PI Firms Fucked Themselves, Part III: You Taught Google the Wrong Fucking Number

Imagine a personal injury firm that wants Google to get smarter. It connects its CRM, imports signed-client events, assigns conversion values, and commits a large budget to automated bidding.

Signed cases rise. Reported conversion value rises. Everybody has a reason to believe the machine is working.

There is one small fucking problem. What if the number the firm taught Google to optimize never represented the economics management thought it represented?

Google may have performed exactly as configured.

Part II separated the clients paying the firm today from the clients being acquired today. This article examines the estimate connecting those two generations: what management believes a new client will be worth before the actual outcome exists.

The useful question is whether that estimate can support the decisions being made with it. I would test the number before asking the machine to pursue more of it.

Google Predicts the Outcome You Define

Google's automated bidding uses historical information and auction context to predict conversions or conversion value and set bids. The advertiser chooses goals, values, budgets, and targets. [1, 2]

That makes the business definition consequential. A model can become excellent at pursuing a configured outcome without independently establishing that the outcome represents collected fees or profit.

Google permits conversion values representing economic outcomes or proxies such as a lead score. Its value-based bidding guidance recognizes a tradeoff between how informative a conversion is and how long it takes to report. [6, 7]

That is reasonable. A law firm needs to make decisions before every matter resolves. A qualified lead or signed-client event may provide useful information much sooner than a collected fee.

But useful information is not certainty. And an estimate does not become money because it appears in the revenue-colored column of a dashboard.

Sophisticated Tracking Does Not Eliminate Time

A call can happen at noon. A retainer might be signed later that day. The firm may then spend months developing facts, handling the matter, and waiting for a result.

At the time of the advertising auction, nobody knows the final fee or the exact cost of producing it. A useful expected value may be possible. The final outcome has not happened.

Google supports importing later offline events, including a contract signing, and connecting them to an earlier advertising interaction. Its conversion-cycle definition includes reporting delay for imported events, and its reporting has processing delays of its own. [3, 4, 5]

Those are separate clocks: time until the event occurs, time until the firm reports it, platform processing, and time until the business knows the economic result it ultimately cares about.

There are practical import limits too. Google's current offline-conversion FAQ distinguishes a 90-day GCLID window from a 63-day window for enhanced-conversion imports using user-provided data. Method and settings matter. Conversion adjustments also have limits; they do not create unlimited time to attach future economics to an old click. [10]

A case producing fees much later may therefore need an earlier event for bidding and a longer internal history for business evaluation. Those two jobs do not need to run on the same clock.

The mistake would be treating the earlier event as proof that the later economics are already known.

Now Give the Estimate a Budget

Suppose a hypothetical firm assigns each signed PI client a conversion value of $50,000. It spends $100,000 on advertising and the campaign is credited with ten signed clients.

The system reports $500,000 in conversion value. Divide that assigned value by advertising spend and the ratio is 5.0, or 500%. Google's conversion-value-to-cost metric performs that division. [9]

It is not 500% net profit. It is not proof that the firm collected half a million dollars. It is ten recorded outcomes multiplied by the values the firm supplied.

Target ROAS uses predicted conversion value to pursue the advertiser's return target. It does not independently audit what “$50,000” means inside the business. [8]

If signed-client volume then rises, the model may be doing exactly what was requested. Management may increase the budget because the report appears to validate the decision.

Eighteen months later, some matters produce no fee, some produce modest fees, and one produces a large result. The original value might have been a useful estimate. It might have overstated expected fees. It might have represented gross recovery when management thought it represented the firm's share. The label alone cannot answer that.

Show me how the fucking number was constructed.

One Average Can Conceal Very Different Businesses

Suppose two clients each receive the same initial $50,000 value. One eventually produces $10,000 in collected attorney fees; the other produces $100,000. If each cost $15,000 in paid media, one is already below media cost and the other has substantial room before additional expenses.

That variation does not, by itself, prove the shared forecast was wrong. An expected value is not a promise about every case. It has to be tested across the population it is being used to buy.

Now consider a different invented group of ten matters. Nine average about $17,222 in fees. One produces $345,000. Together they generate roughly $500,000, or a $50,000 average.

The average is mathematically reasonable. The largest case produces about 69% of the total. Whether that is a sound acquisition model depends on how repeatable the distribution is, what the work costs, and whether the firm can carry the risk and timing.

I want to see that distribution, not just the fucking average. A model can rank promising matters well while overstating their dollar value. It can fit yesterday's case mix and mislead after the acquisition mix changes.

And fee value is not the whole answer. Intake, handling costs, unreimbursed expenses, unsuccessful matters, and overhead still affect what the business retains.

