Mateo’s Seven Untraceable Jobs. Payroll and Next Week’s Budget Are at Risk.

An AI-native growth platform can summarize performance and still fail a basic accountability test: which campaign produced each booked job? If seven bookings cannot be traced to specific ads, messages, referrals, or organic visits, the platform offers attribution confidence without defensible evidence.

At 7:40 on a Thursday morning, Mateo, the fictional owner of a small plumbing company in Phoenix, had yesterday’s calendar open beside a cooling breakfast burrito. Seven jobs had been booked. The platform’s summary called the day’s marketing effective and recommended putting more money into the strongest campaign.

Mateo clicked into the first booking. No campaign identifier. The second showed “digital.” Two were marked “unknown.” Another appeared to credit paid search, but the record did not show the ad, keyword, landing page, or first tracked visit behind that conclusion.

Payroll was due soon, and Mateo had planned to move part of next week’s budget into whichever campaign had worked. A confident summary could push that money toward the wrong source. Seven jobs were real. The explanation for them was not yet defensible.

A confident answer can hide a weak evidence chain

Local-service marketing attribution is difficult for ordinary reasons. A customer might see an ad, return through search, call from a different device, then book after receiving a text. A receptionist may create the appointment manually. Tracking parameters disappear. Consent settings limit collection. Call forwarding and customer relationship records may use different identifiers.

AI does not remove those gaps. It can make them harder to notice by turning incomplete records into polished language.

A defensible attribution should let an operator move backward from the booking to the evidence supporting the source. That might include a campaign identifier, timestamped visit, tracked call, form submission, conversation record, booking event, and the rules used to connect them. Not every journey will be complete. The system should say so plainly.

The useful distinction is between three kinds of output:

• Reported facts come directly from recorded events, such as seven completed bookings.
• Analysis applies declared rules to those facts, such as assigning a booking to the most recent tracked campaign.
• Inference fills a gap, such as estimating that an untagged phone call probably followed a search ad.

Combining all three under one confident label prevents the buyer from judging the evidence.

Attribution should survive a booking-level audit

Mateo stopped looking at the dashboard total and exported the seven booking records. That changed the conversation. Instead of asking whether the platform’s summary sounded plausible, he could ask what supported each row.

A practical audit starts with a small sample, not a quarterly report. Choose several recent bookings from different entry points, then try to reconstruct each path using only information the platform retains. Record where the chain breaks.

For every booking, check whether you can identify:

• the original source and the source receiving credit;
• the campaign, ad, message, or referral when applicable;
• the first tracked interaction and the conversion event;
• any identity match connecting calls, forms, messages, and appointments;
• the attribution rule and lookback period used;
• any manual edits, imported records, or model-generated conclusions;
• the raw events available for export or independent review.

A source label alone proves little. “Paid search” may be a direct observation, a last-touch rule, a CRM default, or a model’s estimate. Those paths carry different levels of confidence and should appear differently in the interface and export.

This is the same evidence problem that appears when an AI product supplies a decision without the record behind it. In Maya’s missing audit trail, the weak point is not the fluency of the answer. It is the inability to inspect how the system reached it.

Test the uncertainty before increasing spend

Current market context makes this test timely. Thryv launched an AI-native growth platform for local service businesses while reporting that SaaS represented 76% of quarterly revenue. SaaS revenue was $114.5 million, down 0.5% year over year, and the company announced a restructuring plan.

Those facts do not establish whether its attribution is strong or weak. They do show why buyers should separate a company’s product positioning, financial reporting, and restructuring announcement from hands-on evidence about a specific workflow. Each answers a different question.

Before changing campaign budgets, run a controlled check. Use distinct tracking details for a small campaign, place test inquiries through the routes customers actually use, and follow each record into the calendar or CRM. Include awkward cases: a caller who later books by text, a returning visitor, and an appointment entered by staff.

Then compare the platform’s explanation with the event trail. A useful system should expose missing evidence, conflicting signals, and attribution-rule changes. It should permit “unknown” when the record cannot support more. That answer feels less satisfying than a complete chart, but it protects the next budget decision.

Make the next budget decision from records

By late morning, Mateo had divided the seven bookings into three groups. Some had traceable campaign records. Some had partial paths. The rest remained unknown. He did not discover a perfect source for every job, and the platform did not become useless.

What changed was the decision standard.

He moved budget only toward the campaign supported by booking-level records. He flagged the partial paths for a tracking repair and left the unknown bookings unassigned. The following Thursday, the calendar still showed booked work, but the spreadsheet beside it contained something more useful than a confident summary: a source, an evidence trail, or an honest blank for every row.