Web design agencies attribution methods for buyers

This guide helps people using a web design agency directory assess how agencies describe measurement and reporting. Web design agencies attribution methods split reported credit across channels, while causal evaluation asks what would have happened without the change. The two questions look alike in a client report and need very different evidence. The case below is held constant on purpose, so each method is read against the same fields.

What to take away

Attribution divides credit for recorded events, while causal evaluation asks what would have happened without the change.
Rule-based models fit allocation decisions, such as which channel receives reported credit.
Only a randomised or quasi-experimental design supports an effect estimate, and only when assignment and outcome capture hold.
Hold one case constant and fill the same fields for every method before you compare.
Report the eligible population beside the number, or the result cannot be reproduced.

Define the case before comparing methods

Start with the decision, not the model. A web design agency's England retailer client wants the reported credit for an October 2026 checkout change split across channels. A rule-based method can answer that allocation question. For directory buyers, the same distinction helps test a listing's measurement claims: is the agency describing credit allocation or claiming that a change caused an outcome?

When comparing directory listings, use the same questions for each agency: what event is measured, which attribution rule is used, what evidence supports any causal claim, and whether consent loss and exclusions are disclosed. A listing's wording is not evidence by itself; use the method and gates below to frame follow-up questions, not to rank agencies without comparable records.

Choose method by decision

Does the decision concern allocation?

Yes

Use a rule-based method

Does it need an effect estimate?

Write down the completion event, the eligible population, numerator and denominator, the unit of analysis, and the period with its timezone. Add duplicate rules, channel and device exclusions, the source system and query version — the Google Analytics 4 property and its attribution report version, for example — and the data owner.

Record consent coverage and the loss from blocked or refused tracking. The ICO's cookies guidance explains why storing or accessing information on a device needs its own permission check in the UK. Our competitor research on agency data handling shows how to run the surrounding desk work without hoarding personal data.

Compare the methods on the same fields

Compare the methods

Question answered

Last-touch rule: Which recorded source receives credit immediately before the event?
Fixed multi-touch rule: How should recorded credit be divided among included touches?
Contribution analysis: Is the observed pattern consistent with a stated theory of change?
Randomised evaluation: What difference arises under assigned exposure versus control?
Quasi-experimental evaluation: What difference can a non-random comparison estimate?

Required evidence

Last-touch rule: Defined window, source taxonomy, identity and duplicate rules
Fixed multi-touch rule: Common exposure window and predeclared weights
Contribution analysis: Delivery evidence, sequence, alternatives and corroborating sources
Randomised evaluation: Ethical feasible assignment, outcome capture and valid analysis
Quasi-experimental evaluation: Credible comparison, baseline and design-specific diagnostics

Main assumption

Last-touch rule: The chosen final touch deserves all assigned credit
Fixed multi-touch rule: Editorial weights reflect the reporting purpose
Contribution analysis: Rival explanations can be examined credibly
Randomised evaluation: Assignment and interference conditions hold
Quasi-experimental evaluation: The chosen counterfactual assumptions are defensible

Output and limit

Last-touch rule: Reproducible allocation, not causal effect
Fixed multi-touch rule: Descriptive shares that shift with the rule
Contribution analysis: Bounded contribution account, not numerical impact
Randomised evaluation: An effect estimate for the studied population, with uncertainty
Quasi-experimental evaluation: Conditional estimate, credibility depends on assumptions

HM Treasury's Magenta Book separates process, impact and value-for-money evaluation. Its analytical methods annex covers experimental and quasi-experimental designs. Both apply to government evaluation, so a commercial website project needs a specialist adaptation.

Run the benchmark comparability checks before you assemble the table, so a rule-based share is never placed beside a causal estimate.

Example: one case filled in

Illustrative figures for a hypothetical England retailer client of a web design agency, not a benchmark or a client result.

One checkout step changed, 1 to 31 October 2026. Eligible sessions were 12,400 before and 11,900 after. Completed enquiries were 620 before and 744 after, an observed rise of 124.

Under a last-touch rule, paid search receives credit for 431 of the 744 post-change completions, organic search 201 and direct 112. A fixed multi-touch rule with predeclared weights moves paid search to 372 and organic to 260.

One case filled in

Method: Last-touch, then fixed multi-touch
Paid search completions credited: 431, then 372
Organic completions credited: 201, then 260
Effect estimate: Not estimable from this design

The 124 is an observed difference, not incremental gain. No assignment, no comparison group, and consent loss was never quantified. That is why the causal rows carry no figures for this case: without a design, the same data cannot produce an effect estimate.

Apply non-compensating gates

A gate is non-compensating when one failure cannot be offset by strength elsewhere. Run five checks in order and record the date of each.

Reconstruct the source query and its version from the platform audit log.
Quantify consent loss by comparing eligible sessions either side of the consent change.
Test cross-device duplication against a logged-in cohort whose true size you know.
List concurrent campaigns, seasonality and product changes that overlap the window.
Mark any failed field as failed, then stop before calculating a share or a return.

On the case above, consent loss was never quantified, so the second gate stays failed. Rule-based models can still support operational reporting while these failures are prominent. Keep them out of any claim about incrementality.

Privacy, PECR, security, accessibility and contract review stay separate from the method choice. The ICO enforces the UK GDPR and PECR; accessibility on a client site is measured against the WCAG 2.2 standard and sits under the Equality Act 2010; security baselines such as Cyber Essentials belong in the contract rather than in the measurement.

Before you act

Where an agency passes these numbers on, the commercial contract records settle who owns the query history and the audit trail.

Define the completion event and eligible population before comparison.
State the decision, assumptions and uncertainty before choosing a method.
Name the ICO and PECR sign-off owner in England before any tracking tag goes live.
Check measurement loss, cross-device exposure and concurrent campaigns.
Define any value or cost field, including VAT, before a return calculation.
Revalidate every assumption before the planned 2027 analysis.

Common questions

How do attribution and causal evaluation differ?

Attribution allocates credit for recorded events. Causal evaluation asks what would have happened without the change, and needs a design that supports that answer.

Which method fits an allocation decision rather than an effect estimate?

Use a rule-based model when the decision splits reported credit. Use a randomised or quasi-experimental design when the decision needs a difference.

What does the worked example show?

One case produces two credit splits under two rules, and no effect estimate at all. Changing the rule changes the reported number, not the data.

What if the query history cannot be reconstructed?

Place the report on hold. Without the source query and its version the comparison cannot be reproduced, so revalidate before the 2027 analysis.