Why the event decides what the budget buys
Google documents conversion tracking as the way an account learns what happened after someone interacted with an ad, and Smart Bidding as bidding at auction time on the predicted likelihood and value of a conversion. The Smart Bidding strategies guide draws the consequence: the model learns from what is reported, a conversion it never sees is one it will not pursue, and it treats every included action as equally worth chasing unless values distinguish them. Meta's delivery system and TikTok's work the same way at the level that matters here. An ad set or ad group is optimised for an Optimisation EventThe conversion event an ad set or campaign is told to optimise delivery towards, such as a purchase, lead or install., and delivery looks for the people most likely to produce it.
The practical result is that the event is a purchasing instruction. Bids, budgets and targets decide how much is bought; the event decides what. A target cost per acquisition (CPACost of each conversion.) set carefully from the performance economics is wasted if the "acquisition" the platform counts is a newsletter sign-up or a page view that fires on refresh. The optimisation hub names the right event as the first of the conditions automated bidding needs; this guide is how to meet that condition on every platform at once.
The failure is rarely dramatic. Reports look healthy, cost per result falls, and nothing in any one platform's interface says that the platform has learned to buy the wrong thing. It shows up later, in the business's own records, as results that do not turn into revenue. The defence is a deliberate design, written down once and applied to every platform.
One written definition of each result
Start outside the platforms. List the results the business actually wants from paid media, define each in plain words, and say where it is recorded: the order system, the booking system, the CRM. A purchase is a completed order, not a visit to the confirmation page. A qualified lead is a lead the sales team has accepted, not a form submission. A booking is a confirmed appointment. These definitions are the business's, and they belong in the tracking plan beside the analytics events; the measurement plan framework is the general shape of that document.
Then build a signal register: one row per platform event, mapped to the definition it is meant to represent. The register is short and it settles arguments before they start.
| Field | What it records | Why it matters |
|---|---|---|
| Business result | The definition the event stands for, and the system that records it | Every platform figure is reconciled against this |
| Platform and event name | The conversion action, standard event or custom event on each platform | Shows where one result has several names, or one name several meanings |
| Role | Optimised for, observed only, or reporting only | Separates what steers bids from what is merely counted |
| Value | Revenue, margin, an expected value per type, or none, and where it comes from | Decides what value-based bidding maximises |
| Source | Browser tag, server event, offline import or analytics import | Shows where double counting can enter |
| Deduplication | The identifier each copy carries | Proves each result is counted once per platform |
| Owner and last check | Who maintains the event and when it was last verified | Keeps the register true after the next site release |
Google's own vocabulary fits this: the conversion action clean-up tutorial describes primary actions, which feed the Conversions column and drive bidding, and secondary actions, which are observed but do not. The same split applies on every platform, even where it has another name: a small set of events that steer bids, and everything else recorded for reporting and diagnosis.
Choosing the event each platform optimises for
The ideal optimisation event is the business result itself. It is also the event that arrives least often, and platforms learn from volume. The learning phase guide explains that a Meta ad set leaves learning only once enough optimisation events accumulate, and that a rare event can leave it learning indefinitely; the guide to bidding with few conversions sets out Google's options when conversions are scarce, one of which is to bid on a more frequent action that tracks the real goal. Choosing the event is therefore a trade between closeness to value and frequency.
| Situation | Event to optimise for | What to watch |
|---|---|---|
| The real result is frequent enough for the platform to learn | The real result | Nothing extra; this is the default |
| The real result is too rare on this platform or campaign | A more frequent step that reliably precedes it (an add to cart, a started checkout, a submitted enquiry) | Real results per optimised event, read from the business's own data |
| Leads are plentiful but quality varies widely | The qualified-lead outcome fed back from the CRM, once it arrives soon enough to steer | Lead-to-customer rate by campaign |
| The structure splits results across too many campaigns | The real result, after consolidating | Whether consolidation alone solves the volume problem |
Two rules keep a proxy honest. First, judge every campaign on the real result, never on the proxy it optimises for; the optimisation event test is built on exactly that rule, because an arm optimising for add to cart will nearly always win on add to cart. Second, before reaching for a proxy, check whether the structure is the real problem. The campaign structure guide explains why splitting spend across many small campaigns starves each of results, and consolidation often gives the real event enough volume to be used directly.
