Your Google Ads dashboard reports a cost per lead every month, and that number feels like an answer. It is not. Your dashboard counts what the platform observed, and observation is a long way from causation. A DUI firm can spend $40,000 in a quarter, watch the reported cost per lead drop, and still have no idea whether the campaign produced a single extra signed retainer. That gap is where marketing mix models earn their money. Smart firms now rebuild their Google Ads for DUI leads around modeled contribution rather than platform-reported credit.
Here is where this gets interesting. Ben Vigneron published a piece in Search Engine Land on September 10, 2026. His argument is short. Running more than one marketing mix model against the same data is the only honest way to make a paid media decision. One model, he says, is a first opinion. Change the decay window or the priors, and the same history tells a different story. This article runs that argument through a criminal defense account. The channels are radio, billboards, and search. The outcome is a signed DUI case, and the budget is far too small to survive a wrong answer.
Legal Leads Group builds and manages lead generation campaigns for law firms nationwide, and criminal defense is one of the practice areas the team works in daily. The agency was founded by attorneys, accountants, developers, and programmers. That mix matters when the question stops being creative and starts being statistical.
Want a second read on what your campaigns are really producing? Call Legal Leads Group at (805) 273-8791 for a free campaign review. The team will tell you what your data can prove and what it cannot.
What a Marketing Mix Model Needs Before a DUI Law Firm Can Trust One
Start with the entry requirements, because most firms skip them and then wonder why the output looks strange. Google publishes its marketing mix modeling package, Meridian, as open source, and the documentation is refreshingly blunt about what the model expects. Legal Leads Group runs this check before any measurement conversation with a criminal defense client, because a model fed thin data will still produce a confident chart. The chart will just be wrong.
Google’s guidance on how to collect and organize MMM data asks for weekly numbers, channel-level aggregation, and a specific amount of history. Meta’s Robyn documentation asks for something similar. Neither one is trying to gatekeep. They are describing what a regression needs before it can separate advertising from everything else that moves your case volume. A DUI firm that clears these requirements can model. A firm that does not clear them can still measure, just with cheaper tools.
How Much History a Criminal Defense Firm Needs Before the First Model Runs
Google states the requirement plainly. Historical data should run a minimum of two years of weekly data for geo-level models and three years for national-level models. Read that as arithmetic rather than best practice. It describes how many observations you need to estimate a set of parameters without the model inventing a story. Most DUI firms have Google Ads history going back further than that, but the offline spend records rarely go back that far.
Why a National-Level Model Asks for Three Years of Weekly Data
Google’s own example does the math out loud. Two years of weekly data give you 104 data points. If the model is estimating 10 parameters, that is four data points per parameter, and Google calls that sample size too low to estimate the model reliably. Push to three years, and you have 156 data points against those same 10 parameters, which works out to nearly 15 observations each. The difference between those two numbers is the difference between a finding and a guess.
How Geo-Level Data Lowers the History Requirement for Multi-Office Firms
Here is the part that favors DUI firms specifically. Criminal defense practices are usually county-anchored, and many run three or four office locations across neighboring counties. Every county becomes its own row of weekly observations, which multiplies the data available without waiting another year. A firm with offices in four counties and two years of weekly history has roughly the observation count that a single-market firm needs three years to reach.
What Counts as a Channel and What Counts as a Campaign in a DUI Ad Account
Google is equally direct about granularity, and this one trips up marketers who love campaign-level reporting. The documentation says outright that it generally does not recommend running at the campaign level because MMM is a macro tool that works well at the channel level. Every campaign inside Google Ads gets summed into one channel line. Your seven search campaigns become one number.
For a typical DUI practice, the roll-up usually produces a short and honest list.
- Paid search across every Google Ads search campaign, branded and nonbranded together.
- Local Services Ads, which are priced per lead, cannot be controlled the way search can.
- Paid social, meaning Meta and anything running on Instagram.
- Broadcast and outdoor, covering radio spots, television, and billboards.
- Organic and referral, which carry no spend but absolutely move case volume.
That last line matters more than firms expect. Meridian supports organic media variables for unpaid activity and non-media treatment variables for marketing with no direct media cost. Leave referral relationships out and the model hands that credit to whichever paid channel was running.
