A Google penalty for AI content is usually a reassessment, not a penalty.

A manual action is a human decision against a specific policy, recorded in the search console's manual actions report and accompanied by a message. It is applied deliberately and can be revoked after a review request.

An algorithmic reassessment is not an event at all. It is the continuous output of quality systems responding to the site's pages, and it produces gradual or date-aligned declines with no message and no report entry.

People say penalty for both, then look for a rule that was broken. If the report is empty, no rule was broken. Something about the pages changed the way they are judged, or something about the competition did.

Get this distinction right first. It decides whether your next three months are spent on a review request or on the pages themselves.

Spam policy does not mention models, and it does not need to. It describes behaviours that generated publishing makes cheap, and those behaviours can trigger action.

It appears in the manual actions report with a named section, such as spammy automatically generated content or thin content with little added value. It can affect specific directories or the whole site.

The scope matters more than the wording. A directory-scoped action means the affected pages were clustered together, which is a strong hint about what caused it and where the fix belongs.

Actions are applied to a site, not to a page, in most cases, because the assessment is about a pattern. Removing one page almost never resolves a pattern-level finding.

If you have access to the report, check it before doing anything else. It is the only definitive statement about whether a human penalty exists, and it takes one minute.

There is no policy against producing content with a model, and no field in a page that records how it was written. A manual action naming AI content as such does not exist in the published sections.

Diagnosing one anyway has a cost. It sends you towards rewording, which changes nothing that is measured, and away from the actual causes: missing evidence, duplication across your own URLs, and pages that answer nothing.

It also produces a false negative, which is worse. If you rewrite everything to sound human and traffic does not return, you conclude the algorithm is opaque, and the real problem keeps compounding.

The productive version of the diagnosis is narrower: which pages lost visibility, what do the results above them have that they do not, and is anything on them out of date or unsourced.

Start with the queries, not the pages. Losing impressions means the site stopped being considered, while losing position means it is being considered and ranked lower. The two point at different fixes.

Then list the pages with the steepest declines and check what they have in common: the same template, the same publication batch, the same topic cluster. Commonality is the cause, nine times in ten.

Compare each page against the three results above it. Not against your own ideal, against what is actually there. If those pages contain a table, a measurement or a named practitioner and yours contains an overview, the gap is visible.

Write the findings down before touching the site. Recovery work done from memory gets repeated, and repeated partial fixes look like the algorithm being unpredictable.

Work in this order. Each step makes the next one meaningful, and skipping ahead is the most common cause of a failed review request.

Rewriting the text to evade stylistic detection changes nothing that is measured, and it costs the only budget you have for adding evidence.

Deleting everything is the second most common mistake. A site with fewer but still unreliable pages has the same problem, minus its best material and its links.

Adding author bios without authors is transparent. A profile with no other work, no contact route and no history reads as decoration, and decoration is worse than absence because it looks like an attempt.

Filing repeated review requests after each small edit resets your place and signals that nothing substantial changed. One request, after a real change, is the strategy.

An algorithmic decline is not reversed by a fix in days. The systems reassess a site over weeks, and the effect usually appears as a slow recovery across a quarter rather than a switch.

A manual action review has its own queue. Weeks, not hours, is the normal expectation, and a rejected request means another cycle after further work.

Plan for one quarter of consolidation before judging the outcome. Judging at two weeks produces a second round of changes on top of the first, and then nothing can be attributed to anything.

Keep the record. In three months you will need to know which change preceded the movement, and the site itself will not tell you.

The guardrails are boring and they work. A cap on how many new URLs can be published in a week, tied to how many can be researched properly.

One evidence requirement per page: a measurement, an example, a limitation, a comparison or a decision rule, decided in the brief. Pages that cannot meet it do not get written.

One owner per cluster, accountable for whether the set is useful rather than for shipping it. Individual page ownership produces pages that nobody maintains.

A pruning review on a schedule, using click data. Sites rarely collapse from one bad batch; they drift because nobody removes anything.

Write one page that states the diagnosis, the evidence for it, the changes made with dates, and what result you expect. That document is what makes the work legible to anyone who asks later, including you.

Include the pages you removed and why, because the subtraction is the part people forget and the part that usually matters most.

Publish nothing new during recovery except pages that meet the evidence requirement. Adding content while consolidating muddles the signal and the outcome.

Then measure one thing: whether the surviving pages keep their traffic. Everything else is a story about a penalty.