Google Says You Have Conversions. Sales Says the Leads Are Shit.

Congratulations. The Number Went Up.

Your advertisements are running.

Google Ads is reporting conversions.

The cost per conversion is down, and somebody wants to increase the budget.

Then you talk to the people answering the phones.

Wrong customers. Wrong services. Wrong side of the legal dispute. People who misunderstood the offer. People the business cannot serve.

Marketing says the campaign is working.

Sales says the leads are shit.

Before either side wins the argument, establish what the fuck Google is counting.

Because a conversion reported by Google is not a certificate of commercial value.

It may represent a completed purchase. It may represent a submitted form. Depending on the setup, it may reflect a call-duration threshold, a click on a telephone number, an estimated call outcome, or a local action such as requesting directions. Some conversion reporting also incorporates modeling and credit for interactions that did not involve an ad click. [1–5]

Those are different kinds of evidence.

Google can be right about the conversion under its rules. Sales can be right that the inquiry cannot be used. The business still has to decide whether it bought anything worth buying.

This is not an accusation that Google fabricated the number. It is a warning not to ask the number to answer a question it was never designed to answer.

If you treat them as interchangeable proof that advertising produced customers, the misunderstanding starts before the AI makes its next bidding decision.

Then you can give the system permission to pursue more of the thing you misunderstood.

What Does Google Actually Deem a Conversion?

There are three separate questions: what happened, how it was measured, and which advertising interaction receives credit.

For many conversion actions, the advertiser defines the event and configures the measurement. Google also provides automatically measured actions and modeled reporting. Calling something a “Google conversion” does not tell you which process produced it. [1, 3, 4]

Here are the distinctions I would want a business owner to understand.

Website and app actions. Purchases, registrations, downloads, and other configured actions can be measured. A successfully recorded form submission establishes that the tracked event occurred; it does not establish that the person meets your sales criteria. [1]

Calls measured by duration. Some Google call-tracking methods let you choose a minimum call length. Meeting that threshold can count as a conversion. A long conversation with someone you cannot help is still a long conversation. [2]

Clicks and estimated calls. Tracking a click on a phone number on a mobile website can measure the click without measuring the call itself. Separately, Google documents a method for call-ad and asset clicks without a forwarding number in which conversion counting relies on Google's estimate of whether a meaningful call occurred. Neither is equivalent to an employee confirming a qualified prospect. [2]

Local actions. Eligible location-asset activity can automatically create Google-hosted conversion goals. Directions, call clicks, website clicks, and other local engagements have different reporting and bidding eligibility. Not every reported local action is used for bidding. A directions request is not proof of a visit or purchase. [3]

Modeled conversions. Google uses observed data to estimate missing connections between advertising interactions and conversions. It says these estimates enter reporting only when confidence is sufficiently high and that modeling can inform bidding. This addresses measurement gaps; it does not give you an individually verified customer record for every modeled contribution. [4]

View-through and engaged-view conversions. These can credit an ad exposure or qualifying video engagement followed by a conversion within the relevant window, without requiring an ad click. Watching a video alone is not the same thing as an engaged-view conversion. Reporting and bidding treatment vary by campaign and settings. [5]

Imported downstream outcomes. A business can also supply eligible outcomes recorded elsewhere, such as qualified leads or completed sales. That can move measurement closer to the business result, provided the underlying records are reliable and the upload is appropriate. [15]

This is a practical map of the distinctions, not every specialized conversion feature Google offers.

And the point is not that every modeled or non-click conversion is fake. It is that observation, estimation, attribution, qualification, and profitability answer different questions.

“Google counted it” is where the investigation starts.

A Conversion Count Is Not a Customer Count

One person might submit a form, call, and later purchase. If you track those as separate actions, adding them together does not produce three customers.

Google's “One” and “Every” counting options operate at the conversion-action level after an ad interaction. Choosing “One” does not create universal deduplication across every action and customer record in your business. [1]

Then there is the column you are reading.

Primary actions generally appear in “Conversions” and guide bidding when their standard goal is selected for the campaign. Secondary actions generally appear in “All conversions” for observation. A significant exception: actions included in a custom goal can be used for bidding even when labeled secondary. [6]

So don't stop at “we made that secondary.” Check the campaign's actual goals.

