A 30-Day PMax Asset Group Test That Cut CPA by 22 Percent

One Campaign, Too Many Jobs

In March 2026, I audited Google Ads for a 14-person B2B SaaS company selling compliance workflow software to IT and operations teams. The product helped teams track vendor security reviews, SOC 2 evidence, and recurring policy attestations. Average contract value was $8,400. The sales cycle ran 31 to 46 days.

The account looked tidy at first glance. One Performance Max campaign. One branded Search campaign. One non-brand Search campaign. A small LinkedIn retargeting pool. Organic was pulling weight too, mostly from 19 comparison pages and a few vendor-risk templates tracked in Ahrefs.

PMax was the problem child.

In February 2026, the campaign spent $17,420 and produced 61 qualified demo requests at a $285 CPA. January had been $251. December 2025 had been $232. The founder wanted to know whether we should cut PMax and move the money back into Search.

I kept PMax live. Search was already covering the obvious demand. The non-brand campaign had 92.4 percent search impression share on the terms we cared about, and branded Search was basically rent. PMax was still finding new accounts. We had just fed it mush.

The campaign had four asset groups named by funnel stage: Awareness, Consideration, Demo, and Competitors. Inside Google Ads Editor 2.9, the structure fell apart. The same demo CTA showed up in every group. The same two product screenshots appeared in three groups. Audience signals overlapped across HubSpot customer lists, website visitors, a compliance software custom segment, and URLs from Vanta, Drata, and Secureframe.

Every group was chasing the same buyer with a different label.

The Test

The test ran from March 4 to April 2, 2026. I kept the budget flat at $600 per day because changing budget and structure at the same time makes the result harder to trust. The bidding strategy stayed on Maximize conversions with a target CPA of $240.

The conversion setup stayed fixed too. Google Ads counted a primary conversion called Demo Request - Qualified. The event came through Google Tag Manager Server-Side and landed in HubSpot. A lead counted only if it used a business email, had 50 or more employees, and selected a pain point tied to vendor reviews, audit prep, or policy management. Calendly bookings were tracked as secondary conversions.

The real change was asset group architecture.

We rebuilt four funnel-stage groups into five intent groups: vendor security review automation, SOC 2 evidence collection, policy attestation reminders, alternatives to spreadsheet tracking, and competitor comparison traffic. Each group got its own final URL cluster, image set, headlines, long headlines, descriptions, audience signals, and sitelinks.

I left final URL expansion on, but I added exclusions for the blog archive, pricing page, login page, careers page, and five thin tag pages that had picked up junk traffic in February. PMax will find cheap clicks in neglected corners if you let it.

For the SOC 2 group, the page cluster included /soc-2-evidence-collection/, /templates/soc-2-control-owner-checklist/, and /compare/vanta-alternative/. Headlines came from Gong calls and HubSpot notes: “Stop Chasing Control Owners,” “SOC 2 Evidence Without Spreadsheets,” and “Audit Prep That Starts in January.” The images were two product screenshots exported from Figma at 1200 by 628, one 1:1 dashboard crop, and one customer-style visual showing a security questionnaire queue.

No stock photos. No fake laptop smiles.

The spreadsheet group pointed to /vendor-risk-spreadsheet-alternative/, /blog/vendor-risk-management-template/, and /demo/. One long headline was “When the Vendor Risk Sheet Has 19 Owners.” That line came from a real sales note, which is usually where the best ad copy hides.

Competitor traffic got the most careful treatment. We used custom segment URLs from Vanta, Drata, OneTrust, Whistic, and SecurityScorecard, but we avoided competitor names in the ads. The copy focused on switching pain: implementation time, evidence handoffs, and stakeholder reminders.

The Numbers

I compared February 3 to March 3 against March 4 to April 2. Same 30-day length. Nearly the same weekday mix. Demo requests for this account clustered on Tuesday, Wednesday, and Thursday, so the calendar was clean enough for an operator-level read.

The campaign spent $17,987 during the test and generated 81 qualified demo requests. CPA fell from $285 to $222, a 22.1 percent reduction. Qualified volume rose 32.8 percent while spend increased 3.3 percent.

That was the dashboard win. The CRM numbers mattered more.

Demo-to-opportunity rate moved from 24.6 percent in the February pre-period to 28.4 percent during the test. By April 19, 2026, HubSpot showed 23 opportunities from test-period PMax leads. The pre-period produced 15 opportunities from 61 qualified demo requests.

