# AI visibility monitoring: a trial you can inspect

Protocol: `seoi-monitoring-retention-v1`  
Prepared: 7 October 2026  
Status: **Not run. No product results or timed workflow observations are included.**

Use this packet to check whether a permitted tool can collect and retain the evidence you need. The example uses our own business, SEO Informatica, and our public website. It is a declared example, not a representative buyer-prompt sample. We have a commercial interest in the business being observed.

## The decision

Can a reviewer reconstruct two collections of the same four questions from the tool's retained answers, sources, collection conditions and exports?

This evaluates evidence retention and review workflow. It does not establish a best tool, national visibility, ranking effects, revenue or a general accuracy score.

## Fixed inputs

| Input | Requested condition |
| --- | --- |
| Business | SEO Informatica |
| Website | `https://seoinformatica.com/` |
| Brand definition | Exact business name `SEO Informatica`; record variants separately |
| Own-domain definition | Host `seoinformatica.com` or its `www` subdomain; a name mention is not a domain citation |
| Reader | Established B2B consultancy evaluating SEO support |
| Surface | Google AI Mode, desktop |
| Market and language | United States, English |
| Collection | Four questions on each of two consecutive UTC dates |
| Question order | Q1, Q2, Q3, Q4 |
| Original source | [Our public service description](https://seoinformatica.com/ai-seo/), read 7 October 2026 |

The requested context is not proof of the actual collection context. Record what the product exposes. Unknown model versions, device settings or collection times stay unknown. If a required condition cannot be matched, retain the mismatch and do not describe that record as fully matched.

### Exact questions

| ID | Type | Question |
| --- | --- | --- |
| Q1 | Unbranded | Which AI SEO agencies help established B2B service businesses improve service pages and review buyer enquiries in the United States? |
| Q2 | Unbranded | Which SEO providers combine research, writing and implementation for a B2B consultancy selling in the United States? |
| Q3 | Branded | What services does SEO Informatica provide for an established B2B consultancy? |
| Q4 | Branded | What should a US-based B2B consultancy verify before hiring SEO Informatica for AI SEO? |

Keep the wording unchanged for this version. Branded and unbranded questions answer different questions, so do not combine them into a visibility percentage.

## Activate before collecting

For every participating product, fill these fields in the JSON method packet before launching a collection:

- Tool, exact plan/component and permission to use it.
- Included allowance remaining, its unit, any overage setting and whether this task can stay within the permitted allowance.
- Actual available surface, country, language, device and cadence controls.
- Day 1 and Day 2 as two consecutive UTC dates, identical across participating tools. Dates are currently blank.
- Any context the tool cannot expose or match, and whether that prevents participation.
- Protocol version, activation timestamp and reviewer identifier without private account details.

Four questions across two days create eight **planned observations per tool**. That is not a vendor credit estimate and does not promise eight returned answers. Check the product's billing unit separately before collection. Use only the access and allowance your team has approved.

The input set is fixed before collection. Save the completed activation packet before running anything. Do not choose the dates, questions or inclusion rules after reading the answers. Changing a material condition creates a new version with its reason.

### Documented access limits

Reviewed 7 October 2026; these are documentation findings, not account checks.

- [Ahrefs](https://help.ahrefs.com/en/articles/13192745-how-to-set-up-custom-prompts-to-track-brand-visibility-in-ai-assistants) describes custom prompts, location and cadence controls. First response data can take up to 24 hours. Verify entitlement, context and usable exports in the actual account.
- [Semrush](https://www.semrush.com/kb/1503-prompt-tracking) requires AI Visibility Toolkit access and documents desktop Google AI Mode tracking, snapshots and report exports. Verify actual export fields and remaining allowance.
- [Profound Trial](https://www.tryprofound.com/pricing) uses recommended prompts that cannot be customized and provides no exports. It cannot perform this exact task on those terms. This says nothing about the quality of another Profound plan.

## Collect and retain

1. Retain setup settings and the completed method packet. Do not silently substitute an index result, different surface or different question.
2. Use one planned collection per question per declared date. Record the product's reported collection date/time and timezone separately from when you retrieved it. If time is not shown, leave it unknown.
3. Keep the full answer and displayed source links where access permits. Retain the original file or screenshot and its SHA-256 digest. An unavailable answer stays unavailable, never “not mentioned.”
4. Inspect the brand identity and own-domain citation separately. A name appearing, a linked citation and an explicit recommendation are different observations. The worksheet records the first two; use review notes for any recommendation claim and retain its exact context.
5. Attempt the normal permitted export. Keep the original file, format, field names and hash. Record whether prompt, answer, source URL, engine and collection time survived. If only a summary exports, say so. Do not infer retention from the export button's presence.
6. Keep every planned row, including pending, failed, unavailable and context-mismatched records. An optional diagnostic retry gets a new run ID and links to the original; it never replaces a primary row or erases a failure.

An independently generated answer can differ legitimately between products. Shared inputs do not establish the same request, model version, infrastructure or generation time. Differences require inspection, not an automatic error label.

## Record effort consistently

**Setup active time:** start when configuration begins after access is confirmed; stop when the declared campaign is saved and the first collection is requested or scheduled. Pause the timer for waiting and unrelated work. Record setup once per tool in `setup_sessions`, with start/end times, pauses and active seconds. Account-access investigation is separate.

**Review active time:** in Q1-to-Q4 order, start when opening one retained record; stop after reading its answer, inspecting the cited pages needed for the decision and completing that row. Pause for interruptions. Record active seconds on that observation. If a record is unavailable, retain any troubleshooting effort separately in notes.

**Wait elapsed time:** record request/schedule and first-visible timestamps separately where shown. It is not active labour. Do not compare a tool's waiting period with another tool's hands-on setup time.

Do not fill durations from memory, use “zero” for missing measurements or extrapolate this small exercise into a general speed ranking.

## Use the worksheet

The CSV begins with eight planned rows. Duplicate the set once for each eligible product and give every actual run a unique tool-prefixed ID. Blank observed fields mean unrecorded. Preserve the supplied `not_run` status until an actual attempt occurs.

Use these status values: `not_run`, `pending`, `answer_retained`, `answer_visible_not_retained`, `failed`, `unavailable`, `context_mismatch`. A failed or missing collection does not mean the business was absent from an answer. Leave `brand_mentioned` and `own_domain_cited` blank until an answer can be inspected, then use `yes` or `no`.

Use `not_attempted`, `retained`, `unavailable` or `failed` for `export_status`. Keep unsupported fields empty and describe them in `missing_fields`. Store multiple URLs as a newline-separated cell. CSV readers must preserve quoted commas, quotes and line breaks; treat imported free text as text rather than a spreadsheet formula. JSON keeps missing observations as `null` or absent records, never invented values.

## Review the result

Report planned, attempted, visible and retained records separately, with the period and conditions. State which necessary evidence could and could not be reconstructed. Explain the setup/review effort actually measured and the action the retained evidence supports. Show failures and mismatches alongside usable records.

Do not pool proprietary visibility scores, treat a source citation as a customer visit, or attribute subsequent enquiries to the trial. A small workflow test supports a bounded buying decision, not a universal winner.

Use public business facts only. Do not place credentials, client details or private account information in this packet. Retain raw evidence locally; review permission and necessary redactions before any public release.

## Packet files

- `tool-trial-protocol-2026-10-07.md`: method and fixed inputs.
- `tool-trial-protocol-2026-10-07.csv`: eight planned observation rows.
- `tool-trial-protocol-2026-10-07.json`: method, source references, planned rows and empty actual-observation records.