Two ad accounts, $287k spent, and what it bought us.

We run paid acquisition for two auto insurance lead generation brands. We are not an insurer and we do not sell policies. We buy clicks, send them to a form, and sell the completed form to a downstream buyer. Revenue is per lead, not per policy.

That single fact drives everything else in this document. Our ceiling is set by what a lead is worth to the buyer, and that number is roughly $19. Every campaign in these accounts is, in the end, an attempt to buy a form fill for less than that.

There is no era before this team

Worth knowing up front, because it changes how you read everything else: all 432 campaigns across both accounts carry the vlad. prefix, with zero exceptions. Nothing here was inherited from a previous agency or a previous build. The first vlad. campaign went live on 7 May 2025 on BudgetAutoRates, and that date is the start of this engagement.

So the history in this document is roughly fifteen months of one team's work, not a mix of eras. Every result below, good and bad, belongs to us.

One lead sells more than once

A detail that confuses everyone at first: our buyer reports — conversions sold against — leads delivered. The ratio is about 1.6. A single form fill is resold to roughly one and a half buyers on average, and the revenue we are credited with is the sum of those sales. So do not read "conversions sold" as "leads". The lead count is the smaller number.

The bar above is the whole business in one object. We spent — and — came back. That is a blended — return. The target is 0.95×, which is where this operation breaks even after the downstream repricing is accounted for.

We are not close to it, and we have never been close to it in aggregate. Anyone who tells you a tweak to bids will fix this has not looked at the distribution below.

Two systems disagree slightly and we show both. PowerFox Media is the buyer and therefore the truth for revenue. Google Ads is the truth for structure and configuration. Where a PFM figure exists we use it, and every row in this document carries a source tag.

They agree closely, which is the main reason to trust either. On cost the gap is 1.4% for CAQ and 0.1% for BAR. PFM reports 4% to 8% more revenue than Google Ads does, because the buyer keeps restating lead values upward after the fact and those restatements do not all make it back into Google.

Twenty one campaigns in the entire history of both accounts returned 0.95× or better, carrying $43,639 of spend, about 15% of the book. Sixteen of those spent at least $500 and so are worth taking seriously as evidence rather than noise. They are listed below. Several returned well over 1.0×, and one returned 1.92×.

All sixteen are BudgetAutoRates Demand Gen video campaigns. All sixteen ran between August and December 2025. And every single one of them is now paused or removed.

What they had in common

• Demand Gen, video first. Not search. Not display.
• Broad audience targeting, marked BroadAud in the name, rather than tight intent segments.
• One state per campaign. Look at the suffixes: VA, PA, NY, MI, TX, CO, SC, FL, CA, AZ, IL. The winners were geographically split, one cell per state.
• An age floor of 25+, 35+ or 45+, never unbounded.
• A long form video asset per cell, tagged VSL or NewsUrg or FormerEmp.

Not for performance. On 6 June 2026 all 28 BudgetAutoRates VSL Demand Gen campaigns were removed in a single operation, by operator instruction, because the video sales letter format was ruled to violate internal compliance. The removal record is in the repository at .planning/scripts/bar_vsl_removal_results.json.

The likely specific problem is the state comparative framing. Thirteen of the twenty eight were VSLWorstState cells, which assert that a named state is the "worst" for something. That is an unsubstantiated comparative claim in a regulated vertical, and it is the kind of thing that costs an account rather than an ad.

The effect shows up cleanly in the current run rate. Over the last 30 days the two accounts together spent $127,580 and returned $40,247. That is 0.316×, well below the 0.499× lifetime blend.

In other words, the accounts are not slowly improving. They are running worse than their own history, because the cohort that pulled the lifetime average up is gone and what replaced it performs at a third of break even. At current pace that is $4,253 of spend and $2,911 of unrecovered spend per day.

Every campaign name begins vlad. and that prefix is load bearing: our tooling refuses to mutate anything that does not start with it. It is the safety catch that stops an automated job touching a campaign it did not create. All 432 campaigns satisfy it today, so the guard has never actually had to block anything. Treat that as a property worth preserving, not as proof the guard is unnecessary.

The two accounts are almost mirror images. CAQ's money went into Search and Display. BAR's went almost entirely into Demand Gen.

BAR has been deleting its best work

Sort BAR's spend by campaign status and the ordering is backwards from what you would expect in a healthy account:

• REMOVED: $35,527 at 1.023×
• PAUSED: $71,874 at 0.632×
• ENABLED: $4,761 at 0.334×

The deleted campaigns are the best performing bucket in the account and the live ones are the worst. That is section 03 showing up in a different cut of the data. CAQ does not have this pattern: its statuses run 0.366× enabled, 0.329× paused, 0.291× removed, which is the normal shape of an account pruning its losers.

A campaign family is one idea, usually run as several cells at once. This is the list of ideas that were tested, with what each cost and returned. Sixty six campaigns were built and never spent a dollar, so they appear nowhere below.

