Summary
Key takeaways
- More data does not improve decisions when events, definitions and commercial outcomes cannot be reconciled.
- A tracked event becomes decision-grade only after it is captured correctly, attributed consistently and validated against an operational or financial record.
- Platform conversions are useful optimisation signals, but they are not automatically equivalent to unique leads, completed sales, revenue or profit.
- Advertising budget should be scaled only after the measurement system has passed explicit tests for completeness, duplication, latency, identity and value.
A company can have sophisticated dashboards, several advertising platforms, server-side tracking, a customer relationship management system and weekly performance meetings, yet still be unable to answer a basic commercial question with confidence:
Which marketing activity created customers worth more than it cost?
The difficulty is rarely a complete absence of data. Modern acquisition systems produce data constantly. The difficulty is that the numbers describe different objects, use different rules and update on different schedules.
Google Ads may report conversions. Google Analytics may report key events and attributed conversions. Meta may show results influenced by browser and server events. The CRM contains leads, opportunities and sales. Finance records invoices, refunds, payment dates and margin. Each system can be internally coherent while disagreeing with the others.
This does not mean every discrepancy is a tracking failure. It means the organisation needs a disciplined way to distinguish an expected difference from a measurement defect.
Without that discipline, apparent precision becomes dangerous. A campaign with the lowest reported cost per lead may generate the weakest opportunities. A channel with fewer attributed conversions may influence high-value sales that close weeks later. A duplicated purchase event can make automated bidding more aggressive. A missing consent signal can reduce observable conversions and increase modelling. A sales team may change lead stages without preserving the identifiers required to connect outcomes back to acquisition.
The company does not merely have a reporting problem. It has a capital-allocation problem.
The operating principle: a measurement system is useful only when its signals can be traced, reconciled and connected to commercial value.
More data does not create measurement integrity
A common response to uncertainty is to add another tool. The company installs a dashboard, connects a warehouse, creates new events or purchases an attribution platform. These additions may be useful, but they do not repair unclear definitions or broken data lineage.
If three systems disagree because one counts form submissions, another counts unique leads and a third counts sales, centralising the numbers does not make them equivalent. It simply arranges the disagreement more attractively.
Measurement integrity begins before implementation. It starts by defining the economic event the organisation wants to understand and the operational signals that can represent progress toward it.
A marketing director may care about revenue from new customers. The advertising platform may need a faster and more frequent signal to optimise delivery. The website can observe a submitted form immediately. The CRM may identify a qualified opportunity after two days. Finance may confirm collected revenue after thirty days.
All four signals can be valid. They are not interchangeable.
The chain becomes unreliable when a stage is undefined, disconnected or assumed to mean more than it actually demonstrates.
Event, key event, conversion, lead, sale and value are different objects
The language of measurement causes avoidable confusion because familiar terms are used differently across platforms and organisations.
Google Analytics distinguishes ordinary events from key events, which represent actions important to the business. Google now uses “conversion” for important actions used to measure advertising performance and optimise bidding across Analytics and Google Ads. The change was intended to align conversion reporting between the products, but alignment between two interfaces does not automatically align either interface with the CRM or financial ledger. Google’s explanation of conversions and key events is therefore useful as a platform definition, not as a substitute for the company’s commercial definitions.
| Object | What it records | What it does not automatically prove |
|---|---|---|
| Event | A detected interaction or system occurrence | That the action was valuable or unique |
| Key event | An Analytics event marked as important | That advertising caused it or that revenue followed |
| Advertising conversion | An action used for reporting or bidding | That it equals a unique customer or final sale |
| Lead | A person or organisation entering a commercial process | That the lead is qualified, reachable or incremental |
| Opportunity | A lead accepted into a defined sales stage | That it will close or produce margin |
| Sale | A completed commercial agreement or transaction | That payment was collected or retained |
| Value | Revenue, contribution margin or another economic measure | That the value was caused by one credited channel |
A measurement specification should state which system owns each definition, which identifier connects it to earlier stages, when it becomes final and whether it is used for reporting, optimisation or both.
This distinction is especially important for lead-generation businesses. Treating every form submission as an equally valuable primary conversion can teach automated bidding systems to find people who submit forms cheaply rather than people who become profitable customers. Google’s current guidance supports importing later offline outcomes and using first-party data through enhanced conversions for leads, precisely because meaningful actions often occur after the website interaction. Google Ads describes enhanced conversions for leads as a way to supplement offline conversion data and improve attribution back to advertising interactions.
