What the guidance actually says
Major search engines have converged on a consistent position: automation is not itself a violation, and content produced primarily to manipulate rankings is. The method of production is treated as evidence about intent rather than as the offence.
Applied to imagery, that means a generated illustration on a page written by somebody with expertise is unremarkable. It also means a page of generated images with generated text, published at volume, is exactly the pattern the guidance was written for.
The distinction is not always comfortable, because it depends on judgement rather than on a detectable property. But it does mean the common worry, that using a generated header image will cost a page its ranking, is misplaced.
What has changed is the surrounding machinery. Image search, provenance metadata and platform labelling all now interact with imagery in ways they did not, and those are worth understanding separately from ranking.
Where images genuinely affect search performance
The mechanisms that matter are mostly the ones that always mattered, and none of them is about how a picture was made.
- Page speed. A large unoptimised image costs more ranking than its provenance ever will.
- Alt text and context. Whether the picture is described accurately and sits near text it relates to.
- Image search visibility. Original photography can rank in image results and be cited; a generic generated illustration competes with thousands of near-identical ones.
- Uniqueness. A stock image used on four hundred sites adds nothing distinguishing to any of them, generated or not.
- Usefulness. A screenshot, a chart or a diagram that helps a reader understand something is worth more than any decorative image.
The third and fifth points are where a real opportunity sits. Original photographs, annotated screenshots and purpose-built diagrams are hard to replicate at scale, which is exactly what makes them valuable when almost everything else is easy to produce.
Where provenance is starting to matter
Separately from ranking, image provenance metadata is being read by more of the ecosystem each year. Search results, social platforms and content management systems increasingly surface a label where a credential survives inside a file.
That creates a mild advantage for publishers who preserve provenance rather than stripping it. A photograph carrying a valid credential can assert what it is; one that has had its metadata removed by a build pipeline cannot, and increasingly that silence is noticeable.
The practical step is small and usually overlooked: check whether your own image pipeline strips metadata on upload or on optimisation. Many do by default, which quietly discards the only provenance your photographs had.
What to do about a site that already has generated imagery
Almost every publisher of any size has some, usually arriving through agencies, stock libraries or a designer using generative fill without mentioning it. Finding out is more useful than worrying about it.
-
Audit the pages that carry factual claims
Case studies, testimonials, news, before-and-after imagery and anything depicting a real person or event. This is where a generated picture is an accuracy problem rather than a stylistic one.
-
Leave decorative imagery alone
Backgrounds, abstract headers and concept illustration are not misleading anybody and are not a ranking risk.
-
Replace what claims to be evidence
A photograph illustrating something you did should be a photograph of what you did.
-
Add original imagery where you can
Screenshots, diagrams and photographs of your own work are the images that compete well now precisely because they cannot be produced from a prompt.
-
Fix the pipeline
Stop stripping metadata on upload, and record which images came from where.
The fourth step is the one worth investing in. As generated illustration becomes universal, the images that carry weight are the ones that could only have come from somebody who was actually there and actually did the work.
The images that are getting harder to replace
It is worth being concrete about what still cannot be prompted, because that is where the remaining advantage sits for anybody publishing seriously.
Annotated screenshots of a real interface, with the arrow pointing at the thing that actually confused somebody. A generator can produce a plausible interface; it cannot produce yours, and it cannot know which part of it people get stuck on.
Photographs of your own work, premises, team and process. These carry the one property generated imagery structurally lacks, which is that somebody was there. For any business selling trust rather than volume, that is the whole point.
Diagrams built from your own data. A chart drawn from numbers you collected is unique, citable and impossible to reproduce from a prompt, and it is the format most likely to be quoted by somebody else.
None of this is an argument against using generation for decoration. It is an argument for spending the effort you save on the images that only you can make.
A final practical note for anyone auditing a large site. Sort the work by how much the image is load-bearing rather than by how many pages you have. Ten case studies matter more than a thousand blog headers, and finishing the ten is worth more than starting the thousand.
Most teams find the audit surfaces a process problem rather than a content problem: an agency brief that never mentioned imagery provenance, or a stock subscription with no policy attached. Fixing the brief prevents the next thousand images from needing an audit at all.