And don't present “All conversions” as if it were a list of new paying customers.

Different dates, counting rules, attribution windows, and measurement methods also complicate reconciliation. A discrepancy between Google and the CRM calls for investigation, not an immediate accusation that one side is lying.

What Google's AI Is Optimizing

Google describes Smart Bidding as auction-time optimization for conversions or conversion value. Strategies include Maximize conversions, Target CPA, Maximize conversion value, and Target ROAS. [7]

That is powerful machinery.

But look at the object of the optimization.

If the selected outcome is an inquiry, the system is pursuing inquiries. If the value supplied is gross purchase revenue, it is not automatically receiving your margin after fulfillment, returns, and service costs.

A value-based strategy can distinguish values you provide. It cannot turn arbitrary values into sound economics.

Give every form submission the same estimated value, regardless of whether it becomes useful business, and you have left an important distinction out of the objective.

My concern is the feedback that follows: if commercially useless inquiries satisfy the selected goal, the bidding system can receive a success signal when the business experienced a failure.

It may find more useful customers. It may also pursue patterns that produce the measured action without enough of the value you intended.

That is a risk to test, not a claim that every automated campaign behaves identically.

The disadvantage is that an efficient pursuit of a weak objective can make the business worse faster.

The Other Disadvantages Don't Disappear With Correct Tracking

Accurate tracking matters. It does not remove prediction error, delayed feedback, incomplete visibility, or the cost of trying things that don't work.

Your budget is available to be spent. Google says Maximize conversions aims to spend the budget to obtain conversions and warns that switching from an underspending campaign can significantly increase spend. A budget you casually treated as a ceiling can become a much more active spending instruction. [8]

A target is not a guarantee. Target CPA seeks an average cost per conversion. Individual conversions can cost more or less, and the achieved average can differ from the target. It is not a contractual maximum acquisition price for each acceptable customer. [8]

Meaningful outcomes can arrive late. Google accounts for conversion delays and can use information beyond a new campaign's own history. But waiting for reliable downstream evidence still takes time. A business with sparse, slow outcomes has a harder evaluation problem than one with frequent, promptly recorded purchases. [9]

Learning has a cost. The business pays for advertising while uncertainty is being resolved. “It needs more data” should come with a defined question, an affordable test, and a decision point. It should not become an unlimited extension of an unsuccessful experiment.

You cannot inspect every detail. Reporting exists, but it has limits. Google says some low-activity search terms are omitted for privacy and aggregated in insights. That limits your ability to inspect every query behind the spend. [10]

These are reasons to govern automation carefully. None requires alleging that Google deliberately sent you bad leads.

Dynamic Ads Add Another Layer of Delegation

Bidding decides what to bid. Other features can affect which searches trigger ads, what the ads say, and where people land.

Those are separate decisions. Their risks can interact.

“Dynamic ads” is too loose a label to evaluate them properly.

Dynamic Search Ads. DSA uses website content to match searches, generate headlines, and select landing pages. Google recommends clear page titles and content, regular search-term review, and exclusions; it does not recommend DSA for rapidly changing websites. Its current documentation also announces an automatic upgrade to AI Max beginning in February 2027. [11]

The practical risk is that an informational page can be relevant to a search without representing an offer you want to advertise to that searcher.

Responsive search ads. Google assembles combinations of supplied headlines and descriptions. Not every supplied line appears every time. Pinning can preserve essential wording in supported positions, but the available combinations and display rules matter. Don't put the qualification that defines your customer in an optional line and assume everyone sees it. [12]

AI Max text customization and Final URL expansion. In Search, these settings can generate text and replace the supplied destination with another relevant page. Google says Final URL expansion is enabled by default when you opt into AI Max and can be disabled. Critically, its documentation says pinned responsive-search assets are not used when an expanded destination is selected; it also flags a pinning exception for URL inclusions. Pinning alone is therefore not a universal safeguard across these features. [13]

Performance Max text customization. Its Search inventory can also use generated text and expanded destinations under the relevant settings. That is another implementation to review, not another name for an ordinary responsive search ad. [14]

The attraction is less manual work and broader discovery.

