Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Growth Marketing › Growth Fundamentals

Hiring and Evaluating a Growth Marketing Agency

How to hire and judge a growth marketing agency or consultant: what the business keeps, the brief, an opening round, questions, warning signs, reports and exit.
Guideintermediate20 min readUpdated

What a growth provider is hired to do

An agency, consultant or freelancer can be engaged to take over part or all of the operating model that the growth marketing pillar lays out: loops for how customers arrive and come back, a north star with the economics beneath it, a backlog of experiments run on a weekly rhythm, and the builds and channel work those experiments call for. The in-house against agency comparison settles whether to buy help in; this guide picks up after that choice, with the purchase itself: the scope, the provider and the means of telling good work from busy work. The agency selection guide asks for the same written scope from a freelance specialist as from an agency, so the scope, questions, reports and exit points here apply to a consultant or freelancer as well, as general practice.

As general practice, the table below sets out what a provider can be asked to do and what the business keeps.

JobWhat the provider can be asked to doWhat stays with the business
The model and the north starDraft the model sentence, map the loops with their leak points, propose a north star and its input treeThe decision to adopt a north star or change it, and the definition behind it
Cohorts and economicsBuild the cohort table and the unit economics view from the business's data and keep them refreshedThe definitions of margin and acquisition cost, which finance has to accept, and the ledger the spend comes from
Backlog and cadenceKeep the backlog ranked, prepare the pre-read for the weekly hour, draft the log entriesOwnership of the backlog and the log, and the decision on what runs next
TestsWrite the plans, build the variants, run the tests and supply the counts for the read-outsAppointing who calls each result, to meet the gate set out below, and the decision to ship
Product changesSpecify them and write the hypothesisEngineering capacity, release dates and the product itself
Channel and lifecycle work the tests needRun campaigns and messages inside the business's accountsThe accounts, the budget and the brand frame the tests work inside

How many rows to hand over is the first scoping decision, because each row the provider takes on enlarges the scope and the reports it must account for. As general practice, keep one named person on the business side for each entry in the right-hand column.

What stays in the business's hands

The agency selection guide's position is that the ad accounts, the analytics properties, the customer data and the paid-for creative should all be the business's. Growth work adds a further class of record: the ones the model itself is kept in. As general practice, each row below names a record, the place it should sit from day one and what a proposal ought to demonstrate about how it will be kept there.

RecordHome from day oneEvidence to ask forGuide
The growth model document and the north star definitionThe business's own document store, each version datedWho drafts and approves a new version and how a change is recorded beside the historyChanging a north star metric
The cohort table and the queries that build itThe business's warehouse or spreadsheetHow it is refreshed, who can open the queries and how a change of definition is flaggedThe cohort analysis tutorial
The unit economics table and its definitionsBeside the numbers, with spend taken, as general practice, from the finance ledgerWhether the provider's own fee counts in acquisition cost, and who reconciles the table with financeThe LTV and CAC guide
The backlog and the learning logOne place the business can open, losses and inconclusive results includedWhat each entry carries, in the form the cadence guide lists, drawn from the read-out templateThe experimentation cadence guide and the experiment results report template
Test plans, variants and trackingThe business's own code, tag container and testing accountsThe accounts and records the conversion agency guide lists, which this page does not repeatThe guide to hiring a conversion agency
The rule that credits customers to channels and loopsWritten down, with each change recorded and datedWho applies it, and that the party being credited is not the only one who can change itThe metrics layer and definitions guide
Channel, lifecycle and analytics accountsIn the business's name, with the provider holding the access it needsWhich accounts the provider needs and through what accessThe agency onboarding checklist

Two of these requirements are best treated as pass-or-fail gates, as general practice, and kept out of the scoring: the definitions, the data and the learning log sit in the business's records, and every result is called by the business's analyst or by an independent reader who does not report to the provider's build team and whose calls the business can check against the plan. The cadence guide wants the reader of results to sit outside the reporting line of whoever's variants are being read, and the experimentation programme playbook settles that the analyst alone declares a result; the independent reader is this page's general-practice alternative. A proposal that fails either gate is set aside whatever it would have scored elsewhere; the weighted scoring matrix framework makes the same separation between must-haves and weighted criteria.