GA4 engagement rate on PMax landing page sessions rose from 47.8 percent to 56.9 percent. The SOC 2 evidence group did the heaviest lifting. It spent $4,210, produced 24 qualified demo requests, and landed at a $175 CPA. The old Consideration asset group, which had swallowed much of that traffic before the rebuild, had run at $263 in February.

The spreadsheet alternative group looked weaker in Google Ads. Its CPA was $246. HubSpot told a better story: six opportunities from 19 qualified demo requests. That is the kind of group founders cut too early when they only read the ad platform.

Competitor traffic stayed messy. CPA improved from $319 to $271, but the notes were uneven. Three prospects were real switchers. Four were early researchers sent by managers to look at options. I kept the group live, tightened the audience signals, and excluded two softer pages.

Why It Worked

The old structure asked PMax to infer too much. “Consideration” is a deck label, not a buyer problem.

A security lead does not wake up thinking she is in the consideration stage. She thinks the auditor is asking for evidence again, the spreadsheet owner left in November, and the vendor review queue has 37 open items. When the asset group, landing page, and audience signal all point at that same mess, PMax has a cleaner job.

Performance Max still punishes vague inputs in 2026. Google has added more reporting since the early PMax rollout years of 2021 and 2022, but operators still do not get full query control or placement control. Asset group reporting helps. It does not rescue a campaign built from interchangeable headlines and homepage traffic.

I see the same failure in Shopify, SaaS, and local services accounts. Asset groups get names like Top Funnel, Remarketing, Core, Promo, and General. Then the machine blends traffic and creative into gray paste.

Name the pain.

For a running shoe brand, that might mean separating marathon training shoes, wide toe box running shoes, and recovery sandals, even if all three buyers can land on the same collection. For a dental group, same-week emergency appointments and Invisalign consults need different proof. For this SaaS account, SOC 2 evidence collection and spreadsheet replacement were different jobs.

Google needed that difference spelled out.

What I Would Repeat

I would rebuild around jobs-to-be-done language again. I would also keep the 30-day window. Shorter tests get jumpy in B2B because one enterprise lead can warp the readout. Longer tests are fine, but by day 30 we had enough spend and enough qualified conversions to make a decision.

I would use sales-call language again too. Gong, HubSpot notes, and actual demo recordings beat whiteboard copy. The 19-owner spreadsheet line worked because it sounded like the buyer’s week. Specific without trying too hard.

I would prune URLs before launch. Final URL expansion can work when the site has clean pages. This client had a few useful blog posts and a few thin posts from 2023 that had never been touched again. February traffic found them.

I would also read the CRM before calling the test. Google Ads said $222 CPA. HubSpot said 23 opportunities. If CPA had dropped 22 percent while opportunity rate fell in half, I would have rebuilt the campaign again.

What I Would Change

I would pass asset group data into HubSpot from day one. We had utm_campaign and utm_content, but we did not add a dedicated pmax_asset_group field until March 18. That made the first two weeks annoying to analyze. We had to stitch Google Ads exports to landing page sessions and form timestamps.

I would also prepare more image variants. We had eight strong visuals across five groups. By week three, Google had clear favorites, and two groups leaned too hard on the same dashboard crop. Next time I would build at least four unique images per group: one product close-up, one workflow view, one proof-oriented graphic, and one customer-context image.

Incrementality still sits under the whole story. Did PMax create demand, harvest demand, or take credit for demand organic and Search would have caught? This account did not have enough volume for a clean holdout without starving active pipeline. The rise in non-brand landing page sessions, the cleaner search category mix, and the HubSpot opportunity count made me comfortable. Comfortable is useful. It is not proof.

The Operator Lesson

A 22 percent CPA reduction sounds like a bidding story. This one was not. We did not change bid strategy, budget, attribution model, or conversion rules.

We made the asset groups less vague.

Performance Max can mix inventory across Search, YouTube, Display, Discover, Gmail, and Maps, but it still needs sharp inputs. Give it one homepage, six generic headlines, and an audience signal called buyers, and it will spend the money. Give it five distinct buyer pains, matching pages, and copy pulled from real calls, and it has a chance to find the pocket you meant.

For this account, the pocket was SOC 2 evidence collection. Thirty days was enough to find it. The next budget conversation got much easier.