The strategic point is broader than one platform feature. The optimisation signal should move as close to realised value as data quality, volume and delay reasonably allow.
Why platform dashboards disagree
Two platforms can report different numbers without either being “wrong” in a narrow technical sense. They may be answering different questions.
Attribution models assign credit according to rules or algorithms. Google Analytics uses data-driven attribution by default and calculates contribution using the property’s available path data. The model can consider factors such as interaction order, device type and time to the key event, and attributed results may be updated after the event is recorded. Google’s attribution documentation makes clear that attribution is a model of credit allocation, not a direct reading from the financial ledger.
Conversion windows also change what can be counted. Google Ads defines a conversion window as the period after an ad interaction during which a later action can receive credit. Different actions can use different click-through, engaged-view and view-through windows. A shorter window records fewer delayed outcomes; a longer window can include more of the buying cycle. Google Ads documents these effects directly.
Counting settings create another difference. Google Ads allows an action to count every valid conversion after an interaction or only one. “Every” is often appropriate for purchases, while “one” may be appropriate when the business wants one unique lead per ad click. The official counting guidance also explains that this choice affects reported conversions and future bidding data.
Other differences arise from reporting date, timezone, channel eligibility, cross-device matching, consent, ad blockers, identifier loss, modelled conversions, invalid traffic treatment and data-processing latency. Google Analytics notes that standard report processing can take 24 to 48 hours and that figures may change during that period. Its data-freshness documentation is a useful reminder that an intraday screenshot is not necessarily a settled record.
The correct objective is therefore not to force every dashboard to display the same total. It is to explain each material difference and reconcile the systems to a common set of underlying outcomes.
The most dangerous discrepancy is an unexplained one
A ten percent difference between analytics and a CRM may be acceptable when the cause is documented. A two percent difference may be unacceptable when nobody knows what created it.
This changes how a measurement audit should be conducted. The audit should not begin with a promise to make every number match. It should begin with a reconciliation model.
For each important outcome, identify the real-world action, the source-of-truth record, the event or import representing it, the unique identifier, the timestamp and timezone, the counting rule, the attribution rule, the expected processing delay and the acceptable discrepancy.
The specification should remain compact enough to use. The purpose is not administrative theatre. It is to make differences explainable.
The Intent & Proof measurement-integrity framework
Intent & Proof uses five connected stages:
The sequence is deliberate. Signal defines what matters. Capture records it. Attribution assigns credit. Validation compares the record with reality. Decision limits the conclusions the organisation is entitled to draw.
An error at an early stage contaminates every stage after it. Better attribution cannot repair an event that fires twice. CRM reconciliation cannot recover an identifier that was never captured. A polished dashboard cannot convert an unqualified lead into revenue.
Signal: define the action before tracking it
The first question is not “Which tag should fire?” It is “Which observable action represents the decision we need to make?”
A business may need several layers of signal:
| Signal level | Example | Suitable use |
|---|---|---|
| Attention | Engaged visit or product view | Experience analysis and audience building |
| Intent | Pricing view, booking start or detailed enquiry | Funnel diagnosis and secondary optimisation |
| Commitment | Submitted application, booked consultation or checkout | Primary optimisation when downstream data is unavailable |
| Qualification | Sales-accepted lead or verified appointment | Higher-quality optimisation and budget comparison |
| Economic outcome | Purchase, collected revenue or contribution margin | Commercial reporting and value-based optimisation |
The nearest signal is not always the best signal. Revenue is economically meaningful but may arrive too slowly or too infrequently for a campaign to learn. A form submission is immediate and frequent but may be weakly related to value.
The practical task is to select the strongest signal that remains sufficiently frequent, timely and reliable for its intended use. Earlier signals can remain visible without receiving the same status.
This is where marketing and sales must agree on definitions. A qualified lead cannot mean “someone sales liked” on Monday and “anyone with a phone number” on Friday. Stage criteria should be explicit enough that the same record would be classified consistently by another person.
The principle connects directly with our guide to turning local searches into real medical appointments. A profile view, website click, call, booking and attended appointment describe different stages. Optimising local acquisition around the first observable action while ignoring appointment quality can produce impressive activity and disappointing economics.
