Three systems, three numbers, and nobody is lying
The pattern repeats in almost every performance review. The ad platform reports one lead count, the analytics property reports a smaller one, and the CRM has no record of most of them. The meeting becomes an argument about which number to believe and ends with an opinion instead of a decision.
All three are correct inside their own definitions. An ad platform counts a conversion it can attribute to a click, within a window it sets, from data it can see on its side. Analytics counts an on-site event under a different attribution model. The CRM recognises only records that actually arrived, survived deduplication and passed whatever validation sits on the form. Three different objects, three time windows, three attribution rules.
Unification is therefore not the project of making those figures agree. They will not, and anyone promising otherwise has not run one of these programmes. It is the project of deciding which system settles which decision, threading a single key from anonymous visit through to recognised revenue, and knowing exactly where that thread breaks and why.
Model the objects before you shortlist the tools
When we are brought into a stalled integration, the cause is almost never a broken API. It is that nobody ever wrote down what a lead is. Marketing counts form submissions, sales counts conversations worth having, and the board reads both numbers assuming they describe the same thing.
Start with one page describing the chain: anonymous session, identified contact, account, opportunity, recognised revenue. Give each object four lines — which system is the system of record, which key ties it to the object before it, the exact moment it is created, and who is permitted to change it.
One rule does more work here than any tool decision: no object has two owners. The CRM owns contacts and opportunities. Analytics owns sessions and on-site behaviour. The ad platform owns clicks and cost. Everything else receives a read-only copy. The moment two systems write to the same field, you have signed up for a permanent reconciliation job that nobody budgeted for.
The hard joint is contact to opportunity, because it crosses the marketing and sales boundary. Ask the sales lead to write the entry criterion for the first pipeline stage in testable language — not 'the prospect is interested', but a condition two people could apply to the same record and reach the same answer. If that sentence cannot be written, no amount of integration will rescue the reporting.
The lead-source contract: written once, never overwritten
Integration without field discipline produces data that is worse than none, because it looks authoritative. The working minimum we implement: first-touch source, medium and campaign; the same three for last touch; first landing page; the click identifier from every platform you spend on; a form identifier; consent state; and a timestamp on both touches.
Two rules govern them. First-touch fields are written once at record creation and never overwritten, however many times the person returns. Last-touch fields update on each subsequent conversion. Collapsing the two is the most common reason a company discovers one quarter that every enquiry now appears to come from branded search.
Populate those hidden fields from a first-party cookie holding the earliest values the browser saw, not from the URL at the moment of submission. Someone who arrives from a paid campaign, returns a fortnight later by typing your name, then completes the form genuinely came from the campaign. Reading the live URL credits your own brand and hides what actually created the demand.
Then lock them. Attribution fields should be read-only in the sales interface. If you want the qualitative signal — and it is worth having — add a separate 'how did you hear about us' field for a rep or the prospect to complete, and use it for colour, never for arithmetic. One field serving both purposes serves neither.
Close the loop: send the sales outcome back to the spend
While ad platforms only learn from form fills, they optimise toward reducing the cost of a form fill — not toward the quality of the customer behind it. Closing the loop means feeding back what happened next: whether the record qualified, became an opportunity, and closed. The major platforms support this through offline conversion imports and server-side event APIs.
Four preconditions before it works. The click identifier must be stored on the contact record at creation. The stage you send back must be defined once and left alone. The lag between click and outcome must fall inside the import window the platform accepts. And the event must occur often enough for the bidding system to learn from it.
That last condition kills most attempts. If your sales cycle is long and you close a handful of deals a quarter, sending 'closed won' gives the platform almost nothing to work with. The practical route is to send back the earliest stage that is both frequent enough monthly and demonstrably correlated with revenue in your own history — usually a sales-accepted stage — and keep closed-won in internal reporting, where you are measuring impact rather than training an algorithm.
Send nothing you cannot defend. Every event leaving for an external platform should be reconstructable from a named field on a real record, so that when finance questions the figure six months later you can show precisely where it came from.
Identity and consent come before the first connector
Unification rests on a key that survives across systems. In practice: a normalised email address — lower-cased and trimmed — as the contact key; the email domain as the first-pass key for grouping contacts under one account; and your internal user ID wherever the customer signs in, which is also what analytics tools use to join behaviour across devices. Write the merge rule for conflicts before the first conflict occurs, not after.
Consent is not a setting inside a tool. It is a field that travels with the record: what this person agreed to, when, through which interface, and which version of the wording they saw. When consent is withdrawn, the effect has to propagate to every downstream system — and that only works if the field was part of the data model from day one rather than retrofitted.
In Saudi Arabia, the Personal Data Protection Law and its implementing regulations are published by the Saudi Data and AI Authority (SDAIA), which is the authority to work from. How those provisions apply to your particular tracking and integration design is a matter for a qualified adviser to review. We are a technical delivery team and this article is not legal advice.
One engineering habit avoids an entire category of trouble: keep personal data out of page URLs, query strings and the event names you send to analytics tools. Use opaque identifiers and hold the mapping between identifier and person inside your own systems.