How Many Channels Can a DUI Attorney Model Without Breaking the Math
Every channel you add costs you statistical power, and there is no way to buy more of it. Robyn’s own analyst guide sets the ratio at roughly one independent variable per ten observations. It recommends that rows outnumber columns by seven to ten times. Treat that as a hard ceiling. Your history caps your channel count, and no amount of enthusiasm changes the arithmetic.
The Ten Observations per Variable Rule Every MMM Runs On
Run the numbers on a single-market DUI firm with two years of weekly data. That is 104 rows. At ten observations per variable, the model supports roughly ten variables total, and seasonality controls, trend, and the intercept consume several of those before a single ad channel gets counted. Five or six media channels is a realistic ceiling. A firm trying to model twelve is not modeling anymore.
What Gets Dropped When a DUI Firm Runs Six Channels on Thin History
Something has to go, and the right something is usually the split you were most attached to. Firms love separating branded search from nonbranded search, because the two behave nothing alike in a criminal defense account. That split costs a full channel slot. If dropping it means you can keep radio and billboards in the model as separate lines, drop it and settle the branded question with a test instead.
Why Flat Monthly Budgets Give a DUI Marketing Model Nothing to Learn From
This is the requirement almost nobody talks about, and it quietly ruins more models than bad data ever does. Google warns that insufficient variation in media spend adversely affects national models. If your firm has spent $12,000 a month on search for three straight years, the model has never seen what happens at $8,000 or at $18,000. It will extrapolate the saturation curve from the shape of the math rather than from anything you did.
The fix costs nothing except nerve. Pull search back by 30% in one county for three weeks, push it up by 40% in another, and let the model watch. A firm running the same budget every month for two years has paid a great deal of money for a flat line.
What a DUI Law Firm Should Do When the Data Falls Short
Plenty of firms will read the requirements above and realize they do not qualify yet. That finding is useful rather than discouraging. Start logging weekly spend by channel today, including the radio invoices and the billboard contracts that currently live in a folder nobody opens. In two years, you will have a model. In the meantime, run geo tests, which cost less, answer one question at a time, and never need three years of history to work.
Why a Single Model Gives Branded Search Too Much Credit in a DUI Ad Account
Picture a Tuesday morning just before seven. A man books out of county jail after an overnight DUI arrest, gets his phone back, and types a law firm name into Google. He saw that name on a billboard three months ago and heard it on the radio during his commute for a year before that. He clicks the branded ad at the top of the page, calls, and signs a week later. Which channel produced that case?
Your Google Ads account has an answer ready. Branded search converted him, so branded search takes the credit, and its cost per lead looks spectacular. A single marketing mix model may agree, and that agreement is the trap. Branded search correlates tightly with signed cases in almost every criminal defense account, and correlation is what a poorly constrained model rewards.
How a DUI Arrest Sends Someone to Search a Firm Name Instead of a Keyword
Criminal defense demand behaves unlike almost any other legal vertical. Nobody plans a DUI arrest, so there is no research phase, no comparison shopping window, and no slow build of intent. Demand arrives all at once, at a strange hour, attached to a person who is frightened and holding a court date. When that person searches, they very often search for a name rather than a service.
Release Hours and the Searches That Follow an Overnight Booking
NHTSA reports that fatal crashes involving an alcohol-impaired driver happen at a rate three times higher at night than during the day. DUI enforcement tracks those same hours, which is why patrols run overnight. Bookings land after midnight, releases follow in the early morning, and the search that comes next arrives in a window most law firms are not staffed for. Those searches carry enormous intent and almost no keyword variety.
Referrals From Bail Agents, Family Members, and Former Clients
A large share of branded searches in a DUI account never started with an ad at all. Think about how the firm’s name reaches those people. A bail agent names three options. Someone’s cousin who hired you two years ago sends a text. A public defender mentions a name in the hallway. Every one of those people searches the firm name, clicks the branded ad because it sits above the organic listing, and gets counted as paid search performance.
Why Ridge Regression Hands Branded Search More Credit Than It Earned
Robyn, Meta’s open-source MMM package, uses ridge regression with an evolutionary search across hyperparameters. Ridge is fast; it runs without a Bayesian background, and it makes a strong first baseline. It also credits whatever correlates most tightly with conversions, and in a DUI account, that is branded search almost every time.