Capture: verify that the event represents the action
Once the signal is defined, the implementation must record it correctly.
A capture test asks whether the event fires at the correct moment, with the correct parameters, once for each valid action and not for invalid actions. It also checks whether the event survives realistic conditions such as mobile navigation, form errors, payment retries, consent choices, browser restrictions and server responses.
Google Analytics provides Realtime and DebugView specifically to verify collection during implementation. DebugView displays events and user properties from a debug device as they are collected, which makes it useful for inspecting event names, sequence and parameters. Google’s DebugView documentation should be treated as the beginning of validation, not the end. Seeing an event arrive proves transmission. It does not prove uniqueness, business validity or reconciliation.
Duplicate purchases are a familiar example. A confirmation page can reload, a browser event can be sent alongside a server event, or two tag-manager triggers can respond to the same action. Google Ads recommends using a unique transaction or order ID so repeated records for the same transaction can be recognised and deduplicated. The identifier must be generated dynamically and remain consistent across tag and server uploads. Google’s transaction-ID guidance also distinguishes deduplication from the separate choice between counting one or every conversion.
For lead generation, uniqueness is more complicated. Two forms from one person may be duplicates, separate service enquiries or legitimate follow-ups. The business must decide which object it wants to count and where that decision is made. A browser tag cannot reliably infer whether two records became one sales opportunity. The CRM usually needs to perform that reconciliation.
Meta similarly recommends using the Conversions API alongside the pixel to create a more reliable connection between website, server, CRM and offline events. Meta states that Conversions API data can be less affected by browser loading errors, connectivity problems and ad blockers, while also supporting later customer-journey actions. Meta’s official Conversions API overview is useful here, but adding a server connection still requires consistent event definitions and deduplication. Two unreliable feeds do not form one reliable system merely because one travelled through a server.
Attribution: understand the rule before accepting the credit
Attribution answers a narrower question than many dashboards imply. It assigns credit among observable interactions according to a model.
It does not automatically establish what would have happened without the advertising.
This distinction matters when several platforms claim the same sale. Google, Meta and another channel may each have observed an eligible interaction within their respective windows. Each platform can assign credit under its own rules. Adding the reported conversions together can therefore exceed the number of actual sales without any literal duplicate transaction inside an individual platform.
A central analytics model can create a more consistent cross-channel view, but it still depends on observable paths, identity resolution and modelling assumptions. A CRM can show which leads became customers, but its “source” field may reflect first touch, last touch, self-reporting, a salesperson’s edit or whichever integration wrote last.
The organisation should document three separate concepts.
Operational attribution supports recurring reporting and campaign optimisation.
Commercial reconciliation connects attributed records to unique opportunities, sales and value.
Causal measurement estimates the incremental effect of marketing through experiments or other counterfactual methods.
Confusing these concepts creates false certainty. Attribution is useful, but it is not a laboratory result wearing a dashboard.
A practical measurement system may use platform attribution for bidding, a consistent cross-channel model for management reporting and periodic experiments for larger allocation decisions. The models need not be identical. Their roles must be explicit.
Validation: connect measurement to a source of truth
A conversion becomes commercially meaningful when it can be matched to an operational or financial outcome.
For ecommerce, the source of truth may be the order system, adjusted for cancellations, refunds, tax, shipping and margin. For lead generation, it may be the CRM with documented stage criteria. For appointments, it may be the booking system combined with attendance and treatment value. For subscription businesses, it may be activated accounts, retained revenue or contribution after acquisition cost.
Validation requires identifiers. Useful identifiers may include transaction ID, lead ID, opportunity ID, booking ID and customer ID. Advertising click identifiers and hashed first-party data can help platforms match later outcomes, but the business still needs its own stable internal key.
Reconciliation should be performed by cohort and identifier, not only by comparing totals. Two systems can display the same monthly number while containing different underlying customers. That coincidence is not integrity.
Start with a sample of individual records. Trace a real conversion from the originating interaction through the website event, platform record, CRM stage and final economic outcome. Then test the reverse direction: select completed sales and determine how many can be connected back to their acquisition records.
The two directions reveal different failures. Forward tracing exposes false positives and duplicated or low-quality conversions. Reverse tracing exposes missing capture, identifier loss and uncredited outcomes.
