An AI monitoring export still contains every answer, yet its citation column suddenly becomes empty. The collection service has changed its response format. The original URLs are present inside a different object, while the export routine continues reading the former field.
For a buyer, this is a useful acceptance question. Can the system recognize the input contract it understands, preserve an unfamiliar response and explain which fields remain unresolved? A parser that substitutes an empty list whenever a field is missing can turn a format change into an apparent loss of citations. The examples below use invented payloads to show how to test that boundary before a reporting change reaches a customer.
Keep the response contract beside the collection label
A platform name describes only part of a record. A collector also depends on an API route, a documented response structure and the code that translates that structure into an export. Each can change on its own schedule.
Google's current Search grounding guide shows citation annotations attached to text content, with URL and text-range information. Its GenerateContent reference separately describes candidate grounding metadata, including grounding chunks and supports. These documented structures illustrate why an integration needs to identify the interface and representation it actually consumes. Google Search grounding guide, GenerateContent reference
A buyer therefore needs more than a column labeled “Gemini.” Ask which collection route produced the record and which response contract the exporter interpreted. A consumer-app observation and an API response also need their own labels. An adapter test can establish how software processes supplied data, while a consumer recommendation requires an actual consumer answer and its supporting context.
The related model-version issue concerns comparability of observations across collection conditions. A response-contract issue concerns whether the integration preserved what arrived in the record. Both deserve a dated change record. Versioning an AI visibility baseline
Define the business record before mapping provider fields
Start with the information the customer needs to retain. A small normalized answer record might contain an answer identifier, the complete text, the observed source URLs, a source-field status and a reference to the retained input.
The source-field status matters. A successfully read empty array can mean that the specified field contains zero URLs. An explicit unavailable value can mean the source information was unavailable under that contract. A missing required field can mean that the input no longer conforms to the expected structure.
Those meanings should come from a documented agreement between producer and consumer. JSON provides arrays, objects and null values. The business meaning of each value belongs in the integration contract and data dictionary. RFC 8259
Keep the original response with an approved retention and access policy. The normalized record is a useful reporting representation, and the original input lets a reviewer inspect how that representation was derived. Preserving one alongside the other also makes a later parser correction reviewable.
Two fictional formats can express the same answer
The following layouts are editorial fixtures. Their version labels and fields were invented for this explanation.
The first layout keeps the answer and source array together.
```json
{
"schema_version": "fixture/1",
"answer_id": "A1",
"text": "A fictional answer.",
"source_urls": [
"https://example.com/spec"
]
}
```
The second layout nests the answer and represents each source as an object.
```json
{
"schema_version": "fixture/2",
"answer": {
"id": "A1",
"text": "A fictional answer.",
"sources": [
{
"url": "https://example.com/spec"
}
]
}
}
```
A correctly specified adapter can produce the same normalized answer from these two fixtures. It identifies the supported version, obtains the answer identifier and text, then reads the appropriate source collection. The reference URL remains the same observed string.
An adapter written only for the first layout needs a declared treatment for the second. If it reads `source_urls` with a default of an empty array, it can produce a plausible-looking record while losing the URL. The absence of an exception would give the buyer little assurance about the result.
Give absence and unavailability different outcomes
Suppose the fictional second contract explicitly allows `sources` to be either an array or null. An array means the source field was provided and interpreted. Null means that the source information is unavailable. The field itself remains required under this particular contract.
| Fixture input | Interpretation under the declared contract | Export treatment |
|---|---|---|
| Array containing one URL object | Source field observed with one URL | Retain the URL and observed status |
| Empty array | Source field observed with zero URLs | Retain the empty array and observed status |
| Explicit null | Source information unavailable | Retain unavailable status |
| Required field missing | Input contract unresolved | Retain the raw record for review |
| Source value has the wrong type | Input contract invalid for this adapter | Record the parsing problem |
| Version is unsupported | No accepted mapping for that version | Retain the input without inventing normalized values |
JSON Schema documentation explains that listing a property does not automatically make it required. It also distinguishes a missing property from a property whose value is null. Those details are useful when documenting the fixture, but the allowed null value still requires an explicit business meaning. JSON Schema object properties and required fields
Apply the rule to the fields used by the calculation. An added optional metadata field can be preserved in the raw input while the established mapping continues working. Requiring every harmless extension to stop the export would create a different integration problem. The useful boundary is the set of fields and types on which the accepted output actually depends.
