Synthetic media markers such as C2PA content credentials and SynthID are becoming industry standards. They are used to differentiate user (human) generated authentic content from AI generated content. It is important to know that AI models to not treat synthetic media markers as hard penalty. Instead, they look for if the markers are present or not, if present, are they tampered with.
If a Synthetic media marker is present, AI models will keep a note of it. If these markers are in excess on the brand related content, then the brand is not considered original enough and will generally have a low EEAT (Experience, Expertise, Authoritativeness and Trustworthiness) score.
If a brand’s content synthetic media markers are tampered with, then generally AI model may penalise the brand and ignore it altogether for competitors which have authentic content. AI models are trained to opt for content which can be trusted and does not lead to hallucinations.
Based on the synthetic media markers and the information they carry, LLMs assign the EEAT score to brand’s website page and brand’s digital footprint as a whole which in turn decides their inline citations within LLM answers. Low EEAT score affects AI visibility, recommendations and citations negatively.
Entity Authority and Brand Attribution
When an LLM evaluates a content source to answer a prompt, it prioritises the C2PA credentials, checks if the manifest is intact or tampered with. C2PA is a standard and not a legal requirement. However, given the high influx on AI generated content on the web, it is best to sign all content using C2PA credentials.
If a brand has original human generated content signed with C2PA, it will explicitly show AI detectors that your content, example, images are generated using a camera and not AI. It may happen that your tool like the camera may not sign C2PA attributes explicitly, then your EXIF data becomes an important indicator of the source. You should not try to strip your images of the EXIF data.
If a brand uploads unlabelled synthetic media or one stripped off of C2PA and EXIF, an LLM looking for content on the web will assign that brand a lower grounding / EEAT core, decreasing the likelihood of it being cited.
Multi-Modal Retrieval Augmented Generation (RAG)
Modern search augmented LLMs (like Perplexity, Gemini and ChatGPT) and Gen AI search engines (AI Mode and AI Overviews) process both text and embedded media when trying to answer user queries.
If a brand uses AI generated images without or altered synthetic media markers, multimodal RAG pipelines will exclude these images from image carousel answers or visual citations to avoid hallucinations or deceptive visual context.
Even if a brand’s website consists of high number of AI generated content with proper C2PA and SynthID watermarks. RAG systems will categorise them as low effort synthetic content. If all a brand is doing is putting out low quality content which is actually a derivative of content already present on the web, then again this lowers the website’s and brand’s overall EEAT (Experience, Expertise. Authoritativeness, Trustworthiness) score, leading to fewer recommendations and citations.
AI companies, while training their AI models on the internet data use C2PA and SynthID markers to check for the originality of the content. During web scraping, synthetic data is filtered out. This is done to prevent ‘model collapse’. LLMs are meant to train on human generated original content for new information. LLMs use content available on the internet for its knowledge. If they train on synthetic content which was generated using AI itself, they are just going in a loop, recycling information and polluting the web.