Enhanced conversions and offline imports can improve platform matching. Google explains that enhanced conversions use hashed first-party customer data and offline information to supplement existing conversion measurement. Its enhanced-conversions overview describes the matching role clearly. The business should still compare imported records with the CRM totals and monitor diagnostics, because a successful upload is not the same as complete or correctly classified data.
Modelled data should be labelled, not feared or disguised
Privacy choices, browser restrictions, cross-device behaviour and technical loss mean that some outcomes cannot be directly observed. Platforms may use modelling to estimate missing conversions.
Google Analytics states that modelled key events estimate online actions that cannot be observed directly and that attributed data can continue updating after the event is first recorded. The official explanation of modelled key events describes modelling as a way to improve reporting without identifying individual users.
The correct response is neither to reject every modelled result nor to treat it as equivalent to an observed transaction.
Management reporting should distinguish, where available, directly observed events, imported or matched offline events, modelled events, unattributed outcomes and unresolved discrepancies.
This allows a decision-maker to understand the evidence base behind a number. A blended total may still be useful for campaign optimisation, while a finance reconciliation may rely only on confirmed records. Again, the same number does not need to perform every job.
Data thresholds and sampling can also affect what appears in reports. Google Analytics applies privacy thresholds in some contexts and provides a data-quality indicator when reports use thresholding or sampling. Google’s data-quality documentation should be checked before interpreting small differences as market behaviour.
Decision: specify what the evidence permits
The final stage is governance.
A measurement system can be technically sophisticated and still produce poor decisions if nobody defines the threshold for action. Teams often react to small short-term movements, compare incomplete periods or scale spend after a dashboard improvement that has not reached the CRM.
Each decision should state the metric used, the source and maturity of the data, the comparison period or cohort, the expected delay, the acceptable discrepancy, the minimum sample, the commercial constraint and the reversal condition.
For example, “increase campaign budget by 20 percent” might require stable qualified-opportunity cost across two complete cohorts, no unresolved tracking incident, an acceptable match rate between platform and CRM, and sufficient sales capacity to handle additional volume.
The requirement is not bureaucratic caution. It prevents a metric from being interpreted differently depending on which conclusion somebody already preferred.
Common failure modes and what they look like
Measurement failures usually leave patterns.
| Symptom | Likely causes | First validation |
|---|---|---|
| Sudden conversion increase without sales movement | Duplicate firing, new primary action, attribution change | Inspect change history and trace individual events |
| Platform leads exceed CRM leads | Duplicate forms, missing CRM creation, different counting rule | Match lead IDs and timestamps |
| CRM sales exceed attributed conversions | Identifier loss, consent, long sales cycle, offline channel | Reverse-trace completed sales |
| Revenue differs between analytics and finance | Refunds, tax, shipping, currency, timing or missing IDs | Reconcile order IDs and accounting treatment |
| One channel improves after a model change | Attribution redistribution rather than behavioural change | Compare model settings and paths |
| Conversion rate drops after consent changes | Lower observability rather than immediate demand collapse | Segment consented traffic and confirmed outcomes |
| Smart bidding deteriorates after a tracking update | Changed event quality, lag or duplicate imports | Validate primary actions and recent uploads |
| Reports keep changing after month end | Processing, modelling, late CRM stages or refunds | Define a reporting-close schedule |
The most useful investigation asks what changed in the system before assuming that customer behaviour changed. Humans enjoy blaming the market for configuration errors because the market cannot attend the meeting to defend itself.
Technical correctness is not enough
A tag can fire exactly as specified while the specification remains commercially wrong.
Suppose a form submission is recorded once, includes all required parameters and appears correctly in every platform. The implementation is technically sound. But if half the submissions are support requests, job applications or spam, the event is unsuitable as the primary optimisation target.
Conversely, a technically imperfect system may still support a narrow decision if its limitations are stable and understood. A business might know that ten percent of phone enquiries cannot be matched to campaigns, yet still compare qualified opportunity cost across large cohorts when the missing share is consistent.
Measurement integrity is therefore not a binary label. It is fitness for a decision.
The audit must review definitions, incentives and workflows alongside tags and APIs. Sales-stage discipline, form routing, call handling, refunds, manual edits and data ownership can affect measurement as much as JavaScript.
