Google refuses health keywords one at a time, not by page

If you run ads for a hormone clinic, a regenerative practice, a weight loss program or a behavioral health provider, you have probably seen this refusal and been told it was your landing page. It is measurably not your landing page. Here is the test, the raw verdicts, and the three explanations the data eliminates.

Method

What was tested, and what was held constant

Two live Google Ads accounts, both managed by PELORA, both in regulated health categories, unrelated to each other and operated by different businesses in different states. Neither account had audience segments configured. Each was submitted for review in the normal course of launching a campaign. In both cases the refusal string was recorded verbatim from the account rather than paraphrased, and in the second case the per-row verdict was recorded for every keyword in the ad group.

Clients are not named. Nothing here identifies a business, a URL, or a spend figure. What is published is the account's structural profile, the policy string, and the keyword verdicts, because those are the parts that carry the finding.

Results

Both accounts returned the same string

Figure 1swipe the chart

One ad group. One landing page. One set of ad copy. Sixteen verdicts.

Because the page and the copy were the same for every bar above, no classifier of the page or the copy can explain why two differ. The keyword string is the only variable left.

Account B is the controlled test, and it is the part that matters

Every flagged and unflagged keyword below sat in the same ad group, pointed at the same landing page, carried the same ad copy, and went through the same submission. The only variable was the keyword string itself.

The remaining rows, all lab and blood terms, passed. The account removed the two flagged terms, kept the other fourteen, and reached ready to publish.

What the data rules out

Three explanations the result eliminates

1. It is not a new-account misfire

This is the explanation most often given, and it is the easiest to test. The two accounts are as far apart as two accounts get: one is weeks old and unverified with no billing history, the other has been running since 2024 with a verified payer entity and a credit balance. Both returned the identical policy string. Account age, verification state and spend history do not separate them, so they are not the cause.

2. It is not classification of the landing page or the ad copy

This was the serious competing theory, and it deserved to be, because the policy genuinely does reach creative and landing pages through its second remedy option. It cannot be the explanation here. The page and the copy were byte-identical across all sixteen rows. No content classifier can return one verdict for hormone test and a different verdict for hormone panel when the page, the copy, the account and the submission are all the same.

3. Geo-qualification neither saves a term nor damns it

This was our own working theory, and the data killed it. We had recorded internally that adding a city or region to a clinical term let it through where the bare term flagged. The table above contradicts that directly: one geo-qualified term flagged and another passed; one bare term flagged and two passed. We are publishing the correction because a rule that sounds right and is wrong costs somebody a launch date.

What we do not know

There is no rule here, and we are not going to invent one

hormone test flags. hormone panel and hormone blood test do not. hormone testing philadelphia flags. hormone testing near me does not. Nothing in Google's published policy predicts that split. A pattern that fits five observations and has no stated mechanism behind it is a guess wearing a lab coat, and guesses in this category are what get accounts suspended.

What is established is narrower and more useful: the evaluation happens per keyword, and it is not derivable from the published policy text. That means the account is the only oracle, and the only reliable method is to submit and read the verdict.

How to build against it

Five rules that follow directly from the result

1. Treat the keyword list as the gate, not the copy. Budget review time for a keyword-by-keyword verdict at the review step on every regulated-health submission. That is where the decision is actually made.
2. Submit and read the verdict. Never predict it. Two experienced people predicted which terms were safe and both were wrong, including one who built an entire fallback ad group around the prediction. That ad group is exactly what got refused.
3. Drop a flagged term rather than appealing, unless the term is load-bearing. Account B dropped two, kept fourteen, and launched. That is the cheap path and it worked.
4. Never let a "safe" ad group be a theory. If it has not been through review, it is not safe, it is a hypothesis with a budget attached.
5. Do not give a client a launch date that assumes the keyword list will be accepted. Give them a submission date and a verdict date. They are different things and the gap is where the relationship gets damaged.

Sources and method notes

Where every claim comes from

Refusal strings and keyword verdicts were transcribed from the live Google Ads interface in each account, not reconstructed from memory. Policy text was read from Google's own pages on 22 September 2026.

• Google Ads: Personalized advertising policy, which contains the Health in personalized advertising section the refusals were filed under.
• Google Ads: Healthcare and medicines. Read this one carefully, because it is an index, not a policy. It has fourteen sections and the operative rules live on the pages it links to. Treating its summaries as the rule is how we produced six errors in an earlier pass, all caught by an independent check before publication.
• Restricted drug terms, sub-page of the above.
• Prescription drug services, sub-page of the above.

If you have run this same submission and seen a different result, we want the data. Send it to hello@peloramarketing.com and we will publish it here with attribution, including if it contradicts us.