GEO reports: testing isolation between clients
The report has the right logo, but a row, citation or total comes from another client. Checking the cover will not catch that mixture. Nor is it enough for the dashboard selector to show the expected name: the query, stored result and delivered file may follow different paths.
This guide proposes tests for an agency generating its own GEO reports or combining outputs from several tools. The result is a record of cases, synthetic data, expected results and delivery evidence. It does not certify a provider's security or announce a Mentio feature. Tests are limited to authorized systems and environments, without real client data or attempts to access other people's accounts.
Define what each report may share
Write an ownership matrix before preparing tests. Distinguish client, project, brand, market, period and report version. Do not assume a brand equals a client: an agency may manage several brands for one client or analyze a competitor in an authorized study.
Separate private data from deliberately shared references. An aggregate benchmark may be approved for several recipients; another project's answers, private prompts or notes do not gain that permission by being in the same database. Document which dataset each report may use and who accepts that rule.
The guide to GEO for agencies with multiple clients covers service operations. The question here is narrower: whether the report preserves agreed ownership from its inputs through to the file each recipient receives.
Build fixtures that make a mixture visible
Prepare a synthetic dataset for each test context. Use unambiguous markers absent from production and quantities whose expected result can be calculated. Preserve relationships between records too: a case with no structure only checks that a name is copied.
Include local identifiers that coincide across test clients, similar brand names, different periods and projects with no observations. These cases check whether the system relies on an incomplete key or reuses a previous result. Do not introduce secrets or real conversations to make a test more convincing.
Save each expected result before running the case: permitted rows, excluded rows, total, denominator and state when data is absent. Markers detect foreign text; counts detect contamination in aggregates that no longer show their source records.
Review joins before looking at the chart
Follow the path from observation to output table. For each join, identify the key preserving ownership and check that the result retains the authorized context. An identifier unique only within one client is insufficient for joining records across all clients.
Test relationships with several answers or citations per observation. Count units before and after the join and compare them with the expected result. Row multiplication can alter a share without introducing an unfamiliar name; filtering the visible table afterward does not necessarily repair an aggregate calculated earlier.
Include the empty case. The report should show the defined no-data state rather than retain the last available table. Also distinguish an authorized shared benchmark from a foreign row: blocking every competitor name would be an incorrect test for competitive research.
Check caching and background work
The OWASP multitenant security guide separates database, cache and context isolation. For protected caches, it recommends authorization before reads and including attributes that change the result or its authorization. Separating keys does not replace access control.
In the test environment, compare generation without a previous result against a run reusing stored state. Alternate authorized contexts and request the same report again. Its result should still match its fixture, even after another test client generated a similar-looking output.
For queued work, check context when creating, executing and collecting the job. A task finishing after the selector changes should not deposit its result in the new context's report. Retain the job identifier and produced artifact, not just a dashboard screenshot.
Inspect every format leaving the system
The dashboard and its export are different outputs. Check tables, totals and citations in each format actually delivered. In a spreadsheet, inspect auxiliary sheets and included ranges; in a document, inspect appendices and metadata that form part of the file.
| Test | Expected check | Evidence to retain |
|---|---|---|
| Coinciding local identifier | Only authorized-context records appear | Inputs and resulting rows |
| Several citations per observation | Total respects the defined unit | Counts before and after joining |
| Project without data | Agreed empty state appears | Generated empty artifact |
| Previously stored result | Another context's output is not reused | Request order and results |
| Deferred job | Result retains its logical recipient | Linked job and file |
| Export | Each sheet or section preserves ownership | Exact reviewed file |
Check destination as well as content. A correct file sent to the wrong recipient remains an incorrect delivery. Use authorized test inboxes, folders and links; do not send to real clients to validate the workflow. If the flow includes shared links, check access within the agreed scope without expanding permissions to make the test pass.
Use negative controls without mistaking errors for success
Define what each authorized case must exclude. Check that the other fixture's marker is absent and aggregates match permitted inputs. Marker absence does not establish isolation if the file is empty because the generator failed.
Keep "correct result," "mixture detected," "execution failed" and "not checked" separate. A test that could not run is not a passed test. If testing a request outside the permitted context, use only synthetic resources in your authorized environment and require the policy's expected response, without foreign data in error messages.
Do not reduce these checks to searching for the client's name. A mixture may affect denominators, sources, notes or filters without copying that name. The prompt dataset privacy review addresses input-content exposure; this QA checks which report receives each datum during processing and delivery.
Block delivery when a mixture remains unresolved
Write acceptance criteria before closing the review. A confirmed mixture should block the affected report until it is corrected and its path checked again. Do not offset a foreign row with a high percentage of passed tests or conceal the issue by removing that row from the final file.
Assign an owner, known or unresolved cause and reviewed scope. After correcting the issue, repeat the failed case and cases related to the same join, cache or exporter. Retain both evidence versions so the correction can be reviewed without reconstructing it from memory.
AI visibility measurement governance organizes owners and versions. Link this QA's closure to the specific generator version, template and data rules. Do not approve "all future reports" because one delivery passed its tests.
Deliver a record another person can reproduce
Include the ownership matrix, fixture version, executed cases, expected and observed results, final artifacts and limitations. Identify formats and paths left outside scope. The conclusion should concern the tested set and reviewed version, not a universal isolation guarantee.
Before delivering the real report, verify that it matches the approved client, period and version, and is the exact file reviewed. A hash can compare files but does not establish that their content is correct. Keep approval and destination linked to the artifact.
Repeat tests when keys, the data model, deferred processing or an export template change. The review's value lies in detecting a specific regression before delivery, not adding a security checkbox nobody can justify with evidence.
Frequently asked questions
Is checking the client's name and logo enough? No. Compare rows, aggregates, sources and files with authorized inputs. A mixture can alter a result without displaying another client's name.
Can I use real data to make tests more representative? This protocol uses synthetic fixtures in authorized environments. Preserve the relationships needed for testing without introducing private conversations, credentials or real client records.
Does an empty export pass the isolation check? Not necessarily. It must match the expected result; a failure preventing content generation is a failed execution, not evidence of correct isolation.
Does a shared benchmark always count as a mixture? No. Its use must be authorized and documented in the ownership matrix. The test distinguishes permitted shared references from another context's private data.
Does this QA certify a GEO platform's security? No. It checks cases, paths and versions within an authorized scope. It does not replace a full security review or substantiate behavior that has not been tested.
Want to know if AI mentions your brand?
Discover your visibility in ChatGPT, Claude and Gemini in minutes.
Related articles
GEO for Agencies: How to Measure and Report AI Visibility for 10+ Clients
Operational guide for marketing agencies offering GEO services. Pricing, tech stack, client reporting and the mistakes that erode margins.
GEO AnalyticsAI Visibility Measurement Governance: Ownership, QA and Version Control
Set owners, a data dictionary, quality controls, traceability and versioning so your AI visibility measurement remains comparable and auditable.
GEO AnalyticsPrompt datasets: reducing exposure of customer data
Prepare prompts from tickets and CRM records with data minimization, identifier review and outbound checks while preserving purchase intent.