TL;DR
- A source should give buyers facts they can check. A request that an assistant favor a brand or store a sales preference serves a different purpose.
- Microsoft found promotional memory-influence attempts in AI-related URLs from email traffic it reviewed. That finding does not measure how often those attempts succeeded.
- A research benchmark measured use of rewritten sources under different defenses. It did not test hidden instructions and gives no success rate for them.
- Review proposed AI features by their full text, user purpose and evidence. Keep useful summaries and accessible content; remove covert demands for future endorsement.
Review what the feature asks the assistant to do
This matters when an agency offers a small site feature as a quick way to get seen in AI search. The visible button may look harmless. The real question is what the feature sends, what it asks the assistant to do, and whether the buyer can see that request. Read the full plan before you approve the change.
What a Source-Use Benchmark Actually Measured
A September 2026 GEO Defender preprint tested source rewriting, a distinct threat from hidden instructions. The chart shows mean rates across five models. Here, the authors' attack success rate means an answer used the rewritten document as a supporting source. It does not mean the assistant obeyed a hidden command.
| Experimental setup | Answers using the manipulated source (%) |
|---|---|
| No defense | 50.32% |
| Static safety prompt | 46.98% |
| Authors' GEO Defender | 6.20% |
Source: GEO Defender, Table 1, September 2, 2026. Test: 616 instances, seven rewriting methods; one of ten candidate documents replaced. Mean across five models. Source rewriting, not hidden-instruction compliance, memory change or recommendation. Not live prevalence.
This matters to the team buying the work. A citation counter can show that a page entered an answer without showing what the answer did with it. A vendor could use that count to claim it can control which brand an AI tool recommends. Ask what the test counted and read the actual answer before you accept that claim.
The study has a further limit. A rewrite aimed at gaining citations need not add false facts or hidden commands. The study did not test hidden instructions. It cannot show whether they work in any setting. They show why a broad claim such as “AI influence” needs a clear test.
GEO Defender preprint, September 2026, Table 1. Test: 616 instances, seven rewriting methods, one of ten candidate documents replaced. Source rewriting, not hidden-instruction compliance, memory change or recommendation. Not live prevalence. GEO Defender experiment.
Real Attempts and Lab Results Answer Different Questions
Microsoft's February 2026 investigation identified 50 distinct prompt-based attempts by 31 companies to promote themselves by influencing assistant memory, during a 60-day review of AI-related URLs in email traffic. The report describes summary links designed to influence future assistant memory. These counts describe observed attempts, not verified success or a random sample of websites. Microsoft also said some previously reported behaviors could no longer be reproduced after mitigations.
| Source | Actual observation or result | What the method can establish | What it cannot establish |
|---|---|---|---|
| Microsoft, February 2026 | 50 attempts; 31 companies | The tactic appeared in observed email URL traffic | Its web-wide prevalence or stable success rate |
| GEO Defender preprint, September 2026 | Source use: 50.32%, 46.98%, 6.20% across three setups | Behavior changed within a controlled candidate pool | Any hidden-command success rate (not tested); live retrieval prevalence |
Source: Microsoft investigation and GEO Defender experiment. Field observations and controlled outcomes have different bases; do not combine them into a ranking.
For a B2B team, the relevant test is simpler than the security vocabulary suggests: is this page giving the buyer evidence, or trying to change the rules by which the assistant judges evidence? True product facts can still sit beside a bad request for trust. A fact check alone would miss that second issue.
Our guide to AI citation source support covers whether a cited passage backs a claim. That check remains useful. This review asks a separate question: whether content is trying to act as an instruction instead of being assessed as a source.
A Change Review Sheet for Marketing Teams
Use this sheet when a supplier proposes an AI summary, share link or hidden text change. It is an editorial review tool, not a security certification. Record the exact proposed text and intended user task in the change request. Have the site owner inspect the implementation before release.
| Proposed change | Buyer purpose to verify | Approval evidence | Decision if evidence is missing |
|---|---|---|---|
| A page summary | Help a reader grasp the offer | Visible summary matches supported page facts | Revise the summary |
| An AI share control | Carry a chosen task into another tool | Full outgoing text and destination are reviewable | Hold the feature |
| Text for screen readers | Make a control clear to a user | Meaning matches the control's visible function | Fix the label |
| A claim of preferred-vendor status | Explain a real, bounded qualification | Named issuer, scope, date and public proof | Remove the unsupported claim |
| A request for future brand preference | None for the reader; it serves only the brand | No buyer evidence can justify a covert preference demand | Reject that request |
Each change needs a clear owner. The owner should know why the text is there and when it needs to change. Save the approved text with the release record. The next reviewer can check it against what shipped.
Do not approve a change just because its citation count rose. The count could rise while the answer drops a key limit, gets access terms wrong or adds praise with no proof. A sound check reads the answer against the page and the buyer task. It records errors as well as mentions.
Worked Example: A Helpful Summary With an Unhelpful Extra
Consider an illustrative industrial service firm with a capabilities page. A supplier proposes an AI share button. Before release, a reviewer reads its full outgoing text and destination. It asks for a page summary and also seeks future brand preference. No real company or measured result is implied here.
Using the review sheet, the team rejects the preference request because it serves no reader task. It then reviews the revised outgoing text and approves only a plain summary request to the recorded destination. No hidden command or attack text needs to be reproduced to make this decision.
Next, the team checks that the summary preserves the service scope stated on the page. It saves the approved outgoing text, destination and owner in the release record. A later reviewer can compare these with the shipped control. The team does not need to run a live attack or test a hidden prompt to approve this change.
Review an AI feature proposal
If your team is reviewing an agency proposal with unclear AI instructions, SemShift's AI visibility audit can help assess the content, evidence and measurement plan. Bring the proposed text and the buyer question it is meant to answer.