This is also why proof architecture matters. Our guide to turning client work into credible proof argues that evidence becomes useful when context, constraints, intervention and limitations remain visible. Measurement follows the same standard. A performance claim without definitions, time logic and limitations is not decision-grade proof.
Build a reconciliation table before building another dashboard
A minimal reconciliation table can outperform a sophisticated visualisation when the organisation has not yet aligned its data.
For each commercial outcome, include the internal identifier, event timestamp, platform timestamp, acquisition identifiers, attributed source, CRM stage, order or opportunity value, status, refund or cancellation state and last validation date.
The table should answer three questions:
- Did the event occur?
- Was it represented correctly across systems?
- Did it become the outcome used in the decision?
Once these questions can be answered reliably, a dashboard can summarise the system. Before that, the dashboard often hides the exact records required to diagnose it.
Metrics for measurement integrity
The audit needs metrics about data quality, not only campaign performance.
| Integrity metric | What it reveals |
|---|---|
| Capture rate | Share of confirmed outcomes represented by the intended event |
| Match rate | Share of events connected to the source-of-truth record |
| Duplicate rate | Share of records representing an already counted outcome |
| False-positive rate | Share of tracked conversions that fail the business definition |
| Stage-completion rate | Share of leads progressing to qualification, sale or retained value |
| Value coverage | Share of outcomes carrying a valid economic value |
| Latency | Time between real action, platform availability and final validation |
| Unattributed rate | Share of confirmed outcomes without usable acquisition linkage |
| Reconciliation variance | Difference between systems after definitions and timing are aligned |
| Incident recovery time | Time required to detect, correct and validate a measurement failure |
These metrics should be segmented by important source, device, conversion action and implementation path. An overall match rate can conceal a broken mobile form or a server event that fails only for one payment method.
Targets should reflect the business model. A high-volume ecommerce system can demand very strong transaction reconciliation. A long-cycle enterprise sale may accept lower attribution coverage while requiring perfect opportunity and revenue records. The right standard is the one that protects the decisions being made.
A 30-day measurement-integrity plan
The first month should create a reliable map, repair the highest-risk failures and establish a repeatable validation process.
During the first five days, inventory every primary and secondary conversion used in analytics, advertising, CRM and management reports. Record its definition, owner, trigger, count setting, value, window and destination.
The next stage traces representative records in both directions. Test normal and abnormal paths: successful submissions, validation errors, repeated clicks, page reloads, payment retries, consent choices, mobile devices and delayed CRM updates.
Repair should prioritise errors capable of changing optimisation or budget decisions. A decorative dashboard field can wait. A duplicated primary purchase or an unqualified form used for automated bidding cannot.
Reconciliation then compares event-level records with CRM, order and finance data. The team calculates capture, match, duplicate, false-positive and latency metrics by source and action.
The final stage establishes ownership, monitoring and change control. Every new conversion action or implementation change should have a written definition, test evidence, release date, owner and post-release validation period.
| Period | Work | Deliverable |
|---|---|---|
| Days 1–5 | Define outcomes, signals and system ownership | Measurement specification |
| Days 6–10 | Trace representative journeys and edge cases | Event-lineage map |
| Days 11–18 | Correct high-risk definitions and implementation defects | Prioritised fix release |
| Days 19–24 | Reconcile platforms with CRM, orders and finance | Integrity scorecard |
| Days 25–30 | Establish monitoring, ownership and decision rules | Measurement governance plan |
The objective is not a ceremonial “tracking complete” document. It is a system that can demonstrate when it is working and reveal when it is not.
When should advertising budget be increased?
Budget should not wait for perfect attribution. Perfect attribution is not available, and pretending otherwise mainly benefits vendors of expensive diagrams.
Budget can be increased when the evidence is strong enough for the specific decision.
A responsible scaling decision usually requires stable signal quality, technical validation, commercial reconciliation, mature data, sufficient capacity and economics, and a defined rollback rule.
The decision can still contain uncertainty. It should not contain unexplained uncertainty large enough to reverse the conclusion.
The Measurement Integrity Audit
The first Intent & Proof product for this problem is the Measurement Integrity Audit.
It begins with a conversion and value map rather than a tag inventory. The audit identifies which user and commercial actions influence budget, how they are captured, which systems receive them, how credit is assigned and where the records are validated.
The deliverables should include an event-lineage map, conversion-definition register, discrepancy matrix, technical test results, CRM or order reconciliation, integrity metrics, prioritised fixes and a decision-ready measurement specification.