The Search Engine Land piece runs this exact scenario on a synthetic direct-to-consumer dataset, and the spread is worth a close look. Ridge regression assigned 41% of revenue to paid search. Two Bayesian models looking at identical inputs cut that credit to between 19% and 22%, because both treated search as partly downstream of demand that already existed. A holdout test settled the argument at 17% incremental. Those figures come from a retail example rather than a law firm, so read the pattern and never the percentage.
What Bayesian Priors Do to the Same Branded Search Number
A Bayesian model starts with a belief about how large an effect can plausibly be, then updates that belief as the data comes in. Google’s Meridian is Bayesian and geographically hierarchical. PyMC-Marketing, maintained by PyMC Labs under an Apache 2.0 license, is fully Bayesian with priors, model structure, and indirect-effect paths you define yourself. Both will pull an inflated branded search estimate back toward something defensible.
How Search Query Volume Controls Separate Demand From Advertising
Meridian offers a control built for precisely this problem. Google describes it as controlling for organic demand by including search query volume data in the model. Feed the model how many people searched DUI-related terms in your county each week, and the model can tell rising demand apart from rising advertising. Without that control, an enforcement crackdown looks like a great ad campaign.
What a Criminal Defense Firm Should Set as a Starting Priority
Do not invent a prior from optimism. Use whatever you have measured, and if you have measured nothing, start with a wide prior that admits you do not know. Meridian’s GeoX workflow lets a firm convert experiment results into priors directly, which means your first geo test becomes the constraint on your second model. That loop is the whole point.
Why Two Models Can Disagree by More Than Twice on One Channel
Two models can look at one dataset and split it by a factor of two without either one malfunctioning. Ridge with no priors pulling it back will chase correlation. A Bayesian model with demand controls will not. The disagreement is not noise. Two methods are showing you exactly which assumption carries the answer, and a single model can never give you that.
What a Branded Search Holdout Proves for a Criminal Defense Firm
Only a test settles it. Turn branded search off in one county for a defined window, leave it running everywhere else, and watch what happens to signed cases rather than to clicks. If the firm still signs the same volume, those people were finding you anyway, and the branded budget was buying traffic you already owned. Many firms discover the truth sits somewhere in the middle, which is a perfectly good answer.
Volume decides whether you can read the result. A practice signing 12 DUI cases a month cannot detect a 20% swing inside four weeks, because normal variation is larger than the effect. Run longer, run across more counties, or accept that the question stays open. Guidance for structuring Google Ads campaigns for DUI attorneys covers the account mechanics a clean holdout depends on.
How Two Different Branded Search Numbers Change the DUI Budget
The stakes here are not academic. A firm that believes branded search drives 40% of signed cases will protect that budget and underfund everything upstream. A firm that measures 17% will move the difference into the channels creating the demand that branded search collects. Same account, same spend, opposite strategy, and the only thing separating the two is whether anyone ran a second model.
How DUI Enforcement Campaigns Distort What Your Model Credits to Google Ads
On August 18, 2026, NHTSA launched its annual Drive Sober or Get Pulled Over campaign. Officers increased patrols nationwide from August 19 through Labor Day, and a federal media campaign ran through September 7 across television, radio, and digital platforms, including social media. Every DUI firm in the country saw lead volume move during that window. Almost none of them controlled for it.
Think about what a model does with that. Arrests rise because enforcement rises. Searches rise because arrests rise. Whichever ad channel your firm happened to be funding in late August absorbs the credit, and the model reports a channel that suddenly started working. It did not start working. A federal enforcement calendar changed how many arrests happened, and a model without tight seasonality controls cannot tell the difference.
The Enforcement Windows Every DUI Marketing Model Has to Control For
These windows are published and fixed, and they repeat every year, which makes them the easiest confound in legal marketing to control for and the most commonly ignored. Two of them matter more than the rest. Build both into the model as control variables before you read a single response curve. A channel that always spends at a peak will absorb calendar lift under weak controls.
