A streamed AI answer arrives in pieces, so the first readable paragraph can precede the final result. An AI visibility record should preserve the received text, the evidence that collection finished and the reason generation stopped. Connection status, message boundaries and completion of the buyer's task describe different events. Check the provider's declared lifecycle and retain interruptions, limits, refusals and unresolved captures with their original context. Then review the complete answer for the brand, its role and the support for any recommendation. A fluent sentence alone leaves too much of that process unknown.
The first readable text can change meaning
Consider a fictional answer received by a monitoring collector. Its first text fragment contains the sentence “Choose ExampleCo.” A later fragment explains that the sentence is an advertising claim being evaluated. A reviewer who sees the prefix alone can assign the wrong role to the company name.
The problem also appears with conditions. An answer can introduce a provider, then explain a missing requirement or recommend another option. Sources and qualification details may arrive after the opening paragraph. Save the prefix when collection is interrupted, together with the point at which observation stopped.
For Generative Engine Optimization, or GEO, this affects the evidence behind a visibility report. The collector needs to establish what it obtained. The analyst then needs to determine what the obtained answer says. Keeping those decisions identifiable makes it possible to investigate a later disagreement.
Follow the boundaries used by the actual interface
Streaming has several boundaries worth distinguishing during procurement.
| Boundary | Evidence available | Decision it supports |
|---|---|---|
| Connection accepted | Transport response and content type | The declared response channel opened |
| One event parsed | A complete event under that format | This event can be interpreted |
| One content block ended | A provider-specific block boundary | That block reached its declared end |
| A message ended | Message boundary and generation stop reason | The message's ending can be classified |
| A monitoring record accepted | Complete retained evidence and the study's review rules | The observation can receive its declared review status |
Ask which of these boundaries the collection system actually records. A field called “success” needs a definition before it can support a denominator or an acceptance decision.
For an API integration, read the provider's event and response contract. For a consumer website, use evidence available from the actual completed conversation. An API event name cannot establish what happened in a browser session carrying the same provider's brand.
Event framing has its own rules
The HTML standard defines server-sent events using UTF-8 and a line-based format. Blank lines participate in event dispatch, while successive data fields can contribute to one event. A network read can end between the bytes or lines needed to reconstruct it. The parser therefore needs to maintain its state across reads. WHATWG server-sent events
The same specification requires pending incomplete event data to be discarded at end of file. A connection ending partway through an event leaves a different evidence trail from receiving the application's declared completion event. Preserve enough diagnostics to distinguish those cases without placing authentication material in a report.
Use a maintained parser or the provider's supported SDK for the relevant transport. Saving each network chunk as if it were a complete JSON object can lose content or produce misleading parsing errors. Logging only the text displayed by the interface can lose the event that explains why it stopped.
A stream event identifier also needs its stated meaning. It may help an interface resume delivery. Whether it identifies a message, an operation or a scheduled monitoring observation comes from the service's own contract.
A message ending still needs interpretation
Anthropic's Messages API documents a sequence involving message start, content blocks, top-level updates and a final message-stop event. Its stream can also contain error events and additional event types. A collector should retain the relevant events and handle the declared lifecycle for its SDK and API version. Anthropic streaming documentation
The associated stop-reason documentation distinguishes natural endings, configured stop sequences, token or context limits, tool use, paused turns and refusals. These endings carry different implications for the next application decision. In streaming, the documented stop reason arrives through a message update. Anthropic stop-reason documentation
For example, a tool-use ending can leave more application work to perform. A configured stop sequence needs to be interpreted against the request configuration. A natural ending can accompany an empty text response. Keep the returned reason available beside the assembled content.
These are examples from a named API. Other interfaces have their own fields and meanings. Record the contract and version used to map the original response into a monitoring record, so a future reviewer can inspect that mapping.
A small classification exercise
The Python example below runs offline on fictional capture metadata. It labels supplied completion inputs under a limited Messages-style review policy. A supported parser or SDK, configured against the actual response contract, supplies the inputs in a real collection system.
Every example remains in the output, including refusals, empty responses and incomplete observations. Brand, recommendation and study-eligibility labels require a subsequent review under the declared method.
import json
def capture_state(terminal_seen, stop_reason, text):
if terminal_seen is not True:
return "completion_unobserved"
states = {
"end_turn": "natural_end",
"max_tokens": "token_limit_reached",
"model_context_window_exceeded": "context_limit_reached",
"stop_sequence": "custom_stop_requires_review",
"tool_use": "tool_work_pending",
"pause_turn": "paused_turn",
"refusal": "refused",
}
state = states.get(stop_reason, "unknown_stop_reason")
if state == "natural_end" and text == "":
return "natural_end_without_text"
return state
cases = [
(False, None, "A readable prefix.", "completion_unobserved"),
(False, "end_turn", "A readable prefix.", "completion_unobserved"),
(True, "end_turn", "A finished answer.", "natural_end"),
(True, "end_turn", "", "natural_end_without_text"),
(True, "max_tokens", "A sentence cut", "token_limit_reached"),
(True, "model_context_window_exceeded", "Partial", "context_limit_reached"),
(True, "stop_sequence", "Stopped here", "custom_stop_requires_review"),
(True, "tool_use", "", "tool_work_pending"),
(True, "pause_turn", "Working", "paused_turn"),
(True, "refusal", "I cannot provide that answer.", "refused"),
(True, "refusal", "", "refused"),
(True, None, "Text arrived", "unknown_stop_reason"),
(True, "future_reason", "Text arrived", "unknown_stop_reason"),
]
records = []
for number, (seen, reason, text, expected) in enumerate(cases, 1):
state = capture_state(seen, reason, text)
assert state == expected
records.append({
"example": number,
"terminal_seen": seen,
"stop_reason": reason,
"text": text,
"capture_state": state,
})
print(json.dumps(records, ensure_ascii=False, indent=2))
The two first cases deliberately contain the same text. Supplying a stop-reason value without the required observed message boundary leaves the capture unconfirmed under this example's policy. The unknown-reason cases also remain inspectable. A future field value should prompt a mapping review.
