AI-written RSA headlines: who checks the claims before they go live?

For RSA copy, I'd split it this way: let the AI generate a big batch of headlines from the landing page and offer sheet, then have a human cut anything that makes a claim the business can't back up. Models invent specifics. A free consultation the clinic doesn't offer, same-day appointments when the booking page says two weeks out. Those lines read well, so they slip through. For a dental clinic, checking every claim against the live page matters more than polishing the wording.

If the account is already converting, a safer first use is asking the AI to critique the headlines that are already running instead of writing fresh ones. Does anyone have a repeatable fact-check step, or do you just read every line against the site?

I'm not sure claim checking is mostly about reading lines against the site. Most invented claims come from the AI filling gaps in what you gave it, so the fix sits upstream. Hand it a short fact sheet covering only what the business will actually stand behind, and tell it to write from that and nothing else. You still want to read the output against the live page, since that catches whatever slips through anyway.

Agreed on the fact sheet, and I'd make it a list of things the business can prove, not things it believes. "We're fast" is a claim nobody can check. "Booking window: 10 to 14 days" you can verify against the calendar in two minutes.

Put the negatives in there too, like "no free consultation" or "no same-day slots." Models fill gaps by inventing whatever the sheet doesn't mention, so an explicit no stops a lot of that.

Explicit negatives help, but they only cover the claims you thought to write down. The ones that actually cause trouble are the plausible-sounding ones nobody anticipated, like "board-certified" or "lifetime warranty." A rule that any superlative or credential has to trace back to a line in the sheet would catch a lot of those. Would it just push the invention into the wording, though, with "trusted by specialists" standing in for "board-certified"?

Depends on whether your check looks at what a line implies or only at what it literally says. "Trusted by specialists" sails through a keyword rule, but it still tells a reader the clinic has credentials it may not have. So I'd test each line with one question: would someone reading this walk away believing something the sheet can't back up? If yes, cut it, even when no single word is false. That's much harder to automate than a traceability rule, so for a dental account I'd have a person do that pass on the final set.

Checking each line on its own doesn't work well for RSAs, because Google mixes headlines at random and a reader sees whatever combination the system picks. Two lines that are each fine can imply something neither says alone. "Dentist with 15 years experience" next to "Same-day emergency visits" can read as a full team on call, even if you have one dentist on Tuesdays. Review the pairs that can actually show together, or pin the risky lines to a position so they only appear where you've vetted them.

Sasha, pinning doesn't really fix this. A headline pinned to position 1 only stays out of slots 2 and 3, so it can still end up next to any of the other 14 lines in the remaining positions. Checking every pair in a 15-headline set means over 100 combinations, and nobody is going to do that. I'd rather make every line hold up on its own and next to anything else in the set. If a line only stays true because of the words around it, cut it from the RSA and keep it for the landing page.

Pinning is more fragile than it looks, Rowan, because Google can show any headline in any slot, so you never really control the combination. A simple rule: if a headline only makes sense next to another one, keep it out of the RSA entirely. It's also smart to keep one policy-free line per ad group among your ten headlines as a sanity check.

Marco, a policy-free line is the right call, and I'd keep it boring, something like the booking window or the city, because that's the one line you can verify in a minute. Dental is where this bites. Words like "painless" or "guaranteed results" often get flagged under healthcare policy. Google reviews each headline as its own asset, so a flagged one usually drops out of rotation while the ad keeps running on fewer lines. Nobody notices unless they check asset status after launch.