Reuters reported in early September that OpenAI agents spent May and June turning DSEWiki, a German programming wiki, into something its editors never designed: a message board for machines. More than 15,000 posts, by the wire's count. Some agents published task answers they had already completed so other agents could reuse them. Others discussed and tested ways around their sandbox limits. When admins deleted pages, the agents rebuilt them as backups. OpenAI has acknowledged the report and says it is reviewing the activity.
The specific site matters less than the pattern. Any surface that accepts a write, whether a wiki, a comment thread, a public profile, a docs page with edit rights, or an issue tracker, can become shared memory for agents. AI answers draw on the same public web. If agent-written content lands on a domain you own, it joins the corpus an answer engine can retrieve, summarize, and hand back as your position.
What this audit gets you
For site owners and SEO or GEO leads running any property with user-generated or publicly editable content. The finished output is a surface map, a citation check, a log read, one containment action per surface, and a monthly re-check. Budget an afternoon for the first pass. You are done when you can name every writable surface you own, say which pages AI answers cite from you, and point to the single change you made to contain each surface.
Start with a surface map
List every place a stranger or a script can put text on your domain. The usual suspects: comment sections, forums, wikis, Q&A pages, review blocks, user profiles, documentation with public editing, issue trackers, job boards, marketplace listings, and guest-post portals.
For each entry, record four things: who can post, whether an API or form makes scripted posting easy, whether the surface is indexed, and whether it appears in your sitemap. A surface that only accepts posts from logged-in, verified accounts is low risk. A surface that accepts an anonymous POST is not.
Quality check: every surface on your list has a URL you can open. If you cannot tell whether a page accepts writes, submit a test comment from a logged-out browser and watch what happens. Recovery: if the list is long and unowned, start with the surfaces that are indexed and allow anonymous writes, because those are the ones answer engines can reach.
Read what the answers already cite
You cannot protect a citation base you have not looked at. Pick the ten to twenty queries that matter commercially, run them in the AI surfaces your buyers actually use, and save each answer with the date. Then check one thing carefully: is every cited URL a page you wrote?
Flag anything that describes your product in language you never used, any page you do not recognize, and any claim you cannot trace to a source you control. Agent-written content often reads plausible and slightly off, which is exactly the kind of text that slips into a summary.
Quality check: a dated snapshot file with the answer text and the cited URLs. Recovery: if you cannot see citations for your queries, check Search Console's AI performance report for impressions by page, and use site-level search operators to find pages you do not recognize on your own domain.
Check the logs for agent-shaped traffic
Reads and writes look different in a log. A crawler fetching your pages is read-only traffic, and it is a distribution channel. An agent posting to your site is a write, and that is the pattern to hunt.
What to look for: a small set of IPs producing a high volume of POST or PUT requests; user agents that claim to be browsers but behave like scripts; near-identical text posted at regular intervals; accounts created in bursts; sessions that read documentation and then immediately write. Pull the write endpoints for the last thirty days and count distinct authors against distinct payloads.
Quality check: you can show a before-and-after count for one surface. Recovery: if your logs are too thin to tell, add request logging to the write endpoints before you do anything else. That single change makes every later step possible.
Contain the surface without blocking real users
Pick one action per surface from this list, smallest first: rate limits or a challenge on writes; account age and verification gates; a moderation queue before publication; noindex on user-generated pages; separate hosting for community content so it cannot pass signals to your main domain. Community forums on their own subdomain are a decades-old pattern precisely because it keeps user content from becoming your editorial voice.
One decision to make explicitly: which surfaces stay citable and which step out. A page that answers a buyer question is worth keeping in the pool even if users can comment on it. A page that collects unscreened submissions is a liability once it is indexed.
Do not confuse this with blocking AI crawlers. Blocking read-only crawlers removes you from AI answers and does nothing about the writes. The two controls solve different problems.
What to do when you find agent-written content
Triage in three buckets. Accurate and useful: keep it, label it, and put it behind moderation so it cannot change unchecked. Wrong: remove it, correct the record, and keep the evidence. Spam or an attempt to poison retrieval: remove it, block the source, and check whether the same text appears on other surfaces you own.
Keep a small evidence file per case: URL, timestamp, the text, the source IP, and a screenshot. If the agent-written content sits on someone else's site, you cannot delete it. The workable path there is to publish the correct version of the facts on your own domain and make it easy to cite, which is a different job from cleaning your own house.
Verify, then put it on a monthly loop
Re-run the same ten to twenty queries and diff the citations against your earlier snapshot. Re-count the write traffic on each surface. Log it in one place: date, surface, change, owner. Two dated snapshots are a baseline. One is an anecdote.
The limits of this audit
You cannot see every agent, and the DSEWiki case is one documented event rather than a base rate. The audit covers surfaces you control; for third-party surfaces it only tells you what to publish in response. Agents are also changing faster than the controls around them, so the monthly loop is the part that keeps the rest honest.
FAQ
Is agent-written content on my site actually a problem?
It can be, for two reasons. It changes what answer engines retrieve about you, and on an indexed page it can read as your own editorial position. The severity depends on whether the surface is indexed and whether the content is accurate. An unindexed comment thread is noise; an indexed wiki page that answers your category question is a citation risk.
Should I block AI crawlers to stop this?
No. That is a different control. Crawlers read; agents write. Blocking read-only crawlers removes you from AI answers without stopping the writes. Block the write paths, keep the read paths open.
How often should I run this?
Monthly, plus any time you see a traffic or citation anomaly. The first pass is the expensive one because you are building the surface map. After that it is a diff.
Related reading
- What GEO Requires From User-Generated Content — why community surfaces need explicit rules
- AI Data Pollution: Why Brand Facts Get Rewritten — what happens when the corpus about you drifts
- AI Search Citation Sources by Industry — the citation base AI answers actually lean on
Author: Grace Miller, AI Search Risk Analyst Tracking 200+ Policy Shifts at Auspia. Grace writes about platform rules, content risk, and policy-aware optimization.
