An illustrative 8.0x Google Ads ROAS can contain only 5.0x of directly observed value and 3.0x of platform estimation. The estimated 3.0x is not automatically fiction, but treating it like a CRM-confirmed sale is a category error. Signal loss did not make cookieless ROAS tracking unmeasurable; it made ROAS partly modelled, and the practical job is knowing which half of your number is measured and which half is estimated.
That distinction should govern every budget review in 2026. Reported ROAS is a composite metric, so interrogate how its numerator was assembled before approving more spend.
Cookieless ROAS tracking is a split number
Google Ads can include modelled conversions when direct observation is unavailable and modelling requirements are met. An observed conversion is backed by an event or imported outcome the platform received. A modelled conversion is inferred from aggregate patterns because the corresponding event could not be observed directly.
Both can carry conversion value. Both can appear inside the same ROAS figure.
Take an illustrative UK IT consultancy assessing one month. Its five labelled inputs are:
- Monthly Google Ads spend: £8,000
- Total reported conversions: 20
- Total reported conversion value: £64,000
- Observed portion: 12 conversions worth £40,000
- Modelled portion: 8 conversions worth £24,000
The arithmetic is:
Reported conversions = 12 observed + 8 modelled = 20
Reported conversion value = £40,000 observed + £24,000 modelled = £64,000
Reported ROAS = £64,000 ÷ £8,000 = 8.0x
Observed ROAS component = £40,000 ÷ £8,000 = 5.0x
Modelled ROAS component = £24,000 ÷ £8,000 = 3.0x
Observed count share = 12 ÷ 20 = 60%
Modelled count share = 8 ÷ 20 = 40%
Observed value share = £40,000 ÷ £64,000 = 62.5%
Modelled value share = £24,000 ÷ £64,000 = 37.5%The dashboard therefore shows 8.0x reported versus 5.0x observed. Its composition is 5.0x observed versus 3.0x modelled, while the conversion count is 60% observed versus 40% modelled.
That is the split a budget owner needs.
Google Ads may provide aggregate information about modelling impact without identifying which individual conversions were modelled. Where a credible aggregate split is available, record it. Where it is not available, mark the modelled share as unknown rather than assuming 0%.
Observed also does not mean commercially verified. A perfectly observed form submission can still be spam, an existing customer or an unsuitable company. The honest limit here is that modelled share measures exposure to estimation, not total measurement error.
The next check is whether recorded value survives contact with the CRM. That commercial distinction sits behind separating platform ROAS from the return the business actually received.
ROAS is usable only when its observed and modelled components are declared.
Why cookieless ROAS tracking tools disagree
HubSpot records identified contacts and deal stages. Google Ads records attributed ad conversions, including eligible modelling. GA4 records website or app events and may also model missing activity.
Those are different evidence sets, not three replicas of one ledger.
Take a separate illustrative monthly reconciliation: Google Ads reports 26 conversions, GA4 records 21 key events, and HubSpot contains 17 identified enquiries, of which 12 became sales-accepted leads. The meaningful comparisons are 26 platform conversions versus 17 identified enquiries, then 17 enquiries versus 12 sales-accepted outcomes.
Calling the nine-conversion Google Ads–HubSpot gap a tracking failure would be premature. Google Ads may include modelled activity, multiple conversion actions or outcomes that never produced a valid CRM record. HubSpot cannot report an anonymous journey that never supplied an identity.
A tool that reports what happened and a tool that estimates what probably happened will never reconcile. That principle is the main reason apparently competent tools produce different totals.
Do not manipulate settings until every dashboard matches. First align the conversion definition, deduplication rule and eligible campaign scope. Attribution-window mechanics remain a separate issue, covered in our guide to how reporting windows reshape ROAS.
The same discipline applies when several platforms claim the same sale. Use a defined method for comparing channel return without counting one outcome repeatedly, rather than adding every platform’s attributed revenue together.
Tool disagreement reflects different evidence, units and inference.
A budget rule for modelled conversions
Use the higher of the modelled count share and modelled value share. A few inferred high-value conversions can make the count share look safe while dominating the commercial result.
Modelled exposure = higher of:
modelled conversions ÷ reported conversions
or
modelled value ÷ reported conversion valueOur three-band budget rule is deliberately conservative:
- Green — 0% to 20% modelled: modelled conversions are acceptable input to a reversible budget change of up to 20%, provided both validation gates pass.
- Amber — above 20% to 40% modelled: use platform ROAS as supporting evidence only. Cap the budget change at 10% and reject an increase when CRM-qualified pipeline moves in the opposite direction.
- Red — above 40% modelled or unknown: modelled conversions are not acceptable as the sole basis for scaling or cutting spend. Hold the baseline and decide using observed sales-accepted leads or CRM-confirmed pipeline.
Every band has two validation gates: the primary conversion must represent a sales-relevant action, and CRM-confirmed qualified pipeline must agree with the platform’s direction. Failure at either gate overrides the colour.
