Your best Google Ads campaign is your worst one in the CRM

Here is a pattern you have probably seen. Campaign A comes in at a 42 pound cost per lead. Campaign B comes in at 118 pounds. Finance looks at the two numbers and declares Campaign A the winner by a mile. Three weeks later, sales has quietly stopped returning the calls from Campaign A’s leads, because they do not close, while Campaign B’s leads keep turning into opportunities. Both things are true at once. Campaign A is the best campaign in the ads dashboard and the worst one in the CRM.

The reason is not that anyone is doing paid search badly. It is that Smart Bidding optimises toward whatever you tell it a conversion is, and most B2B accounts have told it that a conversion is a form fill. Feed the machine form fills and it will get relentlessly efficient at finding people who fill in forms, which in a considered B2B purchase is a much larger and much worse population than the people who actually buy. This article is the full loop for fixing that: how to feed real sales outcomes back into bidding so the algorithm optimises toward pipeline. It is a technical piece, deliberately, because the advice is everywhere and the mechanism almost nowhere.

Why the platform cannot see what you care about

Start with the thing that explains everything else: the bid model only knows about the events you send it. It has no window into your CRM, no idea which leads your sales team liked, no sense of which ones closed. If the only event crossing back into Google Ads is “someone submitted a form”, then as far as the algorithm is concerned, a form submission is the definition of success, and it will spend your budget chasing more of them regardless of whether they are worth anything.

Now walk the gap that opens up between that form submit and a closed-won deal. In most B2B businesses there are several stages in between, and each one filters out a large share of the previous one. A useful default is four stages worth sending back: the raw lead (someone submitted), the marketing qualified lead (they fit the basic criteria), the sales accepted lead (a salesperson agreed they are worth pursuing), and closed won (they bought). The whole point of feeding outcomes back is to move the algorithm’s target further down that chain, away from the raw lead it can see today and toward the qualified stages that actually correlate with revenue.

The real constraint here is not accuracy, it is lag. A form fill happens in seconds; a closed-won deal in B2B might happen nine months later. Smart Bidding learns from events within a usable time window, so if your only valuable outcome sits most of a year past the click, the model cannot learn from it in time to be useful. This is the central tension of the whole exercise, and it is why the trick is rarely to optimise to closed won directly. It is usually to optimise to a mid-funnel stage that arrives soon enough to teach the model while still correlating strongly with revenue. Hold that idea, because it drives every decision that follows.

The plumbing, in order

Build it in this sequence, because each step depends on the one before.

The foundation is click-ID capture. When someone clicks a Google ad, the landing URL carries a GCLID, the identifier that lets you tie a later outcome back to the specific click and campaign. There are also related identifiers, GBRAID and WBRAID, that appear on clicks involving iOS apps, so a resilient setup captures whichever is present. The job at this stage is to read that identifier from the URL the moment someone lands and hold onto it, because if you lose it here, nothing downstream can work.

Next, get the click ID into the form as a hidden field. Add a hidden field, name it clearly, for example gclid, and populate it from the captured value so that when the form submits, the click ID travels with the submission. Consent state should be captured alongside it in the same moment, because whether you are allowed to use that click ID later depends on whether the visitor gave valid consent, and you need that recorded at the point of capture, not reconstructed afterward. The ICO’s guidance on cookies and similar technologies now covers link decoration and stored identifiers of exactly this kind, so treat consent as a build requirement, not an afterthought.

Then store the click ID against the right object in your CRM. This is where a surprising number of setups quietly fail. The click ID belongs on the record that represents the deal, the opportunity or lead object, not only on the contact, because one contact can generate several enquiries over time and you need each outcome tied to the click that produced it. Store the click ID, the timestamp, and the consent state together on that record so the whole chain stays intact as the deal moves through your pipeline.

Finally, the upload path back to Google. As of 2026 this has changed, and it matters: offline conversion imports and enhanced conversions for leads have moved to Google’s Data Manager, and the older Google Ads API route for these uploads is being retired, so build on the current path rather than a legacy tutorial. Google now recommends enhanced conversions for leads, which supplements the click ID with hashed first-party data such as an email address to improve matching, though including the GCLID where you have it is still recommended. The upload itself can run three ways: a manual sheet you upload periodically, a scheduled import from a sheet, or an automated connection through Data Manager or a supported integration. The full field specification lives in Google’s conversion import documentation, and you should treat that as the source of truth because the details change.