The brief a provider works from

The agency selection guide calls the brief the place where most agency work succeeds or fails. For growth work, as general practice, the closest thing to one is the pillar's first growth plan: one sentence on the model, one north star metric (chosen with the north star and OKRs framework), the economics as they stand, a backlog of experiments and a weekly ritual with a learning log. A business that has written those, however roughly, can give them to every provider; one that has not can write rough versions first or buy the missing ones in the first round described below. As general practice, five points belong in the request itself.

  • The objective and the ceiling. What the programme is for, tied to a business objective as the objectives and KPIs guide asks of any measure, and the largest sum the business will pay a provider for the scope as defined, the fee included.
  • The baseline the business already has. Which of the five items exist, which numbers have been built, such as cohorts or acquisition cost by channel, and the definitions that sit behind them.
  • What must not get worse. The guardrails that sit beside the north star and beside each test's primary metric.
  • Who the provider will work with. The named owner of the programme, the analyst, the finance partner and the product and engineering people whose time tests will need. The selection guide notes that an agency without access to sales data cannot optimise for sales, so say which data the provider can read.
  • What counts as a result. One north star and the measures beneath it, defined before work starts, with the business's own records as the source.

Give every provider an identical brief and an identical set of facts about the business, and agree the scoring criteria before any reply arrives. The agency request for proposal template offers a common brief, question list and scoring grid; the agency scope of work template provides the fields for the agreement.

Scoping the opening round: a baseline before experiments

The pillar tells a small team to begin by building its cohort table and settling on a single north star, with experiments to follow. General practice is to buy that groundwork as a time-boxed first round, with three questions riding on it: does the provider's reading of the business match the business's own records, does its way of calling tests hold up, and can its numbers be reconciled with the finance ledger and the order system. Besides the general scope items in the selection guide, the statement of work for this round should cover the points below.

What the round delivers. The model sentence, the north star and its inputs, the cohort table, the unit economics table, a ranked backlog and the first log entries, each in the business's own records. The weekly growth meeting tutorial lists what a team needs before its weekly meeting starts and is worth reading beside this list.

Which tests, if any. As general practice, only the few that the baseline justifies, each an A/B test or a holdout aimed at the leak the baseline shows, and each with a decision rule written before launch. Also as general practice, no test starts until the tracking behind its primary metric has been checked against the business's own records.

Who calls results. Named in the scope, to meet the gate above. The playbook's reasoning is that winners declared by their own channel owners fill the programme with false positives that fail to hold up once shipped.

The weekly rhythm. Who owns the meeting, who attends, who writes the pre-read and when decisions are written down. As general practice, the business's growth owner keeps the decisions even when the provider prepares the papers.

Access and capacity. The data the provider may read, the accounts it works in, the product and engineering time that tests needing a change will take, and what the provider may put into its own tools; the AI rules for agencies guide covers the last point.

Questions for every shortlisted provider

Put identical written questions to every provider on the shortlist and compare the replies in one grid. As general practice, a reply that attaches a named person to each task, and says what that person needs from the business, deserves more weight than one that stays general.

QuestionWhat a careful answer contains
How will you decide what to work on first?A reading of the business's own cohorts and loops, with the leak chosen by where a feasible change moves the loop most, as the growth loops guide describes, not a list of tactics taken from other companies
Who calls a test, and by what rule?A decision rule written before launch, a named reader who meets the gate above, and inconclusive results recorded as inconclusive
What counts as a customer, and what goes into acquisition cost?A paying customer from the order system, with people, agencies, tools, creative and incentives in the cost, shown by channel as well as blended
How do you read retention and lifetime value?Cohort curves compared at the same age, margin rather than revenue, and no lifetime used for a decision beyond the oldest cohort the business has observed, with any projection shown alongside and labelled as one
How will customers be credited to a loop or a channel?A written rule, applied by a person who does not own the loop, with a stated way of checking it
What do you need from our product and engineering people?Named roles, the time each test will take and the tests that depend on a release
What happens to a win after it ships?As the tutorial on shipping a winning variant sets out for the rebuild, plus a revisit against cohort data a quarter later, logged beside the original call
What do you do for a stage with too little traffic to test, such as a late activation step or a referral prompt?The options the cadence guide gives for low-traffic stages, and for website changes the approach conversion work without an A/B test describes
How would we leave?The model, the log, the queries, the definitions and the variants in the business's hands from the first day