The Labor Day High-Visibility Enforcement Period
The late-summer campaign is the larger of the two. Patrols intensify for roughly three weeks, and the federal paid media runs alongside them. NHTSA layers on companion messaging too. Drive High Get a DUI covers cannabis impairment, and Ride Sober or Get Pulled Over targets motorcyclists. Case volume in that window reflects how many officers were on the road rather than how well your ads performed.
The Winter Holiday Paid Media Flight and Its Fixed Dates
NHTSA runs the second flight from December 16, 2026, through January 1, 2027, and December carries National Impaired Driving Prevention Month on top of it. New Year’s Eve sits inside that window, which is exactly why NHTSA runs paid media straight through January 1. Any model that treats late December as ordinary weeks will hand a channel credit it never earned.
Why a Federal Ad Campaign Competes for the Same Screen
There is a second effect worth catching. NHTSA buys television, radio, and digital inventory during these flights, and so does your firm. Auction pressure rises, your radio slots get more expensive, and your digital impressions cost more at exactly the moment demand peaks. Your cost per click climbing in late December is a market condition rather than a campaign failure.
How to Code an Enforcement Window as a Control Variable
The implementation is simple enough that there is no excuse for skipping it. Add a binary column to your weekly dataset that reads 1 during an enforcement flight and 0 otherwise. Do it for both windows, and add a separate column for any state or county crackdown your market runs. The model then attributes that lift to the calendar rather than to whichever channel was spending.
Why a DUI Firm’s Own Spend Calendar Follows the Enforcement Calendar
This confounding gets worse from here, because most firms deepen it on purpose. Marketing directors know December and Labor Day weekend produce cases, so they push budget into those weeks. Spend rises at the same time arrests rise, driven by the same calendar, and the two series move together perfectly. A model cannot separate two things that always happen at once.
That is collinearity, and it is fatal to a clean read. A firm that has raised holiday spend every December for four years has never shown the model a December without a budget increase. Break the pattern deliberately in at least one market. Guidance on what destroys DUI lead generation campaigns covers the account-level mistakes that compound this one.
How Tighter Seasonality Controls Change Which Channel Gets the Credit
Run the model twice, once with loose seasonality controls and once with tight ones, and watch what moves. If a channel’s contribution collapses when the controls tighten, the calendar was driving that result, and the channel simply happened to be running. That is a diagnostic any firm can run in an afternoon once the dataset exists.
Watch the base and seasonality line too. In the Search Engine Land example, base plus seasonality accounted for between 21% and 25% of outcomes across three different models, which is the share advertising never touched. Suppose your DUI model reports a baseline that keeps growing while paid contribution shrinks. The honest reading is that your reputation and referral network outwork your ad account. That is good news, and the model found it for you.
Why Automatic Holiday Decomposition Misses DUI-Specific Peaks
Robyn uses Meta’s Prophet library to automatically decompose trend, seasonality, and holiday patterns, and that automation is genuinely useful. It also knows Christmas and Thanksgiving, not enforcement flights. Prophet has no idea that patrols surged on August 19, 2026, or that a state grant funded extra checkpoints in your county last March.
So add them yourself. Pull your state highway patrol’s published enforcement schedule, add county-level checkpoint announcements, and code each one. One caution deserves stating plainly. NHTSA publishes fatality counts rather than arrest counts. In 2024, 11,904 people died in alcohol-impaired-driving crashes, roughly 30% of all traffic deaths, and that figure describes fatal crashes rather than case volume. Use it to locate the enforcement calendar, never as a proxy for market size.
Which Offline Channels a DUI Lawyer’s Model Undercounts Without a Longer Decay Window
Robyn’s analyst guide publishes suggested decay values by channel type, and the spread should get every DUI marketer’s attention. Television sits between 0.3 and 0.8. Digital sits between 0 and 0.3. That single parameter, called theta, decides how long the model believes an ad keeps working after someone sees it. It can swing a channel’s entire contribution from meaningful to invisible.
Now apply that to a criminal defense practice. Nobody sees a DUI billboard and calls a lawyer that afternoon. They see it for eleven months, get arrested on a Saturday night, and call on Monday. If the model runs a digital decay window across that billboard spend, the effect has already decayed to nothing by the time the arrest happens. The billboard looks worthless, and it was not.