A few things reliably break this chain, so check them deliberately. Some form embeds and iframes strip URL parameters before your script can read the click ID, so test that the GCLID actually survives to the hidden field. Multi-step forms can lose the value between steps if it is not carried forward. Chat widgets and third-party lead tools often do not capture the click ID at all unless you configure them to. And CRM deduplication rules can overwrite the click ID when a returning contact submits again, silently replacing the identifier that tied an earlier deal to its click. Each of these is quiet, so none of them announces itself; you find them by testing the whole path end to end.

Setting conversion values that mean something

Once outcomes are flowing back, you have to tell Google what each one is worth, and this is where most implementations either give up and use a flat value or make the opposite mistake and use the full deal value. Both are wrong.

Do not upload the full deal value of a closed-won deal on a nine-month cycle. By the time that value arrives it is too late to teach the model efficiently, and it makes the occasional large deal dominate the bidding in a way that distorts everything. Instead, use expected value: the probability of a given stage converting to a sale, multiplied by your average deal size. That way you can assign a sensible value to a mid-funnel event the moment it happens, months before the deal actually closes, which solves the lag problem while still pointing the model at revenue.

Here is a worked example for a business with a 30,000 pound average contract value and the following stage-to-sale rates: 100 percent of raw leads exist by definition, 35 percent become marketing qualified, 18 percent become sales accepted, and 6 percent close. The expected value of each stage is that stage’s probability of eventually closing, multiplied by the 30,000 pound deal size.

The exact numbers matter less than the shape. A raw lead is worth around 1,800 pounds in expected terms, a sales-accepted lead several times that, and the gaps between the stages are what teach the model that some leads are worth far more than others. That is the entire mechanism: the algorithm stops treating every form fill as equal and starts chasing the profile of lead that reaches the valuable stages.

Two rules keep this stable. First, the values need to be relative and steady, not precise. Google’s model cares about the ratios between your stages far more than the absolute figures, so a defensible, consistent estimate beats a spuriously exact one that you tinker with constantly. Second, recalibrate on a slow cadence, quarterly is sensible, because every time you change the values you disturb what the model has learned. Changing values mid-flight resets learning, so set them thoughtfully, let them run, and revisit them on a schedule rather than reacting to every wobble.

Choosing the bid target once outcomes are flowing

With real outcomes and sensible values in place, you can choose a bid strategy that matches what you are trying to do, and the right choice depends mostly on your volume.

If your valuable events are still relatively sparse, stay on target cost per action aimed at a mid-funnel event such as the sales-accepted lead. This gives the model a countable, reasonably frequent target that arrives soon enough to learn from, while sitting far enough down the funnel to filter out the junk that a raw-lead target attracts. It is the workhorse choice for most mid-market B2B accounts, because it threads the lag-versus-signal needle, and it is the kind of setup we build in our paid search work.

Once you have enough valuable events flowing, and reliable values attached to them, you can move to maximise conversion value, optionally with a target return on ad spend. Now the model is optimising toward total expected pipeline value rather than a count of a single event, which is the more powerful setup because it lets the algorithm weigh a smaller number of high-value leads against a larger number of cheap ones. The catch is that it needs volume and value stability to work, so it is where you graduate to, not where you start.

That raises the honest question of the volume floor. Smart Bidding needs a reasonable flow of the event it is optimising toward to have anything to learn from, and if you close twelve deals a quarter, closed won is simply too sparse to be a bid target. This is the most common reason these setups underperform: the business optimises to an event that happens too rarely to teach the model anything. The answer is to move the target up the funnel to the most valuable event that still happens often enough, the sales-accepted or even the marketing-qualified stage, and let the expected-value weighting carry the revenue signal. Optimise to the deepest event you have enough of, not the deepest event you wish you had.

Reading the account afterwards

Once outcomes are flowing, the report that matters is not the default ads view. It is a four-column read: cost, leads, opportunities and revenue, broken out by campaign and by match type. That single view is what exposes the lie in the opening example, because it puts the cheap-lead campaign and the expensive-lead campaign next to their actual downstream outcomes. Building it is often the first thing we do in a digital audit.

What you are looking for is healthy divergence. A campaign with a high cost per lead but a low cost per opportunity is doing exactly what you want, spending more to attract fewer, better-qualified people. A campaign with a low cost per lead and a high cost per opportunity is the trap: cheap volume that does not convert. When you see that pattern clearly, the counterintuitive move is usually the right one. The high cost-per-lead campaign, the one finance wanted to cut, often deserves more budget, because it is the one producing pipeline. The whole exercise exists to earn you the confidence to make that call on evidence rather than on the misleading headline number.

