Web-grounded Benchmark
Run current public-web measurements, understand the score, compare valid baselines and audit billing.
A separate web-grounded method
Benchmark answers a different question from AI Visibility: what does the selected OpenAI, Gemini or Claude model return now when it must search the current public web for an ordinary, non-branded customer question? Each question runs independently through the provider's grounded API with web search required. Aetherion then scores the saved answer and its real source metadata against the site's detected brand and domain.
| Workspace | What it measures | History |
|---|---|---|
| AI Visibility readiness | Whether the site provides crawlable, canonical answer sources. | Latest crawl and readiness evidence. |
| AI Visibility Provider Run | A repeatable provider answer without current web_search; used by Content Brief and before → after retest. | Provider Run and linked task evidence. |
| Benchmark | Current web-grounded discovery, brand presence, domain citations, prominence and public sources. | Last ten web-grounded runs with valid fingerprint comparisons. |
| Monitoring & History | Deterministic changes between completed site scans. | SEO scan comparison and alerts. |
Results are not silently mixed
A strong Benchmark score does not replace AI Visibility readiness or a Provider Run retest. Aetherion keeps the methods, scores, billing and histories separate so a web-grounded result cannot falsely verify a memory-only baseline.
Prepare a fair question set
- Add non-branded keyword targets and run a site crawl so SEO Manager can build deterministic customer questions.
- Confirm the detected brand, site domain and question language shown in Benchmark.
- Review every exact question before approving the billed operation.
- Choose 3, 5 or 10 questions. A larger set improves coverage but increases token and web-search usage.
- The target brand and configured competitor brands are removed from the prompts. They are used only after each answer for deterministic evidence analysis.
Do not insert a branded test question
Benchmark is designed for ordinary discovery questions. If the reviewed prompt or its keyword contains the target or a configured competitor alias, the server blocks the run before the selected provider is called.
Run the Benchmark
Open the site
Go to Workspace → SEO Manager, open the correct SEO Site and choose Benchmark between AI Visibility and Monitoring & History.
Choose coverage
Select 3, 5 or 10 questions and wait for the exact registered estimate and prompt fingerprint.
Review the preview
Read the detected identity, language and every non-branded question. A changed context invalidates the preview and requires a reload.
Confirm the paid operation
The confirmation shows estimated credits, model cost and web-search cost separately. No provider request starts before confirmation.
Inspect grounded evidence
Open each saved answer and review brand presence, domain citation, observed competitors and every clickable source returned by the selected provider.
Repeat with the same questions
Run a later Benchmark with the same fingerprint to receive a valid before → after delta. If the question set changed, the new run becomes a separate baseline.
Understand the 0–100 score
The score is deterministic and uses only completed grounded answers. A partial run remains visibly marked Partial; failed questions do not invent a zero-valued answer or a source.
| Signal | Weight | Meaning |
|---|---|---|
| Brand presence | 45 points | Share of completed answers that mention the detected brand or cite its domain. |
| Domain citation | 35 points | Share of completed answers whose saved web sources include the site's domain. |
| Prominence | 20 points | How early the target appears among the target and configured competitors observed in the answer. |
- Brand mentions count answer text and verified target-domain evidence.
- Domain citations require a saved source URL from the target domain or one of its subdomains.
- Unique sources are deduplicated by full URL across completed questions.
- Observed competitors are derived from configured competitor domains found in the answer; they are not invented by a second AI analysis.
Read answers and sources
- Expand a question to see the exact saved answer returned for that run.
- Every consulted source surfaced by the selected provider is stored with URL, title and domain.
- Source links are displayed visibly and open the public page in a new tab.
- Aetherion does not create a citation when the provider returned no source metadata.
- Answers describe a time-sensitive public-web observation, not a permanent ranking guarantee.
Verify business-critical claims
Public search results and third-party pages can change. Treat Benchmark as measurement evidence and verify important commercial, legal or factual claims against first-party sources before acting on them.
Comparison and history
Benchmark keeps the last ten runs for the SEO Site. A comparison is valid only when both completed runs use the same pipeline version, detected identity, language and exact ordered questions represented by the saved SHA-256 fingerprint.
| State | What Aetherion shows |
|---|---|
| First run or changed fingerprint | Baseline. No artificial delta is calculated. |
| Same fingerprint | Score, brand-mention, domain-citation and unique-source deltas. |
| Partial run | The result and evidence remain visible, with Partial status and completed-question denominator. |
| Failed run | No visibility score is treated as a successful observation; inspect the per-question error and Money Flow. |
Credits and Money Flow
Benchmark has its own billing because current web search is a different provider method. The preview estimate combines expected model tokens with at least one web-search call per question. After completion, Aetherion settles the actual registered token usage and actual number of search actions once for the whole run.
| Cost component | How it is recorded |
|---|---|
| Selected model | Actual input, cached-input and output tokens under the configured model pricing row. |
| Provider web search | A separate verified usage event for each actual web-search action, using the official per-call price. |
| Customer credits | The combined provider cost passes through the active API multiplier and credit value, then appears as one idempotent run settlement. |
Actual usage may exceed the minimum estimate
A grounded answer can perform more than one search action. Aetherion's AI Security budget reserves additional headroom, while Money Flow charges only usage events that were actually registered for the run.
Recover an interrupted Benchmark
- Aetherion stores a Running benchmark with its provider, model and recoverable run ID before the first provider request.
- Double clicks are blocked and the AI Security scope permits only one active Benchmark for the user context.
- After a refresh or lost response, the page polls the same run ID instead of starting another paid operation.
- A confirmed 4xx validation error before the provider call clears the local lock immediately.
- An uncertain network or server error never creates an automatic paid retry.
Review Money Flow after recovery expires
If the protected recovery window expires without a confirmed result, inspect Money Flow and the stored Benchmark history before approving a new run. A browser timeout is not proof that provider usage did not occur.
Open Benchmark
Need project-specific help?
Open the private Help Center or create a support ticket with the affected project, route and request ID.