What Adstock Means for a Billboard on the Route Home From a Bar
Adstock is the modeling term for advertising that lags and decays after exposure rather than converting on contact. Every MMM includes it, and every MMM lets you choose how fast that decay runs. Set the window too short, and slow-building channels vanish from your results entirely. Set it too long, and you keep crediting ads that stopped mattering a season ago. For a DUI practice, the correct window is almost never the one the software picks by default.
Geometric Decay and the Numbers Behind a Decay Window
Geometric adstock is the simpler of the two common approaches and runs on one parameter. A theta of 0.2 means roughly a fifth of this week’s advertising effect carries into next week, then a fifth of that into the week after. The effect is gone within a month. A theta of 0.7 keeps most of the effect alive for two months or longer, which is a completely different picture of the same spend.
Why Broadcast and Digital Channels Need Different Windows
Robyn also offers Weibull adstock, which is more flexible than geometric and better suited to channels whose effect builds before it fades. Outdoor and radio often behave that way. A billboard on a commuter route needs repeated exposure before anyone remembers the name. The effect peaks weeks after the spend starts rather than in the week the contract begins.
How a Short Decay Window Makes Radio Look Worthless to a DUI Firm
The Search Engine Land example shows exactly what this looks like in output. Ridge regression with a short geometric adstock left television at 3% of the contribution. Two models running an eight to ten week decay window put the same channel between 14% and 16%. Three models, one dataset, and a five-fold difference on one line.
Vigneron’s reading of that split is the useful part. When a channel’s estimate follows the decay window that closely, the disagreement is telling you the true effect is slow-building and long-tailed. Neither model is broken. The channel needs a longer window and a longer evaluation period than your quarterly review gives it.
The Reallocation That Follows a Wrong Decay Window
Here is the damage in practice. A DUI firm spending $8,000 a month on radio reads a 3% contribution line and cuts the whole thing. The money moves to search. Six months later, branded search volume is down, the phone rings less, and nobody connects the two because the radio decision closed months earlier. The firm cut the channel that was feeding the channel it kept.
How to Spot an Adstock Artifact Before the Budget Moves
Catch this before the money moves, not after. The tell is simple. Change only the decay window and leave every other input alone. If the channel’s contribution moves dramatically, you are looking at an artifact of the parameter rather than a finding about the channel.
The Rerun That Confirms a Slow-Building Channel
Rerun the model three times on the suspect channel, using short, medium, and long decay windows. Watch which estimates move and which ones hold steady. This takes an afternoon once the dataset exists, and it costs nothing compared to killing a working channel.
What Changes in the Response Curve
The response curve shifts and flattens as the window lengthens. Contribution spreads across more weeks, the weekly peak drops, and the total climbs. That widening pattern is the signature of a channel whose effect builds slowly rather than one that never worked.
What Stays the Same in the Revenue Share
Watch which channels hold their rank no matter what you do to the window. Those rankings are the durable findings. Vigneron is direct about this. Budget decisions should rest on where each model sees diminishing returns and how consistently channels rank. Small differences in revenue share matter far less.
What a Confirmed Artifact Changes in the Media Plan
Never cut a confirmed slow-building channel. Give it longer flights and a longer evaluation window instead. Judge radio on two quarters rather than on one month, and stop comparing its cost per lead against search inside a 30-day report. The two channels operate on different clocks, so a single reporting period will always favor the faster one.
What Saturation Curves Say About Adding Another $10,000 to Search
Saturation is the second half of the modeling story and the half that answers your actual budget question. Every additional dollar of advertising still increases response, just at a declining rate. Robyn models this with a Hill function, and the curve shape tells you whether your next $10,000 buys cases or buys impressions.
Legal keywords make this urgent. WordStream by LocaliQ’s tenth annual benchmark study covered April 2025 through March 2026 across 20 industries. It put legal cost per click at $9.87 and legal cost per lead at $131.63. Both were the highest of any industry measured. Cross-industry averages ran $5.42 and $66.69. At those prices, the gap between the flat part of the curve and the steep part decides your year. Legal Leads Group builds digital marketing programs for DUI attorneys around that question.