The choice can differ by platform. A search campaign capturing existing demand may see plenty of purchases while a paid social campaign prospecting among cold audiences sees few, so search optimises for the purchase and social for a step before it. That is fine, provided the register says so and both are reported on purchases.
Values that mean the same everywhere
A value turns counting into weighing. Value-based strategies on Google, and the highest value and return on ad spend options the Meta bid strategies guide describes, bid more for conversions expected to be worth more, which is only useful if the values describe what the business earns. The Smart Bidding guide puts the failure plainly: a campaign told that every lead is worth the same will optimise for the cheapest leads.
There are three sensible kinds of Conversion ValueThe monetary value an advertiser assigns to a conversion, such as the order's revenue or a lead's expected worth.. Revenue is the simplest for a shop and the figure return on ad spend (ROASRevenue generated per unit of ad spend.) targets are usually set on. Margin is better where products earn very different margins or returns are heavy, for the reasons the contribution margin guide sets out: a revenue target can be met while the business loses money. An expected value per lead type, the value of a closed deal multiplied by the share of that type that closes, is the lead-generation equivalent. Whichever is chosen, it must mean the same on every platform, in the same currency, with or without tax consistently, or the platforms' returns cannot be compared and the targets handed to them will be inconsistent. The ROAS calculator and the CPA calculator check that a target and the values behind it agree. Building that expected value with finance, from deal margin and close rates read from closed cohorts at each stage, is the guide to valuing a lead by pipeline stage. The expected value framework explains the probability weighting behind such values and why they need the business's own close rates.
Browser, server and deduplication
Each platform now offers a browser route and a server route for the same events. Meta documents the Pixel and the Conversions API, TikTok its Pixel and Events API, and Google enhanced conversions, which add hashed first-party data such as an email address to the conversion tag so that more conversions can be matched to ad interactions, as the enhanced conversions guide explains. The platforms describe the routes as complementary: the server reports events that ad blockers, consent choices and short-lived cookies would lose, and the browser supplies identifiers the server does not have.
Running both means each event can arrive twice, and a doubled count flatters every report and teaches delivery that conversions are easier to get than they are. DeduplicationRemoving double counts when one conversion is reported by more than one source. is therefore part of the design, not a clean-up step. The Conversions API guide explains Meta's rule: a browser event and a server event carrying the same event name and the same event ID are counted once, and the cleanest implementation takes the ID from the record that represents the action, such as the order number. The TikTok Pixel and Events API tutorial applies the same logic with TikTok's event identifier.
Google adds one trap of its own. Analytics records key eventsA GA4 event marked as important to the business, the property's measure of success., and those can be imported into Google Ads as conversions; if a native Google Ads tag records the same purchase and both are primary, the account counts it twice, which is one of the failures the clean-up tutorial is written to remove. Choose one source per result per platform, mark it in the register, and leave the other observed or switched off.
Feeding back outcomes the platforms cannot see
The results that matter most for lead generation happen after the click, in the sales process, where no tag can see them. The guide to connecting the CRM to ads and analytics covers the routes: an Offline Conversion ImportUploading outcomes recorded outside the website, such as closed deals, to an ad platform. that sends click identifiers with the outcome, enhanced conversions for leads, which match on the hashed contact details captured at the form, and the Conversions API, which Meta documents as the route for CRM outcomes. It also makes the point that decides the design: send the outcomes the platform cannot see, not every form fill it already reports.
The same guide describes a pattern that works on most programmes. Send two outcomes. A qualified-lead conversion, which arrives soon enough after the click to steer bidding, becomes the optimisation event; a won-deal conversion with its value arrives too late and too rarely to steer bids, so it is used for reporting and for checking that the qualified-lead event still predicts revenue. Shops have a smaller version of the same problem with cancelled orders and returns, which the business records long after the platform has counted the purchase; the reconciliation in the reporting guide is where that gap is measured.
