How a Demand Gen Agency Lifted Paid Conversion 15% by Testing the Page, Not the Average

A B2B demand generation agency connected ad campaigns to on-page behaviour, found that one conversion rate was hiding eleven, and turned the worst of them into landing-page tests with the evidence attached.

---

A demand generation agency connected Google Ads, LinkedIn Ads, its analytics platform, and its clients' CRMs to Mindra, then asked AI agents to follow visitors from the ad they clicked to the step where they gave up. The result: a 15% higher paid-traffic conversion rate, 13% lower cost per completed conversion, and half the time spent working out what to test next.

"Our conversion rate was 3.1% for about a year. It turned out we didn't have a conversion rate. We had eleven of them, and two of them were dragging everything else down." — Head of growth, demand generation agency

Key takeaways

• An average conversion rate is not a thing you can optimise. Split by campaign, page and device, one client's 3.1% ranged from 0.6% to 7.2%. Every test run against the blended number was diluted by traffic it could not possibly affect.
• The worst segment wasn't a page problem. It was a promise problem. A "[competitor] alternative" search campaign converted at 0.6% on the same page where brand traffic converted at 4.8%. The ad promised a comparison; the page offered a general pitch.
• Three symptoms, three different tests. Leaving before scrolling means the page contradicted the ad. Starting a form and abandoning it means friction. Reaching the end and leaving means the offer is wrong. Only the second one is fixed by a shorter form.
• A missing-data check runs before any conclusion. Several apparent drop-offs turned out to be a tracking tag that stopped firing after a site release. Treating that as a page problem would have produced a confident test of nothing.
• Agents propose; the team tests. Every proposal arrives with the evidence, the specific change, the success measure and how long to run it. The agency decides what ships.

Why didn't the conversion rate explain anything?

The agency runs paid demand generation for nine B2B software clients, mostly Google Search and LinkedIn, all lead-gen. The reporting was fine. Every client had a conversion rate, tracked weekly, discussed monthly.

It just never told anyone what to do.

The number was an average of unrelated things. One client's paid traffic converted at 3.1%. Inside that were eleven campaign-and-page combinations, spanning 0.6% to 7.2%. A test that improved the weakest segment by half would move the headline number by a rounding error, so tests looked like failures and the team stopped trusting them.

Low conversion has at least three unrelated causes. A visitor who leaves in four seconds without scrolling has a different problem from one who fills in half a form and quits, who has a different problem again from one who reads everything and leaves. The first is message match, the second is friction, the third is the offer. The analytics dashboard reported all three as "did not convert."

Connecting the two halves was manual. Ad platforms knew the campaign. Analytics knew the on-page behaviour. The CRM knew whether the lead was any good. Joining them to answer "which campaign sends people who get stuck on the form" meant a half-day export-and-pivot exercise, so it happened before quarterly reviews, not before tests.

The result was a test backlog ordered by opinion. Someone would suggest a new hero image, it would ship, the blended rate wouldn't move, and nobody could say whether the idea was wrong or merely drowned.

How was the setup configured?

Four connections, made once per client:

• Google Ads and LinkedIn Ads, for campaign, ad group, creative and spend
• The analytics platform, for landing page, scroll depth, form starts, field-level abandonment and completions
• The client's CRM, where available, for whether a conversion became a qualified lead

Then the rules:

That last row is the structural choice. These are clients' websites. The agents never touch a page; they produce proposals good enough that implementing one takes an hour instead of a week of argument.

What was the first request, and what came back?

Pointed at the client whose conversion rate had been flat for a year:

"Follow paid visitors from the campaign they clicked to the point they gave up, and show me where the conversion rate is actually lost. Split it properly."

The reply reframed the account in one table:

Two findings, two different problems.

1. Message match. "Competitor alternative" traffic (1,470 sessions, $11,200 spend) leaves in under six seconds without scrolling. The ad offers a comparison against a named competitor; the page opens with a general product pitch and no mention of them. This is not a page-quality problem, it's a promise the page doesn't keep.

2. Form friction, mobile-specific. Field 4 is "Phone number," required. Desktop completion from form start: 44%. Mobile: 19%. The gap is consistent across every campaign that lands on /demo.

Tracking check: passed. Tag firing on all six segments, no gaps in the window.

A year of flat reporting; two concrete problems in four minutes. Neither was visible in the blended number, because the segments that were working had been quietly carrying the ones that weren't.

What did the AI agents actually build?

Four agents, one chain.

1. The traffic joiner. Every morning it stitches campaign data to on-page behaviour to CRM outcome, and always reports segmented by campaign × page × device, never blended. Segments under the traffic minimum are shown as provisional rather than ranked. This agent is the one that killed the average.