Diminishing Returns in a Single-County DUI Market
Small markets saturate fast, and DUI markets are small by definition. There are only so many arrests in a county in a given week, and once your ads reach most of those people, more budget buys frequency rather than reach. A firm at the flat end of its search curve should move money into a channel that has room left, not push harder into a market that ran out.
Why the Curve Matters More Than the Revenue Share
Firms fixate on the percentage each channel contributed, which is the least useful number the model produces. The percentage describes history, while the curve drives the decision, because it tells you what happens to the next dollar. Two models can disagree by six points on share and still agree precisely on where the curve bends, and that agreement is the finding you act on.
How a DUI Law Firm Feeds Offline Spend Into the Model
None of this works without records, and offline records are where most firms fall apart. Digital spend exports itself. Radio and outdoor live in contracts, insertion orders, and email threads. Build a weekly log now, because the model needs equal granularity from every channel.
Capture these for each offline channel, every week, without exception.
- Gross spend by week, split out by market rather than reported as a national total.
- Flight start and end dates, since a contract month rarely matches a calendar month.
- Gross rating points or estimated impressions, which the vendor will supply.
- Daypart, because overnight radio and morning drive reach different people.
- Creative version, so a message change does not get read as a spend effect.
- Board or station location, mapped to the counties you actually file cases in.
- Any bonus or make-good inventory the vendor added at no charge.
That last item quietly wrecks more models than any other. Unlogged bonus inventory means the model sees exposure with no spend attached, which pushes efficiency estimates in a direction that has no basis in what you paid. Log it as delivered value even when the invoice reads zero.
How Criminal Defense Attorneys Turn Model Disagreement Into Better Google Ads for DUI Leads
So what do you do with two models that will not agree? Most marketers average them, which is the single worst available option. Averaging a 41% estimate and a 19% estimate produces 30%. No model produced that number, and no test supports it. Worse, it hides the information the second model was run to expose.
Vigneron’s instruction is to close the loop with testing rather than averaging, and to treat model disagreement as a ranked list of genuine unknowns. That reframing changes the whole exercise. Your two models did not fail to agree. They handed you a prioritized research agenda, sorted by how much money each open question is putting at risk.
Which Disagreements Can a DUI Firm Settle This Quarter
Sort every disagreement by one question. Can a geographic test answer it within 90 days? Branded search almost always can, because you can suppress it in one country without touching anything else. Radio decay usually can, given a long enough flight. The split between two channels that always scale together usually cannot, and knowing that in advance saves you from designing a test that was never going to produce an answer.
Running a Two-County Geo Holdout Without Pausing the Whole Account
A geo holdout suppresses one channel in one market while every other market runs untouched. Criminal defense firms are unusually well suited to this, because county lines already define how they operate, how they file, and how they buy media. You are not pausing the account. You are turning one variable off in one place and watching what the phones do.
Choosing Matched Markets That Share an Enforcement Schedule
Market selection decides whether the test produces an answer or noise. The two counties need similar case volume, similar demographics, and, above all, the same enforcement environment. A county running a grant-funded checkpoint program against one that is not will diverge for reasons that have nothing to do with your ads.
Why Neighboring Counties Beat Distant Ones
Adjacent counties share weather, share holidays, and usually share the same state patrol district running the same enforcement calendar. Those shared conditions are the point of a matched pair. A test county in your state against a control county 900 miles away introduces every difference you were trying to hold still.
How to Check That Two Counties Share a Media Market
Pull the designated market area for each county before you commit. If both sit inside one DMA, your radio and television spend already reaches both, and suppressing search in one of them does not stop the broadcast from working there. That overlap is a feature for isolating search and a problem if broadcast is the channel you meant to test.
Setting the Holdout Length Against DUI Case Cycle Time
Run the test long enough to capture a full case cycle. A DUI arrest, an arraignment, and a signed retainer play out over weeks, so a two-week holdout measures inquiry volume rather than signed cases. Six to eight weeks is a workable floor for most criminal defense practices, and longer is better when monthly case counts sit in the teens.
What the Test Costs Against What the Wrong Answer Costs
Firms hesitate because a holdout means giving up leads in one county for a defined period. Price that honestly, and the objection usually goes away. Suppressing one channel in one county for six weeks might cost a handful of cases. Reallocating six figures a year based on a model nobody validated costs considerably more, and it keeps costing until someone notices.
