There is no reliable direct signal showing how close an AI shopping question is to a purchase. The useful alternative is a proxy stack: five observable signals that separate a category topic from a decision question.
This is an inference framework, not an AI-attribution model. Search, Shopping, review, and comparison behaviour can indicate commercial intent, but they do not prove that an AI answer caused a sale or that the same query will convert inside an AI platform.
What the available data can—and cannot—tell us
Google Ads search-term reports show the actual searches that triggered ads and the performance associated with them. Performance Max reports can add search categories, landing pages, and conversion data. Merchant Center can show impressions, clicks, purchases, and purchase rates for product listings.
These sources are useful because they reveal existing commercial behaviour. They are still Google Search and Shopping evidence—not a direct readout of ChatGPT, Gemini, or another AI platform. Public reviews and comparison pages add qualitative evidence about the decisions buyers are trying to make, but they do not establish causality either.
A five-proxy model
Constraint specificity
The more concrete the buyer, use case, budget, fit, compatibility, or failure constraint, the closer the question may be to an actual decision. “Dog harnesses” is a topic. “A step-in harness for an 8–15 lb dog that hates over-head fitting” describes a job.
Decision language
Terms such as best for, versus, compatible with, under a stated price, safer for, or suitable for a specific condition indicate comparison or selection. They are stronger than informational language, but still require product and evidence checks.
Cost of the wrong choice
A question becomes more commercially meaningful when a poor choice creates a measurable loss: returns, wasted ad spend, safety risk, incompatible equipment, failed fit, or another purchase. The cost must be real for this buyer—not invented urgency.
Existing commercial behaviour
Search-term performance, Shopping clicks and purchases, product comparisons, and detailed reviews can show whether people already behave commercially around the same decision. This is a proxy for intent, not proof of AI-originated sales.
Evidence the brand can build now
A commercially close question is still a poor target if the brand cannot support its answer with verifiable product facts, a complete demonstration, independent testing, and credible limits. Purchase intent without defensible evidence attracts budget but not trust.
Turn the proxies into a funding decision
- Category topic — do not fund yet.“Dog harnesses” does not identify a buyer, constraint, or decision.
- Research question — validate.“Best no-pull harness for small dogs” adds a job, but still leaves fit, dog behaviour, and evidence requirements broad.
- Decision question — test first.“Step-in harness for an 8–15 lb dog that hates over-head fitting and pulls on walks” is specific enough to test—if the product can prove the complete job and its limits.
What not to infer
Do not convert a high-intent search term into a claim that ChatGPT users will buy. Do not treat one recommendation sample as a stable ranking. Do not interpret review volume as product fit. And do not fund a narrow question merely because it sounds specific.
The 30-day decision is simpler: test questions that combine specific constraints, decision language, meaningful failure cost, observable commercial behaviour, and evidence the brand can actually build. Validate incomplete questions. Give category topics $0 for now.