If fewer than 20 sales-validated outcomes exist in the decision set, treat the CRM comparison as weak and avoid an irreversible reallocation. More rows in Google Ads do not compensate for too little commercial evidence.
Current B2B Google Ads benchmarks can flag an abnormal click or lead cost, but they cannot validate modelled revenue. This separation between platform evidence and commercial evidence is also central to our Google Ads management for B2B growth teams.
Our claim is falsifiable: ROAS with more than 40% modelled value is too fragile to drive scaling alone; prove it wrong with three consecutive closed sales cohorts reconciling within 10% of CRM-confirmed value.
Budget confidence must fall as the modelled share rises.
If the path from ad click to signed revenue still has gaps, Actualyse will trace each hand-off with you — book a call
What Consent Mode and first-party collection do not recover
Consent Mode changes tag behaviour according to the user’s consent choice. Where requirements are met, Google can use available aggregate signals to model gaps. It does not recreate a withheld browser history, reveal an individual who declined consent or turn an inferred conversion into an observed one.
Enhanced conversions can improve matching when lawful, consented first-party identifiers correspond with platform data. Hashing an identifier does not create a match where no usable identity or permission exists.
First-party collection remains valuable because it records business outcomes after someone identifies themselves. Build a five-field measurement spine for every paid enquiry:
- A persistent CRM record ID
- The submitted form and exact timestamp
- The landing page and recorded acquisition source
- The consent state and available advertising click identifier
- The latest lifecycle stage and associated £ value
If fewer than 90 of the latest 100 paid-enquiry records contain all five fields or a documented reason for an unavailable identifier, repair the form-to-CRM mapping before raising spend.
Even complete first-party records leave four blind spots: the unconsented pre-conversion journey, anonymous cross-device activity, exposure inside another platform and intent from people who never identify themselves. Neither Consent Mode nor CRM hygiene restores pre-2023 visibility.
Event design must also reflect the buying journey. In our Lanteria work, a broad software capability set was routed for different stakeholder audiences. AfriCap Hub required catalogue, filtering and registration paths. Our case studies of multi-audience and registration journeys document those architecture decisions, not performance claims.
Collapsing every action into “lead” would discard the intent those journeys were designed to reveal. GDPR and PECR remain boundaries for collection, not inconveniences to route around.
Consent infrastructure narrows gaps without reconstructing lost journeys.
What does not work for cookieless ROAS tracking
We repeatedly see four popular fixes create cleaner reports without creating better evidence.
Forcing GA4 and Google Ads to match
Changing configuration until both totals align removes a useful diagnostic difference. GA4 and Google Ads can use different evidence, scopes and inference, so equality does not prove accuracy.
Perfecting every UTM parameter
UTMs label a visit. They do not prove that advertising caused an opportunity, identify every returning visitor or confirm that sales accepted the lead. Perfect campaign naming cannot repair a missing CRM outcome.
Adding more tags and dashboard precision
Google Tag Manager can deploy measurement tags; it cannot make a weak conversion definition commercially meaningful. A new Looker Studio dashboard displaying ROAS to two decimal places merely gives uncertain inputs sharper typography.
Optimising the consent acceptance rate
Banner dark patterns may increase recorded acceptance, but they do not create trustworthy permission or recover activity from people who still decline. They also add avoidable GDPR and PECR risk without validating revenue.
Cosmetic tracking fixes cannot validate B2B revenue.
FAQ
Should a phone call count as revenue?
Apply a provisional-call rule: count a correctly routed sales call lasting at least 60 seconds as a lead, then replace it with a CRM outcome within 14 days. Without a validated outcome, assign £0 revenue rather than an invented sales value.
What value should a demo request carry?
Use a stage-value decision only after at least 20 comparable opportunities have reached won or lost. Before that threshold, count the demo as a conversion but assign £0 realised revenue; otherwise a speculative value will dominate ROAS.
Should brand and non-brand campaigns share one ROAS?
Separate them whenever brand represents more than 20% of spend or reported conversion value. Use the non-brand result for acquisition-budget decisions because combined ROAS can conceal how much demand paid search merely captured.
When should finance and marketing reconcile the numbers?
Reconcile monthly and open a measurement incident when platform-reported value differs from CRM-confirmed eligible value by more than 15% across two consecutive closed cohorts. Assign one owner to resolve conversion definitions before anyone changes bids.
Named thresholds turn uncertain attribution into controlled decisions.
Summary
Five operating rules determine whether the number deserves budget authority:
- Split every reported ROAS into observed and modelled components before changing spend.
- At 0%–20% modelled exposure, permit controlled changes; above 20%–40%, cap them at 10%; above 40%, do not scale from platform ROAS alone.
- Treat an unavailable modelled share as red, never as an assumed 0%.
- If fewer than 90 of 100 paid-enquiry records contain the five required fields, repair collection before increasing spend.
- Use Consent Mode and first-party data to reduce uncertainty, never to claim lost journeys were recovered.
Cookieless ROAS earns budget authority only after its estimate is exposed.
Actualyse builds measurement and attribution setups that tie B2B ad spend to real revenue. Book a call to talk through where yours stands.