The checks verify this local policy against its declared examples. They establish no property of a commercial collector. A supplier demonstration needs to show how actual transport events and the final message produce the recorded inputs.
Ask for an interrupted example during acceptance
Use an authorized test environment and disposable material to inspect the lifecycle. Agree on a harmless question and retain its intended collection settings before the demonstration starts.
- Inspect a normal completion. Follow the original input through the received events or supported SDK result to the retained message. Locate the ending reason and the associated source information.
- Inspect an interrupted capture. Have the supplier demonstrate a permitted interruption after some text has arrived. Preserve that partial text and the interruption evidence. Establish whether the original result remains recoverable through a supported lookup.
- Inspect a limit ending. Use an agreed test configuration that reaches a documented generation limit. Check that the record retains the reason alongside the received content.
- Inspect a multi-step result. If tools are involved, show which message requests the next step and which subsequent artifact supplies the final answer for review.
- Export the records. Check that the capture states, original reasons, message identities and content remain distinguishable in the receiving file or reporting system.
The exact interruption mechanism belongs to the authorized demonstration. An analyst should have a supported way to inspect the original operation before submitting another request. The request-recovery problem is covered separately in the monitoring timeout guide.
Preserve message and source relationships
A response can contain several content blocks. A collection can also contain several messages and tool results. Agree on which retained object supplies the answer being reviewed, and identify the source information associated with it.
Keep tool output, quoted text and the final assistant answer distinguishable. A company appearing in a tool result or quoted passage needs the appropriate label when the analyst evaluates the answer. Event order alone cannot establish that a particular URL supports the nearby sentence.
An export should let the reviewer locate the message identity, original question, assembled content, collection configuration, observed ending and associated citations. The format can differ across services. Acceptance should demonstrate the relationships required by the intended report.
Maintain the original record when a continuation is collected. A continued response can be linked to its predecessor through the interface's supported relationship and the study's method. Preserve the point where the first capture ended and the circumstances of the continuation.
Keep collection outcomes visible in the report
A complete refusal, a completed answer without the target brand and an interrupted capture require distinct explanations. Preserve all three according to the declared research method. Avoid replacing an inconvenient result with another conversation selected for its favorable answer.
Report how many observations were planned, which were submitted, which have completion evidence and which remain unresolved. Define any rate against the actual set used. The observed prefix of an interrupted answer leaves the rest of that answer unknown.
Completion also leaves factual interpretation to be done. A naturally ended answer can be inaccurate, conditional or irrelevant. Entity and recommendation review should examine the retained full content and supporting sources under the same stated rules used for other observations.
Applying the check when evaluating Semvi
Semvi by Xindar is the monitoring product. Its product team has confirmed ChatGPT and Gemini consumer-web sampling, complete-answer and citation-URL exports, and retention for at least one year with extensions tied to the agreed service term. The public overview also describes API collection as a separate surface. Semvi product scope
Ask to inspect an original completed record and its exported answer-source relationship for the proposed collection route. For an API route, request its ending and interruption handling. For a consumer-web route, request the completion evidence available from that actual interface. The procedure and fictional program above describe evaluation requirements, with product-specific behavior established through an authorized account demonstration.
Frequently asked questions
Can readable text establish that the answer finished?
It establishes that some text was received. Check the declared message lifecycle, ending reason and retained result before assigning a completion label.
Does a message-stop event establish a full buyer-facing answer?
Interpret it with the generation reason and any remaining tool workflow. The reporting method should identify the artifact that actually supplies the answer being reviewed.
Should an interrupted answer count as a negative brand result?
Retain its partial content and unresolved collection state. The missing continuation has not established the brand's presence or absence in the complete answer.
What should happen to a completed refusal?
Preserve the returned content, reason and study label. A refusal is an observed outcome that needs its declared place in the report.
Does the Python exercise prove a monitoring product passed?
It checks fictional inputs against a small local labeling policy. A product needs its own interface, record and export demonstration.
Sources reviewed November 15, 2026. A content block is a provider-defined part of a message. A stop reason explains a declared ending. A completion record connects the received content with the evidence used to classify that ending. The protocol references concern their named interfaces, while the collection and acceptance procedure is Xindar's editorial analysis.