Warning signs in a pitch

Each sign below is a prompt for a further question, not a verdict. As general practice, keep asking until the provider's answer is specific.

  • Tactics promised before the business has been read. The pillar is blunt about hack folklore harming the discipline; a plan should begin with the model and where it leaks.
  • Another company's loop or results offered as the plan. The growth loops guide lists copying another company's loop among its mistakes.
  • Test volume as the headline. The cadence guide's first mistake is velocity as the goal.
  • Only wins in the report. The cadence guide wants every outcome, losses and inconclusive results included, reported in one place.
  • Lifts quoted from the test alone. Wins are revisited against cohort data a quarter later; a proposal that does not mention the revisit has not planned for it.
  • Acquisition cost without the provider's fee, or blended where a channel decision is being made. Media-only and blended figures are both on the LTV and CAC guide's list of mistakes.
  • Lifetime value on revenue, or projected from a whole-base churn rate. The same guide gives the margin and cohort alternatives.
  • A new north star proposed before the definition and inputs have been checked. The guide to changing a north star starts by checking the definition and the inputs, not a leader's taste or a fresh planning cycle.
  • Definitions that only the provider can change. The Goodhart's law framework advises keeping definitions out of the targeted team's hands.
  • No word on handover. The exit part of the scope template is the place to spell it out, with contract wording left to the business's legal advisers, as the selection guide says.

Reading the fee and the incentive

The fee model should fit the work, as the selection guide advises; its models run from the retainer to a performance element. Growth work adds three cautions, each carried over from a guide the site already holds and each general practice when applied to a provider.

Paying more as tests multiply, or per winning test, rewards volume and premature calls. The cadence guide lists velocity as the goal among its mistakes, and the conversion agency guide cautions against a fee per winner.

A performance element tied to sign-ups, to acquisition cost or to the KPI that is the north star, with nothing beside it, is the pattern the Goodhart's law framework describes. Pair any such target with a guardrail metric that must not get worse, such as the retention of the cohort an acquisition target brings in beside cost per customer, and keep the definitions out of the paid party's hands.

A fee that grows with the media budget the provider advises rewards spending more, the objection the selection guide raises to commission on media spend. For acquisition spend, the payback ceiling comes from payback and cash agreed with finance, as the retention and payback guide describes. As general practice, the largest sum the business will pay a provider for the scope, the fee included, is set before the fee model is chosen.

Then, as general practice, test each proposal's total against that sum. The provider's own fee is an acquisition cost, since the LTV and CAC guide's definition adds agencies and tools to media; a proposal that omits its fee, the creative, the tools or the business's own team time and still fits has left those costs unpriced. Also as general practice, ask each provider to split its price into the fee, the build time and any tools, so that the replies can be compared.

Reports the business can check

The table sets each report line beside a record the business already holds.

