1. :brain: _Regeneration still works in aged tissue — at a fraction of the efficiency. An…

3. :vertical_traffic_light: _Google can cut your ad account's impressions by more than half without disapproving a single ad — and one agency's appeal took seven months

Search Engine Land · John Horn · September 4, 2026Patient acquisition

Limited Ad Serving is an account-level status that restricts which auctions your ads may enter — auctions they are technically eligible for. It is not a suspension, a strike, or a disapproval, and Google does not publish the criteria for which auctions you lose. Google introduced it in 2023 for Search and YouTube; in 2026 it refined the policy and expanded it to Gmail, the Play Store and Discover, with broader implementation rolling out through 2028.
Horn's case study: a $3 million authorized-retailer account, with written brand agreements in hand, was limited with no warning. Monthly impressions fell from ~800,000 to ~350,000 — and they only held even that by bidding more aggressively and absorbing higher CPCs and CPAs to buy back some of the lost traffic. The appeal, contracts attached, was denied almost immediately. It took escalation to director-level Google support and seven months to lift; impressions returned to pre-limitation levels right after.
The named risk profile is the uncomfortable part. The most common trigger is targeting company or product names as keywords. Industries flagged most often: affiliates, insurance, consumer services, third-party lead generation, travel and franchises. New accounts are likelier to be flagged, but accounts with millions of dollars and years of history have been too. Separately, Google monitors your online reputation — a high volume of negative feedback or many direct reports about your ads can itself trigger the flag (or an Unacceptable Business Practices suspension).
There is no dashboard for this. Google shows an in-account notification and usually — but does not promise to — send an email. If you dismiss the notification, Horn's stated answer for confirming whether you're still limited is: contact Google support and ask.

Why it matters for an independent practice: Two of the named-risk categories are how practice marketing actually gets run — third-party lead generation and franchise / multi-location structures — and "conquesting" a competing clinic's brand name is a common enough tactic that most people running it don't know it's the top trigger. A practice can lose over half its paid impressions with no violation, no guaranteed notice, and a months-long appeal, while spend keeps flowing at worse CPAs. Three things worth doing this week: (1) actually look at the notification area of every ad account you touch — silence is not evidence of health; (2) treat a sudden impression collapse as a policy event to investigate, not a bidding problem to out-spend, because out-spending it is exactly what that agency had to do for seven months; (3) note that the reputation trigger ties patient-review management directly to paid-search survival — that is not how any practice currently thinks about reviews, and it should be.

Read the originalSearch Engine Land

4. :robot_face: _One video replaced an estimated 380 hands-on demos — and the 66% success score quietly counts the times a human stepped in to rescue it

The Rundown AI · Jennifer Mossalgue · September 6, 2026Buildable AI

Skild AI says its S1 robot foundation model can watch a human do an unfamiliar job once and then attempt it without updating its model weights — the video enters as a prompt, not as training. Demonstrated tasks absent from pretraining included plant potting, pancake and coffee preparation, and kit assembly, some running up to ten minutes across dozens of steps.
The headline comparison, at 100,000 hours of pretraining: demonstration prompting scored 66%, language prompting 9%. Skild estimates matching that single-video result through conventional training would take ~380 teleoperated demonstrations, roughly 50–100 hours — though that crossover was interpolated, not measured. With 2,000 demonstrations, the trained approach reached 86% — beating the single video.
Credit where it's due, and the reason this is the Trusted source: The Rundown checked the timing claim instead of repeating it. Recording of the human demo began at 9:16 p.m. and autonomous execution at 9:27 p.m., so the widely-quoted "11-minute turnaround" is record-to-execute — setup had actually started at 8:54 p.m.
And the number that matters most is the caveat: that 66% averages cumulative success across task steps and includes human recovery interventions. How often the robot completes an entire job unassisted is unresolved. For contrast, Generalist's Aug 19 GEN-1.5 reported 59% across ten short tasks from 3–12 seconds of demonstration, and 83% after ten gradient steps with five minutes of task data.

Why it matters for an independent practice: This is a robotics story with nothing medical in it, and the transferable part isn't the robot — it's the scoring. Every AI receptionist, scribe and intake agent pitched to a practice is measured exactly this way: per-step accuracy, with staff corrections folded quietly into the number. Before anything touches scheduling or documentation, demand the figure Skild's own report leaves open — what share of complete jobs finish with no human intervention, measured on your patients, your payers, your phone system. Second, the shape of the finding is a genuine pilot method: record one real workflow, let the tool attempt it, measure where it breaks, then feed it examples only at the failure points. That's cheaper and far more honest than buying a system on a demo, and it's the same reason to distrust "it handled 66% of calls" as a purchase signal.

Read the originalThe Rundown AI

1. :brain: _Regeneration still works in aged tissue — at a fraction of the efficiency. An Upstate lab named the thing blocking it: inflammation.

3. :vertical_traffic_light: _Google can cut your ad account's impressions by more than half without disapproving a single ad — and one agency's appeal took seven months

4. :robot_face: _One video replaced an estimated 380 hands-on demos — and the 66% success score quietly counts the times a human stepped in to rescue it