Which Test Comes First When Three Disagreements Compete
Run the test attached to the largest dollar disagreement first. If your models split by four points on paid social and by twenty on branded search, branded search wins the quarter. Vigneron makes the same point about efficiency, noting that one or two targeted geographic lift tests can resolve more uncertainty than a quarter of scattershot experimentation.
Sequence matters for a second reason. Each confirmed result narrows the next model, so an early answer on your biggest unknown tightens every later estimate. Approaches to attribution when buying legal leads follow the same logic, since the tracking layer decides what a test can see.
Which Disagreements Observational Data Will Never Settle
Some questions have no answer sitting in your history, and recognizing them early saves months of wasted effort. When two channels always move together, no amount of modeling recovers how the credit between them should split. The data does not contain that answer, and running a fourth model will not put it there. Mark these as permanently open, manage the pair as one budget line, and stop relitigating the split every quarter.
Search and Local Services Ads Scaling Together Every Season
This one is close to universal in criminal defense accounts. Local Services Ads and search ads usually get budgeted together, launched together, and increased together, so their weekly spend curves are nearly identical. Every model will split that combined credit differently, and each will report its own confidence. Only suppression in one market separates them.
Small Channels That Flip Sign Between Models
Watch for a channel that shows a positive contribution in one model and a negative contribution in another. That sign flip marks a fragile, noise-driven estimate rather than a channel doing harm. The correct response is to leave the budget alone and gather more data, because acting on an unstable estimate is how firms cut something that was working.
How to Feed a Confirmed Test Result Back Into the Next Model
A finished test is not the end of the process. Both Meridian and PyMC-Marketing accept experiment results as priors, so a measured number becomes the constraint on your next model run. Google built its GeoX workflow specifically to convert experiment output into Meridian priors, which turn each test into permanent knowledge rather than a one-time report.
Check the rerun for one specific thing. The disagreement you tested should visibly shrink. If two models were 20 points apart on branded search before the holdout and sit 4 points apart after, the calibration worked. If they are still 20 points apart, the test did not measure what you thought it measured, and that is worth knowing before the next budget cycle. Firms already using AI tools across their DUI lead pipeline should hold those channels to the same evidentiary standard.
What to Tell a Managing Partner Who Wants a Single Number
Partners want one number, and a range sounds like hedging until you frame it correctly. A single-model point estimate hides its uncertainty completely. Three models agreeing on a few points, with a test confirming the finding, give you a far stronger basis for moving money. A single unchecked point estimate gives you almost nothing.
Structure the report around evidence strength rather than around channels.
- Confirmed findings, where the models agree and a test backs them, are ready to move the budget on today.
- Convergent findings with no test yet, safe for smaller moves while a test gets scheduled.
- Open disagreements are held flat until the test that settles them runs.
- Unresolvable splits are managed as a combined line rather than argued about every quarter.
One more finding deserves its own warning. Sometimes all your models agree that a channel a partner has personally championed for years is underperforming. Vigneron calls that uncomfortable convergence, and it is the strongest signal any comparison produces, precisely because no assumption you chose is supporting it.
Build Better Google Ads for DUI Leads With Help From Legal Leads Group
Every DUI firm already owns the raw material for this work. Spend history sits in the ad account, case records sit in the practice management system, and the federal government publishes the enforcement calendar every year. What most firms lack is someone who will run the second model, name the disagreements out loud, and design the test that closes the biggest one.
Criminal defense practices across the country rely on Legal Leads Group to run their campaigns, and measurement comes with that work rather than as an add-on. The team will tell you whether you have enough history to model or whether a geo test is the smarter first move. You will also learn which channels your current reporting quietly over-credits. No firm should reallocate a six-figure budget based on one opinion.
Bring the last two years of weekly spend, including the radio invoices and billboard contracts, and the conversation gets specific fast. If those records do not exist yet, that is the first thing worth fixing, and it takes weeks rather than years. Better Google Ads for DUI leads start with knowing which channel actually produced the last hundred signed cases.
Call Legal Leads Group at (805) 273-8791 or reach the team through the contact page. Schedule a free consultation and get a straight answer about what your data can prove.