Reconstruct one small summary
Consider four invented input records. A1 has one source URL. A2 has a successfully interpreted empty source array. A3 has a complete text answer with explicitly unavailable source information. A4 uses an unsupported response version and is retained for interpretation.
Under the fixture's rules, three records can be normalized, and two have interpretable source arrays. One of those two arrays contains a URL. A proposed source-field summary would therefore show one record with a URL out of two records whose source arrays were interpreted, or 50 percent. It would also report one unavailable source field and one unsupported input.
The calculation needs that full description. A headline of “one citation in four answers” would combine an observed zero, an unavailable field and an unresolved input contract without explaining their treatment.
The answer text and recommendation classification need their own handling. A3 can contain a usable answer even though its citation information is unavailable. A4's retained input can be examined by a reviewer while its automatic normalization remains unresolved. A format problem supplies no observation about which supplier an unavailable or uninterpreted answer recommended.
Test the relationships as well as the columns
An integration can preserve the URLs and still attach them to the wrong answer. Keep the answer identity stable through the transformation and check the text associated with that identity.
If the input contains sentence-level support ranges, preserve the text representation and range convention used by the source. A changed answer string can invalidate an otherwise well-formed range. The schema mapping should identify what it preserves and which additional relationships require verification.
Also preserve observed multiplicity until the reporting rule has been declared. Two occurrences of the same URL may represent two displayed citation relationships. A separate unique-URL summary can group them later. Deleting one during normalization can erase information a later claim-level review needs.
This follows the broader data-grain requirement. One input answer, one citation relationship and one publisher summary are different records with different counting rules. Joining AI answer and citation exports
Make a format change a controlled reporting event
Before changing an adapter, retain a small set of authorized old-format inputs and their accepted outputs. Add the new-format fixtures and record which equivalences the change is intended to preserve.
The comparison should identify answer IDs, complete text, source-field statuses, URLs and any declared support relationships. Review differences explicitly. A corrected parser may recover information that the previous export omitted, so a historical summary can change even when the underlying AI answer stays the same.
Retain the former output and explain any restatement. Record the parser version and the affected records, together with the reason for the correction. An export repair then remains distinguishable from a new collection or a change in consumer visibility.
Where a contract is unfamiliar, keep the input in a reviewable state and obtain the mapping specification. Guessing the new field from a similar product can produce a tidy file whose meaning has never been established.
Use a compact buyer acceptance exercise
- Select one authorized input and its corresponding exported answer. Identify the actual collection interface and input-contract version.
- Trace the answer identity, complete text and source information through the documented mapping.
- Test an empty source array, an explicitly unavailable value and a missing required field. Record the treatment of each case.
- Supply an unsupported fixture version and a wrongly typed source value. Check whether the original input remains available for review.
- Add harmless optional metadata to a supported fixture. Confirm the behavior expected by the agreed compatibility policy.
- Reproduce a summary from the interpreted records, with unavailable and unresolved states visible.
- Agree how a future parser correction will affect historical exports and customer reports.
For a team considering Semvi by Xindar, complete-answer and citation-URL export are confirmed product-team statements. The proposed test adds a precise question about the purchased workflow's representation and change handling. Request the actual agreed format and a traceable demonstration before treating those additional details as accepted capabilities. Semvi monitoring scope and buyer verification, Semvi product page
Questions buyers ask
Does an empty URL column prove that the answer displayed no citations?
Interpret it through the export contract and retained input. The field could represent an observed empty collection, unavailable source information or a mapping problem.
Should every newly added input field stop an export?
Use the documented compatibility policy. Validate the fields and relationships required by the accepted output, and preserve additional input information for review.
Can the same model name guarantee the same response structure?
Record the actual API interface, response contract and adapter version. A model label alone leaves those integration details unspecified.
Can a parser correction change an old report?
Yes. It can recover or reinterpret information already present in the original input. Keep the original report and document the affected records and revised calculation.
What did the examples establish?
They establish the expected outcomes of the stated fictional contracts and normalization rules. Vendor behavior requires evidence from the actual authorized integration and export.