Report lineThe business's test
The north star and its inputs for the periodThe business's own records with the definition unchanged since the last report, checked against the measurement change log; where a total moved, split into mix and rate as explaining a change in a metric advises
Retention and lifetime valueThe cohort table at the same ages, built from orders or subscriptions, on margin rather than revenue, with the limit on lifetime value set out in the questions above
New customers by channelThe order system's or CRM's own flag, as the new and returning customers guide explains, with the provider's total reconciled to it
Acquisition cost and payback by channelSpend from the finance ledger, as general practice, with the provider's fee included, set against the cumulative contribution curve for the same channel
Movement in the active baseThe growth accounting tutorial's bridge, which has to balance: new, retained, resurrected and churned
Tests run, won, lost and inconclusiveThe learning log, read against the test reports the conversion agency guide describes, with each guardrail result shown
Shipped winsA revisit a quarter later against cohort data, logged beside the original call
The output of a loopThe written crediting rule, kept by a person other than the loop's owner and, as general practice, checked with a holdout or with customers' own answers as the self-reported attribution guide describes; counted by generation, since a loop valued at its first round of invitations leaves out later generations and one credited with customers still to come counts them before they have arrived

Weekly movements of a handful of customers may be noise, a point the reading small numbers guide makes at length. Where the provider's totals rise and the business's own stay level, the diagnosis of flattering agency reports separates a measurement problem from a performance problem.

Judging the provider and the programme

General practice is to keep two verdicts apart: one on the provider's own work and one on how the programme is performing. The programme's numbers also reflect the product, the offer, the traffic and the business's own capacity, which the provider does not wholly control. The early weeks can only show what the provider governs: access set up in the business's accounts, definitions written down and agreed with finance, the baseline built and reconciled to the ledger and the order system, plans written before launch, a log the business can open and the service levels the scope sets.

The programme's verdict comes later. As general practice, wait until the baseline is built, the first tests have been read against their plans and the shipped wins revisited, then look at three things together: whether the north star is still moving ahead of revenue, whether the cohort curves and the payback by channel have moved in the direction the work aimed at, and whether the log records decisions that changed. The cadence guide measures a programme by decisions the learning changed rather than by a tally of wins.

Score the business too, as the two-way scorecard in the selection guide does. Slow product capacity, changed definitions and decisions not made are the business's own problems, which, as general practice, a provider cannot fix. The underperforming agency tutorial is the place to begin when the verdict is poor, and the pre-renewal agency review is the one to read before renewal.

Changing provider, or bringing the work in-house

The order for collecting accounts and assets and for withdrawing access is the one the agency offboarding checklist gives, set in context by the transitions section of the selection guide. As general practice, a growth programme adds the points below, which an account inventory would not list.

  1. The model and the north star. The current model document and its dated earlier versions, the input tree and the loop maps with their measured leak points.
  2. The cohort table and its queries. The queries, the refresh schedule and the definitions of active and acquired; whoever takes over rebuilds last month's figures before changing anything.
  3. The unit economics table. The definitions, the spend lines and the reconciliation to the ledger, with the provider's own fee shown as a line.
  4. The backlog and the learning log. Both in the business's hands, with the conversion agency guide's decisions on running tests applied to any test still live.
  5. Wins in flight. The revisits that fall due after the switch, handed to a named person.
  6. The crediting rule and any holdout still running. Each holdout is read out as planned or closed early with the closure noted.
  7. The weekly meeting. The owner, the attendees, the pre-read format and the next dates.

Where the provider still controls the ad accounts, the diagnosis of ad account handover works through each account in turn, and the post-switch drop diagnosis covers results that slip after a change. As general practice, bringing the work in-house uses the same list, with the in-house team receiving it.

Weighing growth proposals for a wood-care shop (illustrative)

Illustrative example — not a real result

Kolvanth Woodcare weighs three growth proposals

Kolvanth Woodcare, an invented online seller of wood stains, oils and finishing brushes, has a marketing team of 2 and a part-time analyst, and finds that many first-time buyers do not place a second order. Seeking outside help, it completes as much of the pillar's first plan as it can: a model sentence and a rough north star of customers who reorder within 90 days of a first order, with no cohort table yet. Its ceiling on what a provider may be paid for the first round is £24,000, the provider's own fee included, and three providers receive the same brief.

