Google has been explicit that how content is produced matters less than whether it is useful. That is widely read as permission to publish AI drafts at volume, which is where the traffic disappears.
What actually gets discounted
Not AI authorship. Content that could have been assembled from the top ten results, whoever assembled it. If a model can produce a satisfying answer to the query by reading page one, another page-one-quality answer earns no click.
That test applies equally to a human writer summarising the same sources. AI just makes producing it cheap enough to do at scale, which is why the correlation looks like a penalty.
Where AI genuinely helps
- First drafts from your own material: call transcripts, support tickets, internal docs
- Structural editing: tightening, cutting, reordering an argument you already have
- Producing variants of something already validated, such as ad copy or subject lines
- Summarising primary sources you have gathered, with the sources cited
Where it reliably fails
Anything requiring a fact the model cannot verify. Regulatory content is the sharpest case: an Act with recent amendments produces confident, plausible, outdated output. On DPDP specifically, a model trained before the November 2025 rules will state the position of the Act alone and sound certain about it.
The operational rule
Use AI to shape material you supply. Do not use it to supply the material. A draft built from your own survey data, your own customer calls, or a statute you pasted in is defensible. A draft built from the model general knowledge is the thing that earns nothing and occasionally embarrasses you.
General information, not legal advice. Rules current as of July 2026.