Value Has a Clock and Capacity Has a Cost

Suppose a firm acquires ten clients each month at $15,000 in paid media per client. That is $150,000 leaving each month. If those new cohorts return no fees for the first three or four months, acquisition alone has absorbed $450,000 to $600,000 before their first collections.

That is an invented timing example, not a normal PI timetable. It excludes payroll, case expenses, and other costs of carrying the work.

Two matters can produce the same eventual fee while requiring different amounts of time, capital, and attorney capacity. Increasing volume can also change the cost or speed of servicing the work. The value model needs to be useful for the business the firm can actually operate.

That is why doubling signed cases is not enough to establish success. More volume can improve contribution. It can also buy the firm more obligations than its cash and capacity can comfortably support.

Before management celebrates the bigger pipeline, I want to know what changed in expected contribution, the wait for payment, and the ability to deliver the work.

Getting Credit Is Not the Same as Creating Another Client

There is a separate question underneath even a well-calibrated value model: did the spending create additional business?

Imagine a person receives a referral, searches the firm's name, and clicks an advertisement. A campaign receiving credit for that client does not, by itself, tell us whether the person would otherwise have contacted the firm. Nor does the referral prove the ad had no effect.

That is a question to test, not an excuse to dismiss either channel.

I would distinguish recorded attribution from evidence about what additional spending actually produced. Otherwise a firm can attach an accurate value to a client and still draw an unsupported conclusion about the return from acquiring that client through a particular campaign.

Better tracking helps describe the observed journey. The spending decision still needs a defensible comparison with what would have happened under a different decision.

Your PPC Manager Might Be Fantastic

Give a capable operator a target for qualified leads, signed clients, or reported return, and they may become very good at producing it.

That does not automatically give them access to collected fees, servicing costs, mature case outcomes, or authority over allowable acquisition cost. A good operator should challenge a weak objective. Management still has to provide a defensible one.

The same applies to the CRM. It can transport a value perfectly without making the value economically sound. Integration removes manual work; it does not settle a disagreement about what the business should optimize.

Someone has to connect the operating evidence to the bidding assumptions and have authority to change the decision when they diverge. Otherwise every department can maintain its own correct-looking report while nobody owns the full result.

This is where I would bring marketing, finance, technology, and operations into the same conversation. The work is to decide which signal can support which decision—and stop treating every number as interchangeable.

The Correction Starts With the Meaning of the Number

I would begin with sufficiently mature internal outcomes and work backward to what the firm believed at acquisition. Which event was reported? What value was assigned? When? What did the comparable matters eventually produce, and at what cost?

If the forecast is useful, preserve it and improve what needs improving. If it consistently overstates the relevant economics, examine the definition, inputs, and spending decisions built on it. If the economics deteriorated after the signing because operations changed, address that instead of automatically condemning the acquisition source.

The goal is a usable connection between early signals and business results. Management needs enough information soon enough to act, plus a process for finding out when that information has stopped being reliable.

I am interested in rebuilding that connection. An automated campaign is much easier to evaluate once the firm can explain what success means and why the number deserves its budget.

A Proxy Can Be the Right Answer

The strongest counterargument is that early estimates are necessary and often useful. Waiting for every case to resolve would leave the firm unable to operate a timely acquisition system.

Agreed. Google's guidance explicitly accommodates proxies and the accuracy-versus-delay tradeoff. [6] A reasonably calibrated expected value can be more useful than an economically complete number arriving too late.

A large difference between two individual outcomes does not invalidate a forecast. Nor does one unusually good or bad cohort establish a permanent pattern. A diversified portfolio can make sense even when many matters produce little or nothing.

I would not tear out a functioning model because an actual outcome differed from an estimate. The question is whether it performs well enough across the relevant population to justify the decisions made with it.

That is a much stronger standard than demanding perfect foresight—or accepting every number the CRM exports.

What Would Prove This Wrong?

My hypothesis is that automated bidding can amplify weak economics when early values are treated as established returns and never checked against later outcomes.

Compare predicted and realized outcomes across similar case types and comparable cohort ages. Include zero-fee matters. Keep forecasts for open matters separate from collected fees. Test on cohorts that were not used to construct the model, and examine both useful ranking and dollar calibration.

Then ask whether additional spending produces worthwhile additional contribution at acceptable timing and capacity requirements. A good prediction and a good spending decision are related, but they are not identical tests.

If the values hold up and acquisition remains economically sound, my criticism loses force. If the apparent problem comes from changed operations or a faulty comparison, correct the diagnosis. If values consistently overstate outcomes and spending follows those values, there is a business decision worth changing.

A precise fucking number still needs a defensible meaning.

Part IV examines an earlier decision hiding inside that number: who decided the lead was qualified, and what did they actually know at the time?