The audit is not limited to Google Analytics. It may involve Google Ads, Meta, other acquisition platforms, tag management, server-side collection, consent systems, forms, calls, booking tools, ecommerce, CRM and finance data. The scope is determined by the decision chain, not by whichever interface currently has the most colourful charts.
The broader Measurement Integrity System adds ongoing monitoring, release controls, recurring reconciliation and alerts when important rates move outside expected ranges.
Final principle: measurement should constrain claims
A measurement system should not exist to confirm that marketing performed well. It should define what the organisation can legitimately claim.
A dashboard may suggest that a campaign generated one hundred conversions. The company should be able to explain how many were unique, how many met the commercial definition, how many became customers, what value they produced, which outcomes were modelled or delayed and which uncertainty remains.
That explanation does not weaken performance. It makes performance defensible.
When measurement integrity is weak, more advertising amplifies two things at once: acquisition activity and the cost of being wrong.
The sensible sequence is therefore not traffic first, tracking later.
It is:
More data is not the objective.
Better decisions are.
Sources, further reading and methodology
This guide combines the Intent & Proof Signal → Capture → Attribution → Validation → Decision framework with current first-party platform documentation.
Google Analytics: Conversions versus key events explains the current distinction between important behavioural events and conversions used for advertising measurement.
Google Analytics: Get started with attribution describes available attribution models and the operation of data-driven attribution.
Google Ads: Conversion windows and conversion counting options document two settings that can materially change reported performance.
Google Ads: Use transaction IDs to minimise duplicates explains the role of unique identifiers and the distinction between deduplication and conversion-count settings.
Google Analytics: DebugView, data freshness, modelled key events and data quality provide the operational context required to validate and interpret reported data.
Google Ads: Enhanced conversions and enhanced conversions for leads describe the use of hashed first-party and offline data to improve matching and measurement.
Meta: About Conversions API describes direct server, website, CRM and offline event connections used for optimisation and measurement across Meta technologies.
The article also applies the broader Intent → Proof → Action → Measurement operating model developed in our guides to turning local searches into real medical appointments and turning client work into credible proof. Platform documentation is used to describe product behaviour. The commercial framework, reconciliation method and decision rules are Intent & Proof analysis.
Frequently asked questions
Why do Google Ads and Google Analytics show different conversion numbers?
The systems can use different attribution rules, conversion windows, reporting dates, channel eligibility, identity signals, processing times and counting settings. A difference is not automatically an error. It becomes a problem when the organisation cannot explain the difference or reconcile both systems with the same underlying commercial outcomes.
What is a conversion tracking audit?
A conversion tracking audit examines the complete path from a real customer action to the event recorded in analytics and advertising platforms, the attribution applied to that event, and the final validation against CRM, ecommerce, booking or financial records. It identifies missing events, duplicates, inconsistent definitions, broken identifiers, unsuitable conversion settings and reporting gaps that could distort budget decisions.
What should a business count as a primary conversion?
A primary conversion should represent the closest reliably observable action to commercial value that is suitable for campaign optimisation. For ecommerce this may be a completed purchase with a valid transaction ID and value. For lead generation it may be a qualified lead, booked appointment or accepted opportunity rather than every form submission. Earlier actions can remain secondary signals without being treated as equal outcomes.
How can duplicate conversions be prevented?
The implementation should fire only when the underlying action occurs and use a stable unique identifier wherever the platform supports one. Purchases should carry unique transaction or order IDs, while browser and server events should use consistent identifiers and deduplication rules. Counting settings must also match the business model because preventing duplicate events is different from choosing whether one or every valid conversion after an ad interaction should count.
Is attribution the same as proving that advertising caused a sale?
No. Attribution assigns credit according to a rule or model using observable interactions. It does not by itself prove incremental causality. Attribution is useful for operational reporting and optimisation, while experiments, holdouts, geographic tests and other causal methods are better suited to estimating what would have happened without the advertising.
When is conversion tracking reliable enough to scale advertising?
Tracking is reliable enough for a specific decision when the important event is defined, tested, deduplicated, reconciled with the source-of-truth system, monitored for delays and failures, and understood across stakeholders. Perfection is not required, but the remaining uncertainty must be documented and small enough that it would not reverse the budget decision.
