AI Marketing Playbook #005: Diagnose Campaign Performance Leaks Like a Senior Performance Marketer

Your Google Ads report says conversions are up. LinkedIn says cost per lead is down. GA4 has a third set of numbers, HubSpot has a fourth, and sales says half the leads aren't worth ringing. You've now read four reports, each internally consistent and none of them in agreement, and you still can't answer the only question that mattered when you opened the first one: where is this campaign actually going wrong?

This is one of the places AI earns its keep in B2B marketing — though not in the way it's usually deployed. The standard approach is to sit a marketing lead down for an hour to collect CPCs, conversion rates, MQLs, SQLs and pipeline figures, and type them into a prompt. The flaw in that is obvious once you say it out loud: if you already knew which numbers to pull together and how they connected to one another, you'd be most of the way through the diagnosis on your own. The prompt would only be confirming what you'd worked out while assembling it.

A better approach is to hand AI the evidence you already have and let it work out what the evidence can and cannot support.

Start With the Files, Not the Prompt Fields

Most marketing teams are not short of data. They're short of data in one place. Google Ads holds campaign performance, LinkedIn holds its own version of campaign reporting, GA4 holds website behaviour, and HubSpot or Salesforce holds the leads, opportunities and revenue that the business actually cares about. On top of that there is usually an agency report summarising some of it, produced by people with a reasonable interest in how the summary reads.

Rather than summarising those reports for AI, export them and upload them. In practice the useful set is your paid media exports from Google Ads or LinkedIn, a CRM export showing leads, opportunities and outcomes, landing page or GA4 data where it exists, and optionally the report your agency or internal team normally uses to explain performance. CSV or Excel exports are worth more than anything else here, because structured data can be inspected, joined and recalculated. PDFs and presentations still provide context, but a summarised figure in a deck is a claim, not evidence.

One practical caution before you upload anything: strip out customer information you aren't comfortable sharing. Names, email addresses, phone numbers and anything else personal should come out first, and none of it is needed for the analysis.

The First Question Shouldn't Be "What's Wrong With My Campaigns?"

That's the mistake nearly everyone makes, and it's the reason so many AI performance reviews read impressively and mean nothing. Before diagnosing anything, the model should establish whether the evidence in front of it is good enough to support a diagnosis at all.

Picture a fairly ordinary situation. Google Ads reports 300 conversions. The CRM contains 170 leads. Forty-two of those became qualified leads, and nine became opportunities. But there is no campaign ID, no UTM data and no GCLID connecting those nine opportunities back to individual campaigns. In that scenario the correct response is that Campaign A cannot be judged against Campaign B on pipeline, because nothing in the data links either campaign to an opportunity. Saying "I don't have enough evidence to determine that" is a far more valuable output than a confident ranking assembled from thin air.

In fact, that may well be the most important finding of the entire exercise. Your biggest campaign problem might not be a campaign at all. It might be that you cannot reliably tell which campaigns are producing revenue, which means every optimisation decision made in the last twelve months was taken partly blind.

So the Prompt Is Built Backwards

The prompt below begins by taking inventory of the files. What data do we actually have? Which date ranges overlap? Which datasets can genuinely be connected, and by what identifier? Where do the numbers disagree with each other? What can be calculated from this, and what cannot be concluded from it? Only once those questions are answered does it start hunting for performance leaks.

That order matters, because not every unattractive marketing metric is a problem. A campaign with a €150 cost per lead can be commercially superior to one producing €50 leads, if the expensive leads consistently become sales opportunities and the cheap ones consistently don't. A landing page with a lower conversion rate can be the better page, if the reason it converts less is that it filters out prospects sales would have rejected anyway. And a campaign can look excellent inside Google Ads while producing almost nothing the sales team wants to touch. The job isn't to find the ugliest number on the sheet. It's to find where business value is disappearing.

What the Prompt Looks For

Where the evidence supports it, the analysis follows the path from spend through traffic, conversion, lead, qualified lead, opportunity, pipeline and revenue, and asks at each step whether the drop-off is normal or notable. It then works through six areas separately.

• Paid media efficiency — whether particular campaigns are consuming disproportionate spend without producing corresponding results, and whether performance is deteriorating over time. If this is where you suspect the problem sits, it is worth running a dedicated audit of your Google Ads campaigns alongside this analysis.
• Targeting and lead quality — whether campaigns are generating plenty of conversions but very few prospects that sales actually accepts. Persistent quality problems usually trace back to who you are targeting in the first place, which is a job for validating your ideal customer profile.
• Landing page conversion — whether reasonable traffic is arriving and then failing to take the next step. A focused landing page review will tell you more about that than campaign data can.
• Funnel performance — whether leads enter the CRM and then disappear somewhere between qualification, opportunity and close.
• Measurement and attribution — whether campaign activity can be connected to business outcomes at all.
• Data quality — whether conflicting definitions, missing identifiers or mismatched reporting periods are quietly undermining the whole analysis.