Note the shape of BAR: one family, VAC, is 70 campaigns and carries almost the entire account. Note the shape of CAQ: no family dominates, because CAQ has been a sequence of experiments rather than a scaled winner.

The format table is worth sitting with. BAR Demand Gen video is the single largest block of spend in either account and returns 0.91×, close to break even. CAQ responsive search ads are the second largest and return 0.19×, which is the worst ratio at scale anywhere in the book. We have spent more on search text ads at CAQ than on anything else except BAR's video, and it has returned about nineteen cents on the dollar.

Why CAQ search fails, in one number

CAQ Search spent $76,148 and bought 3,652 clicks. That is $20.85 per click. Compare it to the other channels in the same account:

A search click costs seventeen times what a display click costs and about nine times what a Demand Gen click costs. A lead is worth roughly $19. At $20.85 a click, a single click has to convert at better than one in one to pay for itself, and it converts at roughly one in seven. That is the entire CAQ search problem and no amount of keyword selection closes a gap of that shape.

Below are the highest spending responsive search ads, assembled the way Google serves them. A responsive search ad is a pool of headlines and descriptions, not a fixed ad: Google picks three headlines and two descriptions per impression. What you see is a representative combination honouring the pins we set, not the only one that ran. Expand any card to see the full asset pool.

Reading these numbers correctly

The two accounts look wildly different here and the asymmetry is real. CAQ carries 16,800 keyword placements over 7,462 distinct terms. BAR carries 1,000 placements over 186 terms. CAQ is where the search experimentation happened.

One gotcha that will bite you: a keyword's criterion ID is keyed on the text and match type across the whole account, not per ad group. CAQ's 16,800 placements resolve to only 8,198 distinct criterion IDs because the same term is reused across many ad groups. If you mutate by criterion ID alone you will hit every ad group that shares it. Always address a keyword as adGroupCriteria/<ad_group_id>~<criterion_id>.

Second gotcha: keyword level conversion counts are lower than campaign level counts for the same spend, because Google does not attribute every conversion down to a keyword. Use keyword CPL to compare keywords against each other. Do not compare it against the campaign CPL in section 02.

Strip away the campaign structure and the business is two numbers. We pay — to acquire a lead. We earn — when we sell it. Everything else is detail.

To reach 0.95× from here, one of two things has to happen, and the size of the required move is the point:

• Cost per lead has to fall from $38.45 to about $20.18, a 48% reduction, holding lead value constant.
• Or earnings per lead has to rise from $19.17 to about $36.53, a 91% increase, holding cost constant.

Neither is a bid adjustment. A 48% cut in CPL is not something smart bidding delivers on traffic that is already being bought at auction; a 91% lift in lead value is a commercial negotiation or a different lead product, not an ads change.

The per brand split shows the same story at two scales. BAR buys leads at $27.43 and sells at $20.38. CAQ buys at $51.78 and sells at $17.71. BAR is closer to viable because Demand Gen video buys attention more cheaply than the search auction does, not because BAR's leads are worth much more.

Google keeps a change log with a hard 30 day window. Everything older is gone from the API, and the only record of it is our own repository: 500 dated handover notes in .planning/handovers/ plus the commit history. If you need to know why something was done in May, that is where to look.

Within the window, the tempo is high. 15,125 changes across both accounts in 30 days, almost all of them ours, applied through the API by our own tooling rather than by hand in the web interface.

Reconstructed from handover filenames, git history and campaign create dates. Every entry is anchored to a file or commit that exists in the repository.

236 of the 15,125 changes, 1.6%, came from Google's recommendation system. That sounds negligible. It is not, because of what they changed. They split into two categories and the difference matters:

1. Every mutation is dry run by default. Tools take --apply or --confirm and without it they only print what they would do. Read the dry run. Every time.
2. Nothing gets mutated unless its campaign name starts vlad. This is enforced in code and it is not negotiable.
3. The final URL host must match the brand. CAQ goes to find.cheapestautoquotes.com, BAR to quote.budgetautorates.com. Crossing them sends a brand's traffic to the wrong entity.
4. Copy changes go through compliance before deploy. Not after. The downside here is account termination, not a rejected ad. Note that 36% of what is live today already fails our own scanner (section 07), so "it is already running" is not evidence that something passed.
5. Nothing auto applies bids. Tooling proposes, a human approves, then it applies. If you build something that changes money on its own, you have built the wrong thing.
6. Check for Google auto-apply first when something looks strange. See section 10.
7. Do not compare a paused campaign's lifetime ROAS to a live one's without checking when each ran. The 2025 BAR cohort and the 2026 CAQ cohort operated in different auctions with different creative rules.

Where to look for anything this document does not cover

• .planning/handovers/ is 500 dated notes and is the real change record beyond 30 days.
• cmd/clone-ppc/ holds the operator tooling, including everything used to build this report.
• outputs/onboarding/ holds the raw extract this page was generated from, including every ad and keyword not shown here.
• PFM is the buyer's reporting surface and the arbiter of revenue. Google Ads is the arbiter of structure.