Can AI check a marketing page for claims an advertising regulator might challenge? We compared the best single AI models available with a committee of AI reviewers, scored against real regulator decisions. On real marketing pages, the single models missed about one in four of the claims the regulator flagged. The committee found nearly all of them, and the difference came from the implied claims: the ones the page suggests without saying in words.
What we compared
Both approaches got the same task: read a marketing page and list every claim an advertising regulator could challenge.
The first approach was the single best AI model you can buy today, asked to read the page and list the claims. The second was a committee of AI models, each playing a different reviewer: a regulator’s attorney, the competitor’s lawyer, a scientist, the brand’s own counsel, and someone whose only job is to read the page the way a real customer does.
We then counted how many of the claims that actually mattered each approach caught.
How we scored it against a real regulator
To keep the scoring honest, we didn’t grade ourselves. We used real decisions from an advertising regulator, the BBB’s National Advertising Division (NAD). In these decisions, the exact claims that a real expert board formally challenged are on the public record.
We rebuilt each page as it actually ran. We ran both approaches blind, meaning neither knew which claims the board had flagged, and then checked their answers against the board’s list. The score is the share of the board’s objections each approach found.
Results: single models missed about one in four flagged claims
On real-world marketing pages, the single best AI models missed about one in every four claims the regulator flagged. The committee found nearly all of them.
| On the pages that matter (many claims, real marketing copy) | Real issues caught |
|---|---|
| The committee | ~9 out of 10 |
| Best single premium model | ~8 out of 10 |
| Another leading premium model | ~7 out of 10 |
On the hardest pages, where a single model struggles most, the gap grew. The committee caught 97% of the flagged claims; the single premium model caught 67–80%.
The committee did not win by using a bigger or smarter model than the one it was compared against. It won by reading the page from several points of view at once.
What the single AI model missed: implied claims
The misses weren’t random. The single AI reliably missed claims that aren’t spelled out in plain words, which are often the ones a real customer reacts to most:
- A flag-shaped logo and patriotic imagery that imply “Made in America”, with no evidence to back it up. The regulator flagged it. The single model never mentioned it. Our customer-perception reviewer caught it on every run.
- A customer testimonial (“I feel better now at 70 than I did at 50”) presented as a typical result. The regulator flagged it. The single model skipped it as “just a quote.”
- A slogan, “A Smaller Dose. A Smarter Start.”, implying the product is simply better. The regulator reviewed it on the merits. The single model saw a tagline, not a claim.
The reason comes down to what reading the page requires, not how many claims are on it.
Why a single AI misses them: copying versus reading between the lines
The dividing line isn’t how many claims are on the page. It’s what reading the page requires.
When a page is a tidy bulleted list (“80% of members say X”, “save $200/month”, one explicit claim per line), every approach scores nearly perfect, including a single AI. Reading that page is just transcription: copy the list.
Real marketing pages don’t sell that way. They sell with imagery, social proof, slogans and implication: a logo that suggests where a product comes from, a testimonial that suggests results, a sentence that carries two promises at once. Reading those takes inference. You have to notice the claim nobody wrote in words, and separate promises that are bundled together.
A single AI, told to “list the claims”, does what it’s good at. It transcribes the obvious and moves on. It misses what isn’t literally written, and merges claims that should be listed separately. A committee wins here because its members aren’t all looking for the same thing. One of them only asks: what would a real person take away from this page? That’s the reviewer that catches the flag, the testimonial and the slogan, the claims a single AI misses.
Limitations
A few limits worth noting:
- On simple, explicit pages, one good AI is perfectly fine. The committee’s edge only shows up on pages that sell the way real pages sell. If your page is a spec sheet, a single model will do.
- The single premium model is erratic. On the same page it caught 67% of the flagged claims on one run and 97% on another. The committee gave steady results from run to run, which matters when you’re making a decision based on the result.
- The committee never invented a claim that wasn’t on the page. Everything it flagged was real text or a defensible reading of it. There were no made-up (“hallucinated”) problems to chase down.
- On the cleanest, most exhaustive pages, the premium model is a little more concise. When everything is already caught, a shorter list is nicer. But conciseness isn’t a virtue when it comes from missing three things.
Why this matters for marketing and compliance teams
If you put a page in front of buyers, the claims that move them are rarely the ones in the bullet list. They’re the implied ones, conveyed by an impression, a piece of proof, or wording between the lines. That’s the blind spot of a single AI reviewer, and the strength of a panel that reads from more than one point of view.
We tested this against a real regulator’s own decisions. The panel beat the smartest single model on the pages where reading between the lines mattered most.
Method note: claims sourced verbatim from published National Advertising Division decisions (2024–2026), pages reconstructed as they ran, answers matched blind against the regulator’s flagged claims. Some cases post-date the AI models’ training cutoff, so the models can’t have simply remembered the outcome.