incident involving the tool and when the use cases change. The tutorial on reviewing a marketing tool before it renews sets out the renewal reading step by step, from a calendar built on notice dates to a recorded decision. The automation governance guide covers the ongoing governance of one of the categories where the decision record matters most, because the tool runs on its own once configured.
Common mistakes
- Letting the demo write the requirements. Write them first, with the people who will own, use and govern the tool.
- Scoring from demos. A score without evidence is an impression; the proof of concept with real data is the evidence.
- Marking everything essential. A list that cannot rank anything cannot decide anything.
- Checking integration after purchase. The integration and data facts are discoverable from documentation and a short test before any commitment.
- Comparing licence prices. Total cost includes implementation, integration, owner time, data costs and workarounds; compare that.
- Not scoring the incumbent. A fair score for the current tool is a valid and often cheap outcome.
- No decision record. The renewal in a year is then a negotiation with no memory.