The tradeoff is more decisions delegated to the platform—and more responsibility to verify that the resulting message, audience, and destination still fit the business.

A relevant page is not necessarily the right sales page. A relevant searcher is not necessarily your customer.

Same Legal Topic. Wrong Fucking Client.

Consider a hypothetical employment-law firm representing employers.

An employee searching for help suing an employer and a business owner searching for defense counsel can use overlapping language.

Both are interested in employment law.

The firm serves one side.

Suppose its website includes an educational article about wrongful termination claims. That article may be useful to business owners, employees, students, journalists, and people seeking general information.

Now imagine an automated campaign reaching an employee with a broad employment-law message and sending them to that article. The person requests a consultation.

If the configured goal is the form submission, the campaign can record success before anyone establishes that the firm represents employers.

The topic matched.

The person responded.

The business acquired an inquiry it cannot use.

This is a hypothetical failure mode, not a claim that DSA, AI Max, or Performance Max necessarily produces this result.

It shows why the controls matter: which pages are eligible, what qualifying language appears, which searches are excluded, and what outcome guides the bidding.

More explicit language, appropriate targeting, negative keywords, selected destinations, and reliable qualification can help. Each should be tested against actual inquiries.

And there is a limit that no setting removes: excluding employees does not manufacture employers who currently need counsel.

The addressable demand still has to exist.

A Big City Is Not Infinite Demand

Los Angeles and New York City contain millions of people. That does not mean millions of people currently need the particular service your business sells.

Start narrowing the market. The person needs the correct service, on the correct side of the transaction, in the geography the business serves. They need the authority, budget, timing, facts, and willingness to buy. The company still needs the capacity and economics to serve them.

A specialized employer-side employment firm does not sell to the population of Los Angeles. It sells to the much smaller number of employers who need that kind of counsel now and are a credible fit for the firm.

The same applies outside legal services. A commercial contractor does not acquire the entire population of New York. A specialized medical practice does not serve every patient. A high-ticket service business can operate inside an enormous metropolitan area and still face a small addressable market.

Automation can search more auctions, test more combinations, and identify patterns a person may miss. It cannot manufacture qualified demand that does not exist.

As reach expands, the next available conversion may come from a person who is easier to persuade but less useful to the business. That is a market constraint to measure, not automatic proof that the advertising system malfunctioned.

A large city can contain a very small market.

A Better CPA Can Hide a Worse Business

Here is a hypothetical analysis based on an understanding of how these systems work. These are invented figures, not results from a client account or evidence that automation caused a particular outcome.

Compare two possible campaign outcomes with the same $20,000 advertising spend. Assume sufficient time has passed to evaluate the resulting customers, with consistent definitions and one counted inquiry conversion per prospect.

Outcome B produces twice the conversions at half the reported cost per conversion.

It also produces fewer qualified inquiries and fewer customers.

Assume each customer contributes $3,000 over the same evaluation horizon after delivery and other variable costs, but before this advertising spend and fixed overhead.

Outcome A produces $30,000 in contribution. Subtract advertising and $10,000 remains.

Outcome B produces $12,000 in contribution. Subtract advertising and the result is negative $8,000.

Neither calculation is a complete company profit statement. Additional intake burden could make the second outcome worse.

The point is the direction of the numbers.

The conversion report improved. The economics deteriorated.

This comparison does not establish causation. It establishes why cost per reported conversion cannot settle the argument.

Personal Injury Makes the Feedback Problem Harder

A personal injury firm can receive a call today, sign a client later, and collect a fee much later—if the matter produces a fee at all.

A call is not an accepted case. An accepted case is not collected revenue. Collected revenue is not profit after acquisition costs, case expenses, and the work involved.

For that business, optimizing to calls gives relatively prompt feedback but a weak picture of the final economics. Waiting only for final collections can leave the advertising decision far behind the current market.

That is the operating problem: choose a timely signal with a defensible relationship to useful business, then test that relationship against outcomes as they mature.

Estimate when necessary. Label the estimate. Revisit it.

Do not give every signed case an invented value and call the resulting return on ad spend a financial fact.