The failure modes

Several things go wrong often enough to be worth naming, because most of them are quiet.

The first is human: sales stops updating stages. The entire loop depends on the CRM reflecting reality, so if deals sit in the wrong stage or never get marked closed, the signal you send back degrades and the model learns from noise. This is a process problem before it is a technical one, and it needs an owner. The second is attribution windows shorter than the sales cycle. If your conversion import window is shorter than the time it takes leads to reach the stage you are uploading, those conversions never get counted; note that enhanced conversions for leads will not import an outcome uploaded more than 63 days after the click, which directly shapes how deep a stage you can realistically feed back. The third is double counting, where a GA4 import and native tracking both record the same action and inflate your conversions, quietly telling the model two events happened when one did; Google’s attribution settings documentation is worth reading so you understand which system is crediting what. The fourth is subtler: lead scoring that encodes seniority bias rather than buying signal, so the model learns to chase senior job titles that look impressive and convert no better, instead of the behaviours that actually predict a purchase. Audit for that, because it hides inside a system that looks sophisticated.

The 30-day version for a team with no engineering resource

You do not need a developer or a data team to start. A capable marketer with access to the CRM and the ads account can stand up a working version of this in about a month.

In the first week, get the plumbing minimally working: capture the GCLID into a hidden form field, and make sure it lands on the deal record in your CRM along with the consent state. In the second week, define your stages and work out the expected values using the simple probability-times-deal-size method above, so you know what each stage is worth. In the third week, set up the upload, a scheduled Google Sheet through Data Manager is enough to begin with, and start sending your sales-accepted events back with their values. In the fourth week, leave the bidding alone while data accumulates, then once there is a reasonable flow, point a target cost-per-action strategy at your mid-funnel event and begin watching the four-column report. That is a complete, honest version of the loop, built with a spreadsheet and some care, and it will already stop your account optimising for the leads your sales team ignores. The automated, API-driven version is an upgrade you earn once volume or latency demands it, not a prerequisite for starting. Where it fits in the wider plan is something we work through as part of digital strategy.

FAQs

What conversion value should I upload for a lead that has not closed yet?

An expected value, meaning the stage’s probability of eventually closing multiplied by your average deal size, not the full deal value. For a 30,000 pound average deal where a sales-accepted lead closes a third of the time, you would upload around 10,000 pounds. This lets you value the lead the moment it reaches a stage, months before it closes, which is what makes the signal usable for bidding while the deal is still in progress.

How long after a click can I import an offline conversion into Google Ads?

Import windows are finite and set by Google, so check the current limit before designing a long sales-cycle loop. As of 2026, enhanced conversions for leads will not import an outcome uploaded more than 63 days after the associated click. That window directly constrains how deep a funnel stage you can feed back, which is one reason optimising to a mid-funnel event rather than closed won is usually the right call for long B2B cycles.

Do I need a developer to send CRM outcomes back to Google Ads?

No. A scheduled sheet upload through Google’s Data Manager works for most mid-market accounts, and a capable marketer can set it up without engineering support. API or automated integration matters when your volume is high or when latency becomes a problem, because manual uploads introduce delay. Start with the sheet, prove the loop works and produces better decisions, and upgrade to automation once the volume or the timing genuinely requires it.

Will optimising to opportunities reduce my lead volume?

Usually yes, and that is the point. When you tell the model to find sales-accepted leads or pipeline value rather than raw form fills, it stops chasing cheap, poorly-fitting clicks, so total lead count typically falls while lead quality rises. The number to judge success by is cost per opportunity, or cost per qualified customer, not cost per lead. A smaller number of better leads at a higher cost per lead is often far cheaper per actual customer.

What if consent is refused, can I still upload the conversion?

No. A click ID collected without valid consent must not be used for this, which is why consent state should be stored against the record at the moment of capture, alongside the click ID itself. Under UK rules consent must be freely given, specific, informed and unambiguous, and advertising measurement of this kind sits within that requirement. If consent was refused, exclude that record from your uploads. This is general information, not legal advice, so confirm your own obligations with a suitable adviser.

Last reviewed: August 2026

This article provides general information about paid search measurement and is not legal, compliance or financial advice. Google’s products and import specifications change frequently, and data-protection obligations depend on your circumstances; check the current Google Ads documentation and the ICO’s guidance, and take advice on your specific position before acting on this.