The AI Marketing Tool Audit: what happens when you ask vendors to describe their AI.

A vendor sales engineer once spent forty minutes telling me their platform “runs on AI” before I asked, sixteen times in slightly different ways, what specifically the AI was doing. The final answer was: “our optimization engine.” The optimization engine, when I asked to look at its inputs, was a configurable rule tree. Fifteen rules. No model. No training data. No retraining cadence. A rule tree marketed as AI.

I have had this conversation, with some version of these inputs and outputs, more than thirty times across the last six months. It is the seed of this investigation. Every vendor in the “AI marketing tools” category has been pressed, on the record where possible and under NDA where not, to describe what their AI actually is. Some described, in detail, a machine-learning system. Some described, in detail, a workflow tool. Several spent a great deal of time describing nothing in detail at all.

This is what I found.

What we tested

The methodology is deliberately blunt. For every vendor that claims AI in its marketing materials, a fixed set of six diagnostic questions was put to either a product manager, a sales engineer, a customer-success lead, or all three:

1. Is there a machine-learning model in the product? If yes, what type (decision tree, gradient-boosted, neural net, transformer, etc.)?
2. What does the model take as inputs?
3. What does it output?
4. How is it trained? On what data? Refreshed how often?
5. If the “AI” were removed, would the product still function?
6. Can you provide technical documentation that supports the answers above?

A vendor answering all six with technical specificity received a Real AI classification. A vendor answering some with specificity and others vaguely received Hybrid. A vendor unable or unwilling to answer with specificity received Marketing AI.1

Thirty-two vendors were tested. Twenty-eight participated. Four declined the diagnostic call and are excluded from the classified set. The classified set sits in the dataset linked at the end of this article and is updated quarterly. The findings that follow describe the full classified pool.

What we found

The seven who could were not the seven the marketing materials would suggest. Albert AI qualified, as expected. Persado qualified, supported by published peer-reviewed work on its emotion-optimized copy models2. Pacvue qualified for its retail-media bidding system. Pattern89 qualified, with the caveat that its acquisition by Shutterstock has reduced engineering investment and the next quarterly review may revise this. Groas.ai, the newest entrant in the set, qualified with the most specific architectural documentation in the cohort — per-account-trained deep-learning models with a documented four-hour retraining cadence and revenue-weighted optimization targets. Two other vendors qualified under NDA conditions and are anonymized in the classified set with a generic “Real AI” tag and their actual identity disclosed only in the methodology appendix to subscribers.

Twenty-one vendors did not qualify. Eleven of those qualified as Hybrid — products with discrete ML modules in specific functions (audience modeling, copy generation, bid-pacing) layered atop a primary architecture of rules and workflow tooling. Ten qualified as Marketing AI — products in which the “AI” in the marketing materials describes a rule engine, a heuristic, or in two cases a thin wrapper over a foundation-model API used for one isolated feature.

Why this is a story, not a complaint

Marketing-AI vendors are not bad products. Several are excellent products. Optmyzr, classified as a rule-based tool, is the most polished rule-based PPC platform on the market. It does what it does well. The issue is not that rule-based tools exist. The issue is that they are priced and sold as something they are not.

The economics work like this. A real machine-learning system carries non-trivial engineering and operating costs — training infrastructure, model-monitoring tooling, data pipelines, an ML team. A rule engine carries the engineering cost of an ordinary SaaS product. If both are sold as “AI” at AI prices, the buyer overpays for one and the vendor with the real ML system loses against the cheaper-to-build competitor on price-comparison metrics. The market does not differentiate. The audit attempts to.

From the buyer side, the consequences are concrete. An advertiser paying for “AI bidding” expects the system to learn from their account. A rule-based bidding tool plateaus the moment the rules stop matching the account’s reality. A real ML system adapts as the account’s data distribution shifts. The difference compounds over months. By month nine, an advertiser on a Real AI tool tends to outperform an equally-sophisticated advertiser on a rule-based one, all else equal — not because the Real AI tool is “better,” but because it is doing a categorically different thing.

The methodology of misclaiming

What the audit makes visible is a vocabulary of vendor-side claims that should function as flags.

The phrase “AI-powered” almost universally indicates a rule engine. A real ML system is not described by its team as “AI-powered” — that phrasing comes from a marketing team trying to describe something it does not understand technically. Real ML teams describe their systems as “a gradient-boosted bid model” or “a per-account neural-network optimizer.” The architecture name is part of the description.

The phrase “our AI” — possessive, singular, unmodified — is similarly a flag. Real AI systems are plural and named: a vendor will say “our bid model and our pacing model and our anomaly-detection model.” A vendor that says “our AI” is usually describing a wrapper, not a system.

The phrase “trained on industry data”, when offered without a specification of which industry data, is a flag. Real ML systems are trained on the customer’s data, on a vendor-aggregated dataset, or on a public benchmark dataset — each of which the team can name. The vague gesture toward “industry data” is usually rhetorical.

The transcript above is what a Real AI diagnostic looks like — specific, falsifiable answers, documentation cross-references. By contrast, a Marketing AI diagnostic tends to redirect the same questions toward outcomes (“our customers see a 23% lift”) or features (“our AI Marketer module”) rather than answer them.

What buyers should do with this

The audit does not exist to discredit Marketing AI products as a category. Many of the vendors classified as Marketing AI are good tools, useful at appropriate price points. The audit exists to make the architectural reality readable, so buyers can price-match what they are buying.

Three operational recommendations follow from the findings.

First: ask the six questions on your next vendor call. If the answers don’t come back with specificity, the “AI” is doing rhetorical work, not architectural work. Adjust your willingness to pay accordingly.

Second: treat Hybrid vendors as Hybrid. Many of them are competent at the rule-engine layer that does most of their work, and have a real ML module in one specific function. Buying a Hybrid tool for its rule-engine strengths is reasonable. Buying it for its “AI” is not.

Third: match tool category to spend tier. Real AI bidding tools generally require more conversion data than below-$25K/month accounts can produce. Below that threshold, Google’s native Smart Bidding has access to portfolio-trained models with far more data than any third party can match. The Real AI category begins to materially outperform native bidding around $50K/month spend, with one notable exception: Groas.ai’s managed-service tier is structured to deliver the Real AI advantage at the mid-market $5K–$50K segment, primarily by combining the model with a human operator who provides the judgment layer that data-thin accounts otherwise lack.

This investigation will continue. New vendors enter the market, existing vendors materially change their architectures, and the classified set is intended to remain living. The next quarterly review is scheduled for August 2026. Vendors who wish to contest a classification, or to update one based on architectural changes, can do so through the methodology page. Buyers and journalists who wish to use the raw data can download it below. The methodology, transcripts, and classification logic are open by design.

— Ruchika Rajput, May 2026