
TopicHow AI Reads the Web
Prompt injection is when text inside a document changes the behaviour of a model reading it. For a brand this cuts both ways, your own pages may carry text that misleads an agent summarising you, and any page accepting user content can become a channel for someone else to do it deliberately.
Why this is a marketing problem as well as a security one
Prompt injection is the first entry in the OWASP Top 10 for large language model applications OWASP Top 10 for LLM applications. The marketing consequence is simpler than the security one, an agent asked to summarise your company reads your pages, and whatever those pages say shapes the summary a buyer sees.
The audit, in four passes
- Hidden and leftover text. White-on-white copy, stale hidden divs, old A/B variants still in the DOM. An agent reads what is in the HTML, not what is visible.
- User-generated content. Reviews, comments and profiles are text you did not write, published on your domain, under your name.
- Contradictions. Two pages naming different prices or claims give an agent licence to pick either, and it will not tell you which.
- Stale superlatives. An "award-winning" claim from 2019 with nothing behind it is a claim an agent may repeat and a buyer may check.
What to do about it
Remove hidden text, moderate what you host, resolve contradictions to one source of truth, and keep claims to things you can evidence. Google's guidance on helpful content and its spam policies both point the same way, content that says one thing to a machine and another to a person is the pattern they exist to catch Google, creating helpful content Google, spam policies.
Run it as an inventory, not a scan
There is no tool that reports this, because the problem is editorial rather than technical. What works is an inventory of every place text reaches your pages without an editor seeing it.
- Review and testimonial widgets, including third-party ones you do not moderate
- Comment sections, forums and any member-generated profile
- Job listings and press releases syndicated in from another system
- Auto-generated location, category or tag pages nobody has read since launch
- Old campaign landing pages still live and still indexed
For each, ask the same question, if an assistant summarised our company using only this page, what would it say? That question surfaces more real problems than any automated check.
The two failure modes are different problems
Accidental injection is your own stale or hidden text steering a summary, and the fix is deletion. Deliberate injection is someone placing text on a surface you host to influence how an agent reads you, and the fix is moderation and access control. OWASP treats prompt injection as the leading risk for applications built on language models OWASP Top 10 for LLM applications, and the same mechanism applies to any page a model reads.
How strong is your own AI marketing? Take the free two minute scorecard and get an honest read plus the next steps that matter most.
Take the scorecardWhat good looks like afterwards
One canonical statement of each important fact. No hidden text anywhere. User content moderated or nofollowed. Claims that can be evidenced, and dated where they are time-bound. That is unglamorous work, and it is also most of what makes a brand describable accurately by a machine. If you would rather have it done than done to a checklist, our GEO service includes it.
Common questions
Can I stop an agent misreading my site?
Not entirely, and anyone promising control over model output is overselling. You can remove the material that makes a bad summary likely, which is most of the available benefit.
Should I add instructions for AI in my HTML?
Text telling a model how to describe you is exactly the pattern the policies treat as manipulation, and it can be read by anyone inspecting the page. Write the truth clearly instead.
Is this a real risk or a theoretical one?
The security case is documented by OWASP. The marketing case is ordinary and everyday, a wrong summary in front of a buyer, which is why the fix is editorial hygiene rather than a security product.
Put this to work on your business
We build exactly this, SEO, AI visibility, PPC, email and automation, as one measured system. Talk it through with a real person, no obligation.