Consent, modelling and the gaps they leave
Not every conversion can be observed. Where visitors decline advertising storage, the platforms see fewer conversions directly, and Google describes conversion modelling as its way of estimating the share it cannot match, so that reports and bidding include an estimate. The privacy and consent guide explains consent mode, when modelling is produced and its limits: modelled figures are estimates, they appear only where the documented conditions are met, and a market where almost everyone declines gives the model little to work from.
Three consequences for the signal design follow. Server events and hashed first-party data are sent only where the consent the visitor gave allows it, which the platform guides and the CRM guide each spell out. Modelled conversions are labelled as modelled in any report that includes them, and the business's own records stay the reference. And a sudden change in consent rates, for example after a new banner, is logged as a measurement change, because it moves every platform's reported conversions without anything having happened to demand.
Changing an event without losing what was learned
Changing what a platform optimises for restarts its learning. Google lists changing conversion actions among the changes that put a bid strategy back into a learning period, as the Smart Bidding guide records, and Meta lists changing the optimisation goal or conversion event among the significant edits that reset an ad set's learning phase. A redesign of the signal register is therefore a planned change, not a tidy-up.
- One platform at a time. Changing every platform in the same week makes it impossible to tell which change moved the blended result.
- Outside unusual periods. A change made during a sale or a seasonal peak teaches the system that the unusual period is normal.
- Observe before you switch. Add the new event as observed or secondary first, confirm in the business's records that it counts what the definition says, and only then make it the optimisation event.
- Judge on the business result. The new event will report differently from the old one by design; compare cost per real result before and after, read over a period long enough for late conversions to arrive.
- Record the change. Note it in the decision log and the tracking verification checklist run, so that the next review knows why the conversion column changed.
The conversion goal change checklist turns these rules into a gate to tick before each change on any platform.
A garden-room builder's signal register (illustrative)
Quorvane Garden Rooms aligns its conversion events
Quorvane Garden Rooms designs and builds garden offices and studios. Customers book a free design consultation online, and about one in four consultations leads to a paid deposit within six weeks. The business spends £9,000 a month on Google Search and £6,000 on Meta prospecting. Google's primary conversion is a brochure download and Meta optimises for submissions of an enquiry form; the platforms report 410 and 260 conversions a month, while the CRM records 140 booked consultations and 34 deposits.
The marketing lead writes three definitions (booked consultation, paid deposit, brochure download) and a register. Booked consultation becomes the optimisation event on both platforms, sent from the booking system through enhanced conversions for leads on Google and the Conversions API on Meta, with the booking reference as the event ID so the browser copy is deduplicated. Each booking carries an expected value: £2,400 for a garden office and £1,500 for a studio, a quarter of the average margin on each. Deposits are imported with their actual margin as a reporting-only conversion. Brochure downloads stay as a secondary action.
Google is switched first, in a quiet month, and Meta four weeks later. Reported conversions fall to 150 on Google and 95 on Meta, which looks like a collapse in the interfaces. Read against the CRM, cost per booked consultation falls from £107 to £88 over the following quarter, and cost per deposit from £441 to £352.
Common mistakes
- Letting the tag set-up choose the event. Whatever the installer marked as a conversion becomes what the budget buys; write the definitions first and map every platform to them.
- Optimising for a proxy and judging on it. A campaign optimising for add to cart or brochure downloads will always look good on that event; judge it on the real result in the business's records.
- Different meanings of "conversion" on each platform. If one platform counts purchases and another counts sign-ups, their costs cannot be compared and their targets cannot be set from the same economics.
- No values, or values that mean different things. Without values every conversion is worth the same; with revenue on one platform and margin on another, the returns are not comparable.
- Two primary sources for one result. A native tag and an analytics import, or a pixel and a server event without a shared ID, double the count and teach bidding that results are cheap.
- Sending every form fill back from the CRM. The platform already sees those; send the outcomes it cannot see.
- Switching events everywhere at once. Every platform re-enters learning together and nobody can tell which change did what.