The dashboard worth building — and the tile above it
A unified dashboard is not everything your tools can export. Build one view per recurring decision, not one per system. Three questions cover most marketing and sales teams: where qualified opportunities originate, by channel and campaign rather than by traffic; where they die, measured as stage-to-stage conversion and time in stage; and what each source costs per qualified opportunity and per riyal of pipeline.
Above those three, place a single data-quality tile: the share of new records arriving with a populated source field, the share carrying a click identifier, failed syncs in the last day, and duplicates detected this week. When that tile is red, everything beneath it is decoration — and the team needs to know that before it starts arguing about channel mix.
Two metrics to refuse outright: percentage splits between channels while a material share of records carries no source at all, and month-on-month comparisons that straddle a change in stage definition. Both reliably produce a long discussion and a wrong decision.
Three integration patterns and when each is right
Native connector between two tools. Right when the flow runs one way, the fields are standard and volume is moderate. Low cost and quick to stand up; the limits appear at the first non-standard business rule — merging two records, transforming a field value, or handling a partial failure inside a batch.
An integration platform where flows are assembled visually. Right for a team that wants to change logic without a development cycle, across a moderate number of connections. A practical caution: watch the cost curve as operation counts grow, and require every flow to be documented and safely re-runnable after a failure.
A custom integration layer with server-side event collection. The correct answer when identity and merge logic are genuinely complex, when volume is high, or when you need full control over what leaves your estate. Higher to start, and cheaper once the number of connections and rules exceeds what an off-the-shelf flow can carry.
Whichever pattern you choose, three things are not negotiable: a durable log of every sync you can go back to, the ability to replay a failed batch without creating duplicates, and alerts that reach a person by name rather than a shared inbox nobody opens.
A ninety-day sequence you can actually run
Days one to fifteen — definitions and inventory. Write the object model on one page. Get the sales lead to sign off the entry criterion for the first stage. Inventory the existing CRM fields alongside their real fill rates. The near-duplicate fields under slightly different names that you will uncover here usually justify the fortnight on their own.
Days sixteen to forty — the capture layer. Hidden fields on every form, a first-party cookie preserving first touch, a consent field, attribution fields locked against manual editing, and one agreed tagging convention with whoever runs the ad accounts.
Days forty-one to sixty-five — the loop and the dashboard. Map CRM stages to the events you send back to the ad platforms, then ship version one of the dashboard with its data-quality tile. Do not build the dashboard first. A dashboard sitting on incomplete capture produces worse decisions than no dashboard at all, precisely because people trust it.
Days sixty-six to ninety — run it without touching it. Change no definition and no stage. The rule we repeat with every client: freeze definitions for one full sales cycle before judging any channel, because moving a definition mid-measurement makes the comparison meaningless.
One mistake deserves a direct warning: buying a customer data platform before writing the object model. A platform does not create definitions. It replicates what you already have, faster — including the mess.
Key takeaways
- Write the object model on one page before choosing tools: who owns each object, the joining key, when it is created, who may edit it.
- Separate first-touch from last-touch fields, populate them from a first-party cookie, and lock them against manual editing in the CRM.
- Send back the earliest sales stage that is both frequent and revenue-correlated — not closed-won, if closed-won is rare in your cycle.
- Put a data-quality tile above the dashboard: source fill rate, click-ID coverage, failed syncs and duplicates. Red tile, no decisions.
- Freeze definitions for a full sales cycle before judging channels, and do not buy a customer data platform before the model exists.
Frequently asked questions
Should we start by replacing the CRM?
Usually not. The CRM is rarely the real constraint; undefined objects and missing capture fields are, and both migrate into the new system unchanged. Replace it when it genuinely cannot hold the fields you need or expose a reliable API for reading and writing — not before, and not because the interface feels dated.
How long before we can trust the numbers?
The capture layer can be working within weeks, but trusting comparisons takes a full sales cycle with stable definitions. Timeframes here are estimates that vary with the number of systems and the state of existing data, and should not be presented as a promise. The honest maturity signal is not a polished dashboard — it is a stable share of records arriving with a populated source field.
What do we do with historical records that have no source?
Do not backfill them with a guess or a rule of thumb; you will pay for that in every later report. Classify them explicitly as unattributed, show that share on the dashboard, and calculate channel performance from the date proper capture began. Being clear about the unknown portion is more useful than a complete number that is wrong.
Do we need a customer data platform?
Not before a written object model, disciplined capture fields and a working closed loop exist. A CDP is a unification and distribution layer for what you already have; it does not author definitions. Companies that buy one first tend to discover months later that they replicated the same mess somewhere more expensive.
Sources
- About offline conversion imports — Google Ads Help · https://support.google.com/google-ads/answer/2998031
- Conversions API — Meta for Developers · https://developers.facebook.com/docs/marketing-api/conversions-api
- Measure activity across platforms with User-ID — Google Analytics Help · https://support.google.com/analytics/answer/9213390
- Laws and Regulations — Personal Data Protection Law and its Implementing Regulations — Saudi Data and AI Authority (SDAIA) · https://sdaia.gov.sa/en/SDAIA/about/Pages/RegulationsAndPolicies.aspx