Two gates apply before any scoring. A proposal must keep the data, the definitions and the learning log in Kolvanth's own records, and every result must be called by Kolvanth's analyst or by an independent reader who does not report to the provider's build team and whose calls Kolvanth can check against the plan. Proposal A asks for a fee per new customer, runs its tests in its own tool account, keeps the log and the cohort definitions in that tool, lets the team that builds the variants call the results and reports acquisition cost on media alone. It fails both gates and is set aside whatever its pitch. Proposals B (an agency) and C (a freelancer) pass. Weights fixed before any proposal was opened are depth of the baseline round 25, quality of the plans, read-outs and log 25, reports Kolvanth can test 20, fit with the team and product capacity 15 and shape of the fee against the work 15. Against those criteria B scores 5, 4, 4, 3 and 4, a weighted total of 410 out of a possible 500, which is 82 out of 100, and C scores 4, 3, 3, 2 and 5, a total of 340, which is 68. Reweighting to put the fee at 30 and the baseline at 10, with the other three unchanged, gives B 79 and C 71, so B still leads.

Kolvanth appoints B for a baseline and two tests. B's price of £19,800 fits inside the ceiling and is made up of a retainer of £12,960 over four months and build time for the two tests of £6,840, each shown as its own line. Kolvanth's analyst calls the results.

B's first monthly report puts acquisition cost at £30 per new customer. Kolvanth's order system shows 540 new customers that month and its ledger £16,200 of media, which is £30 a customer, but the ledger also holds B's retainer for the month (£3,240), creative (£1,620) and Kolvanth's own tool costs (£540). Together these lines make £21,600, or £40 per customer before the separately billed build time for the tests and Kolvanth's own team time, so the report is restated with the lines shown and the figure of £40 enters the unit economics table with those lines named. One of the two tests wins on its decision rule: a one-tap reorder link in the dispatch email lifts the share of first-time buyers who reorder within 90 days from 21% to 27% in the test. Once the first cohorts acquired after it shipped have been read at the same 90 days of age, the cohort table shows 23%, and the log records both figures beside the original call. Kolvanth uses the 23% in its model.

The growth fundamentals hub lists the rest of the section's guides and the problems they treat.

Common mistakes

  • Committing to the entire programme at once. A time-boxed first round puts the questions to the test.
  • Letting the provider define the numbers. The north star and the definitions of new customer, margin and acquisition cost belong to the business.
  • Starting with experiments. Without cohorts the business cannot see whether acquired customers stay, and without a north star each test is argued on a different measure.
  • Letting builders mark their own work. A result called by the team that built the variant fails the gate set out above.
  • Paying for volume or for a single number. Pair any target with a guardrail.
  • Leaving the provider's fee out of acquisition cost. A figure without the fee, the creative and the tools flatters the channel.
  • Accepting a win from the test alone. Revisit it against cohort data a quarter later and log both figures.
  • Keeping the log in the provider's tools. The log leaves with the provider unless the business holds it.
  • Signing without a way out. The model, queries, log and definitions should be collectable on the day the contract ends.

Frequently asked questions

Is a growth agency the same as a conversion optimisation agency?

No, although the work overlaps. A conversion provider works on page and funnel tests and the research behind them. A growth provider is asked to work across the journey, including activation, retention and the economics beneath them, and to keep the model and the backlog those tests come from. Where the growth work includes page tests, hold it to the standard the site's guide to hiring a conversion agency sets for them.

Can the agency own the north star and the definitions behind our numbers?

It can draft them, and a good proposal may, but the business should approve them, hold them in its own records and make sure the provider is not the only party that can change them. The north star and the definitions of margin, new customer and acquisition cost decide what the tests and the reports count, and a provider paid on a number it alone defines can move the number without moving the customers. Dated versions and a recorded route for change keep the history the business needs if the provider goes.

What should a first engagement cover?

As general practice, a bounded round that builds or checks the baseline before it runs experiments: the model sentence, a north star with its inputs, the cohort table, unit economics with their definitions and a ranked backlog, all in the business's records. The round can include a few tests aimed at the leak the baseline shows, each with its decision rule written first. It tells the business whether the provider's reading of it matches the business's own records before a larger commitment.

Terms used on this page

Frameworks

Related