The prompt is also written to resist the reflex of blaming advertising. Sometimes the campaign isn't the problem; the campaign is simply the thing with the most detailed reporting, which makes it the easiest place to find something that looks like a cause. Where the campaign itself was never properly defined, the gap usually shows up earlier, in the campaign brief.

What AI Still Cannot Tell You

Uploading more data does not automatically make AI right. It makes AI better informed, which is a different thing. A model can calculate patterns, compare campaigns and flag anomalies across a volume of information far faster than you could by jumping between reports, and that is genuinely useful. What it cannot do is establish cause. If one campaign has a poor opportunity rate, the numbers alone rarely explain whether the reason was targeting, positioning, the offer, sales follow-up, market conditions or something nobody has thought of yet.

For that reason the prompt sorts every conclusion into four categories: OBSERVED, meaning directly supported by the data; INFERRED, meaning strongly suggested by the evidence but not proven; HYPOTHESIS, meaning plausible and worth investigating; and UNKNOWN, meaning the available evidence isn't sufficient. That last category is the one that does the real work. An AI telling you it doesn't know is considerably more useful than an AI producing a well-written explanation for something it has no evidence for.

Copy the Prompt

How to Use It

Export the reports you already have, and resist the urge to assemble every system your company owns before you begin. Start with the best available evidence covering the same period — sixty to ninety days of matched reporting will tell you far more than four excellent reports drawn from four different date ranges. Upload the files to an AI tool capable of analysing spreadsheets and documents, then paste the prompt.

The one thing worth disciplining yourself about is the temptation to explain what you think the problem is first. Let the data inspection happen before you influence the diagnosis, because a model given your hypothesis up front will find evidence for it. When the output comes back, pay particular attention to two sections: what the AI could not determine, and what it recommends investigating first. Those two answers are the point of the whole exercise. You aren't trying to generate another thirty-page marketing report; you're trying to reduce a complicated performance problem to the few things actually worth looking into.

One Final Warning

Don't implement whatever comes back without reading the reasoning behind it. If the analysis tells you to pause a campaign, shift budget, change bidding strategy or rewrite a landing page, look at the evidence supporting that recommendation before you act. A good analysis can always tell you why something deserves investigation; if it can't, it hasn't found anything.

AI is useful for finding patterns in more data than a person can hold in their head at once. It is considerably less useful when we let it turn incomplete evidence into certainty. And if your reports can't tell you whether advertising is producing pipeline, that isn't a reason to ask AI to guess harder. That's the leak you need to fix first — and closing the loop between marketing spend and revenue is where that repair starts.

Frequently Asked Questions

How do I use AI to analyse campaign performance data?

Upload your raw exports rather than typing metrics into a prompt. The most useful set is a paid media export from Google Ads or LinkedIn, a CRM export containing leads, opportunities and outcomes, and GA4 or landing page data covering the same period. Structured formats such as CSV and Excel are considerably more useful than PDFs, because the data can be inspected and recalculated rather than merely read. Then ask the model to inventory the evidence and assess data quality before it diagnoses anything.

Why shouldn't I just tell AI my campaign metrics?

Because deciding which metrics to include is most of the analysis. If you already know that cost per lead should be compared against opportunity rate by campaign, you've done the diagnostic thinking yourself and the model is only confirming it. Uploading the raw files lets the analysis find relationships you weren't looking for, including the ones that contradict your current explanation of performance.

What is a campaign performance leak?

A performance leak is any point in the path from advertising spend to revenue where business value deteriorates faster than it should. It might sit in paid media efficiency, landing page conversion, lead quality, the marketing-to-sales handoff, or opportunity creation. Frequently it sits in measurement itself: if campaign activity cannot be connected to CRM outcomes, every downstream conclusion about performance is unverifiable.

Can AI tell me which campaigns generate pipeline?

Only if a reliable identifier connects campaign activity to CRM records — campaign ID, UTM parameters, GCLID, lead ID or an equivalent. Without one, the honest answer is that individual campaigns cannot be attributed to opportunities, and totals can be compared only in aggregate over matching periods. An AI that ranks campaigns by pipeline in the absence of that identifier is guessing convincingly.

Is a lower cost per lead always better in B2B marketing?

No. Cost per lead measures the price of an enquiry, not the value of one. A campaign producing €150 leads can outperform one producing €50 leads if its leads convert to opportunities at a materially higher rate, and optimising toward the cheaper source can quietly reduce pipeline while every paid media metric improves. Cost per qualified lead and cost per opportunity are the more honest comparisons wherever your data supports calculating them.

What can't AI determine from campaign data?

Cause. AI can identify that a campaign has a weak opportunity rate; it cannot tell you from the numbers alone whether the reason was targeting, positioning, the offer, sales follow-up or market conditions. That's why conclusions should be classified as observed, inferred, hypothesis or unknown, and why "insufficient evidence" is a legitimate finding rather than a failure of the analysis.