2. The drop-off finder. For each weak segment it locates the step where visitors leave and classifies it, because the class determines the test:

Before any of it, the agent checks tracking integrity. Twice in the first two months, an apparent collapse was a tag that stopped firing after a client's release — a page "problem" with no page involved.

3. The test proposer. Each proposal names the change, the evidence, the measure and the duration:

Proposal · "Competitor alternative" → /demo Change: a dedicated page variant whose headline names the competitor and opens with a three-row comparison, above the fold. Evidence: 1,470 sessions, 0.6% conversion, median time on page 5.8s, 91% never scroll past the hero. Brand traffic on the identical page converts at 7.2%, so the page works when the promise matches. Measure: scroll-past-hero rate first, conversion rate second. Scroll is the leading indicator and will move in days rather than weeks. Run for: 14 days or 400 sessions, whichever comes later. Spend at stake: $11,200/month currently going to a six-second bounce.

4. The result reviewer. After the agency ships a test, it measures against the stated goal, controlling for traffic-mix shifts during the window — a conversion lift means nothing if the campaign mix changed underneath it. It also checks the CRM: more form fills that never qualify is not a win. The phone-field test lifted mobile completion from 19% to 31%, and the CRM confirmed those extra leads qualified at the same rate as before, which is what made it safe to roll out everywhere.

Before vs after: what changed?

Across client accounts, compared with the equivalent period before Mindra and using the same conversion goal: a 15% higher paid-traffic conversion rate, 10% less drop-off at the targeted step, 13% lower cost per completed conversion, and 50% less time spent on analysis and test preparation.

Why didn't the analytics tool already do this?

They had a good analytics platform. It was not the obstacle.

Reporting is not segmentation. The tool could segment, in the sense that a human could build the view. Nobody built it every week for nine clients across three dimensions, so the default number everyone looked at was the one that hid the problem.

Classification is the step that turns data into a test. Knowing 1,470 people left is not actionable. Knowing they left in under six seconds without scrolling, from a campaign promising a comparison, onto a page that never mentions the competitor, is a brief. The gap between those two sentences is the whole job.

Proposals with evidence survive the client conversation. "We'd like to try a new headline" competes with everything else the client's team is doing. "This campaign spends $11,200 a month on a six-second bounce, and here's why" gets scheduled. Several of the agency's tests had been suggested a year earlier and declined; the same ideas shipped once they arrived with the numbers.

Frequently asked questions

How do AI agents tell a page problem from an offer problem? By where the visitor stops. Leaving within seconds without scrolling points at message match between the ad and the page. Scrolling but never starting the form points at the offer or the proof. Starting the form and abandoning points at friction, usually a specific field. Converting but never qualifying points at targeting, not the page at all. Each class leads to a different test, which is why the classification matters more than the conversion number.

What stops a drop-off from being mistaken for a website problem? A tracking integrity check runs before any segment is diagnosed. If a tag stopped firing, events are missing for a window, or a release coincides with the change, the agents report that instead of proposing a page test. In this agency's first two months, two apparent collapses were tracking, not behaviour.

Do the agents change the landing pages? No. These are clients' websites, so the agents produce proposals: the specific change, the evidence behind it, the metric that defines success, and how long to run. The agency decides what ships and implements it. The division is deliberate — automated analysis, human judgement on anything client-facing.

How do you know a test actually worked? The review controls for what else moved. A conversion lift during a window when the campaign mix shifted toward higher-intent traffic is not a page win. The agents compare like-for-like segments and, where CRM data is connected, check whether the extra conversions became qualified leads. More form fills of worse quality is a loss wearing a win's clothing.

How much traffic does a segment need before it's worth acting on? Set per client, typically a few hundred sessions or a couple of weeks. Segments below the threshold are reported as provisional and never ranked or tested against. Most false CRO conclusions come from reading a 40-session segment as a trend.

Do I need engineers to set this up? No. Mindra supports 3,000+ integrations, including Google Ads, LinkedIn Ads, analytics platforms and CRMs. You connect the tools, state the conversion goal and testing priorities once in plain language, and ask for what you want in a sentence. This agency's first working agent came from one request: "Show me where the conversion rate is actually lost. Split it properly."

Where Mindra fits

If your conversion rate has been the same number for a year and nobody can say what to change, Mindra is the analyst who splits it until it means something, names the step where people leave, and writes the test. Hire a whole department of AI agents with a single sentence. They connect to your tools, read your data, run on schedules, keep a memory of your rules, and bring you proposals with the evidence attached. It's model-agnostic (Claude, Gemini, GLM, Qwen, DeepSeek), SOC 2 Type II and GDPR compliant, with Zero Data Retention available, full audit trail, RBAC, and SSO. No engineers required.

See more results across industries on our case studies page, or read two related stories: how an agency diagnoses 40 client accounts every morning, and how a creative agency ships 50% more ad variations a month.

Ready to find the segments your average is hiding? Book a demo.