And do not use fees collected from older acquisition cohorts to declare this month's campaign profitable.

The same distinction appears elsewhere. A software trial is not a retained paying account. An e-commerce purchase can become a return. A commercial contractor can receive real calls from residential customers it doesn't serve.

The stages differ. The measurement problem survives.

Sending Google More Data Is a Decision, Too

Google's enhanced conversions for leads can connect eligible offline outcomes to advertising using identifiers and hashed customer data. That can improve the link between an initial inquiry and a later outcome. [15]

It does not repair inconsistent qualification records.

If one employee labels every answered call “qualified” and another requires an accepted opportunity, the upload distributes an unresolved business definition.

There are also boundaries on what may be sent. Google's customer-data policies impose disclosure and consent requirements where applicable and prohibit sensitive-category conversion information in enhanced conversions, including health or medical information. A legal practice should not assume that an injury-related outcome is eligible merely because its CRM can export it. [16]

Hashing is part of the matching method. It is not permission to disregard the policy or other obligations.

Define and review the specific data flow. Internal reporting can remain useful even when a proposed advertising upload is unsuitable.

The objective is better evidence, not the largest possible transfer of customer information.

Sales Can Be Wrong, Too

The criticism of automation does not excuse an operation that fails to answer its inquiries.

A suitable customer who never receives a callback is not proof of bad targeting.

An unassigned form, an unanswered phone, a poor explanation of pricing, or a broken booking process can destroy useful demand after the advertisement did its job.

Separate wrong-fit inquiries from mishandled opportunities. Review outcomes and reasons, not just the sales team's mood at the end of the week.

Google's automation can also perform well. Auction-time bidding can use signals a person cannot practically evaluate for every auction. Dynamic matching can uncover relevant demand that a manually assembled campaign missed. [7, 11]

If it produces better commercial outcomes at an acceptable full cost, use it.

My objection is to granting it the benefit of the doubt indefinitely because its own selected metrics look encouraging.

What I Would Check Before Spending More

First, break the conversion total apart. List each action, source, counting rule, assigned value, and campaign goal. Identify which figures are observed, modeled, imported, or credited after non-click interactions where the reporting allows that distinction. Don't promise a customer-level reconciliation for a modeled total.

Second, inspect the decisions being delegated. Review bidding, actual search terms available in reports, eligible pages, generated assets, required qualifying language, and destination settings. Changing a bid strategy does not by itself fix an advertisement that reaches the wrong buyer.

Third, reconcile the operating outcomes. Establish consistent definitions for qualified, rejected, duplicate, existing customer, accepted, delivered, and paid. Match records where possible, account for timing, and investigate gaps.

Fourth, evaluate the economics. Use contribution, acquisition costs, capacity, and an appropriate outcome horizon. Distinguish expected value from realized value and attribution from evidence of incremental business.

Then run a bounded test of the suspected problem. Where practical, compare a constrained setup with the expanded one while keeping the offer and other material conditions comparable. Agree on the budget, evaluation period, and business measure before the results arrive.

Google itself provides experiments and guidance on evaluating outcomes after conversion delays. Those tools support testing; they do not choose the correct business objective for you. [9]

I would favor expansion if it improved qualified customer acquisition or contribution at an acceptable cost after allowing for timing and uncertainty.

I would constrain or change it if the apparent gains came primarily from low-value actions, wrong-fit inquiries, or assumptions that failed against mature results.

And I would fix intake first if relevant demand was arriving and the business was losing it.

That is what would change my recommendation: evidence about what is actually failing.

Google's Conversion Is Not Your Conclusion

The disadvantage of automated advertising is not simply that a machine makes decisions.

It is that you can delegate targeting, messaging, destinations, and bidding while accepting a definition of success that never adequately represented your business.

Then every improvement in the report looks like a reason to delegate more.

Google can accurately report a conversion under its measurement rules while sales accurately reports an inquiry the business cannot use.

The number can be legitimate.

The purchase of that outcome can still be a bad decision.

Your responsibility is to understand the difference before spending more money to repeat it.

Because the objective was never to collect conversions.

The objective was to acquire business worth having.

Follow the fucking customer.

Then follow the money.