---
title: "AI Visibility Vendor Migration Handover Checklist"
description: "A practical export acceptance table and parallel-run checklist for moving AI visibility measurement between vendors without losing historical prompt, engine, answer, citation, and eligibility definitions."
canonical: https://authoritytech.io/blog/ai-visibility-vendor-migration-handover-checklist
last-updated: 2026-09-15
---

# AI Visibility Vendor Migration Handover Checklist

A practical export acceptance table and parallel-run checklist for moving AI visibility measurement between vendors without losing historical prompt, engine, answer, citation, and eligibility definitions.

Canonical URL: https://authoritytech.io/blog/ai-visibility-vendor-migration-handover-checklist
Published: 2026-09-15
Author: authoritytech
Topic: Machine Relations

An AI visibility vendor migration should preserve measurement identity before it preserves dashboards. The handover needs historical prompts, engine identities, raw answers, citation definitions, source eligibility rules, and exclusion logic to survive intact, or the new tool will show a cleaner report while breaking the trend line the team actually needs.

I would not treat this as a software replacement project. I would treat it as a metrology handover.

The reason is simple: AI visibility is only useful when the question shape, engine surface, answer evidence, and citation rule are stable enough to compare across time. If any of those change during a vendor switch, the team can mistake instrumentation drift for market movement.

This is the operating method I would use before switching an AI visibility measurement vendor. It is a proposed handover checklist, not proof that changing tools earns citations or improves conversions.

## Why an AI visibility vendor migration breaks trend lines

**An AI visibility migration breaks when the new vendor measures a different question, a different engine surface, or a different citation event than the old vendor.** The dashboard can still look precise, but the historical comparison is gone.

The search signal that prompted this checklist came from an existing AuthorityTech page about [BrightEdge alternatives](/blog/brightedge-alternatives-2026-ai-search-visibility). In the Google Search Console export generated September 14, 2026 at 09:05 UTC for August 14 through September 11, the query-page pair for “competitors of brightedge” generated 3,995 impressions, zero clicks, and an average position of 5.56. Those impressions are not unique searches, and they do not prove buyer intent. They show that comparison demand is visible enough to deserve a sharper operational answer.

That answer is not another alternatives list. It is the question behind the list: if a marketing team changes the system that measures AI search visibility, what has to survive so the team can trust the before-and-after view?

The same boundary applies to crawl and model-panel data. AuthorityTech's September 14 crawl log showed 2,898 requests across AI agent, assistant, crawler, and search classes, but those requests are not buyers. The September 14 synthetic engine panel measured 35 monitored prompts with 34 successful runs and one partial run, but synthetic prompts are not organic demand. They are instrumentation.

Migration work fails when teams confuse those categories.

## The export acceptance table for AI visibility measurement handovers

**The export acceptance test should start with raw measurement primitives, not summary metrics.** A new vendor can rebuild a dashboard if it receives the primitives. It cannot rebuild history from aggregate scores alone.

Use this table before approving the export from the old vendor or accepting the import into the new one.

| Handover object | What must export | Acceptance test | Why it matters |
|---|---|---|---|
| Prompt history | Exact prompt text, prompt ID, prompt group, market, language, run date, and run cadence | A sampled prompt reproduces the same text and grouping in the new system | Trend lines break if the question shape changes |
| Engine identity | Engine name, surface, model or product label when available, access method, geography, and run timestamp | The new system can distinguish ChatGPT Browse, Perplexity, Gemini, Claude, Google AI Mode, and Google AI Overview when those surfaces were separately measured | “AI visibility” is not one engine |
| Raw answer body | Full answer text, answer timestamp, and any rendered source blocks | A reviewer can inspect the original answer without relying on a vendor summary | Citation disputes require raw answer evidence |
| Citation events | Cited URL, cited domain, cited title when available, citation position, citation type, and whether the citation was a source, mention, grounding link, or recommendation | The same answer can be rescored from raw events | Share-of-citation definitions must be portable |
| Eligibility rules | Which source types counted, which were excluded, and which answer states were eligible for scoring | A historical answer produces the same eligible and ineligible citation set | Vendor changes often hide rule changes inside “improved scoring” |
| Entity matching | Brand aliases, product names, competitor aliases, canonical domains, and disambiguation rules | Known false positives and false negatives stay documented | Entity drift creates artificial wins and losses |
| Failure states | Timeout, blocked, no-answer, no-source, partial answer, and degraded-run flags | Failed runs stay visible rather than disappearing from the denominator | Silent failures inflate apparent visibility |
| Denominators | Total eligible runs, monitored question-shape runs, successful runs, partial runs, and excluded runs | Every percentage has a reconstructable denominator | Percentages without denominators are not measurement |
| Versioning | Vendor version, scoring version, prompt version, rule version, and export checksum | A future analyst can tell which method produced which number | Migration needs auditability, not just continuity |
| Evidence files | Raw JSON, CSV, screenshots where required, and API payload samples | A technical reviewer can rerun a small fixture without vendor UI access | Screenshots alone are not enough |

This is boring on purpose. The boring fields are what keep a future board slide from lying.

## The parallel-run handover checklist for AI visibility vendors

**A parallel run should prove measurement continuity before the old system is shut off.** The goal is not to make two vendors match perfectly. The goal is to know which differences are caused by method changes rather than market changes.

Run both systems together long enough to answer six questions:

1. **Are the prompts identical?** Lock the exact text, grouping, language, geography, and run cadence before comparing results.
2. **Are the engine surfaces identical?** Do not compare an old “ChatGPT” label to a new “ChatGPT Browse” or “ChatGPT web” label unless the access method is documented.
3. **Are raw answers retained?** The team should be able to open the old and new answer bodies side by side.
4. **Are citation definitions equivalent?** Decide whether a citation means a source link, a mention, a grounding URL, a recommendation, or a qualified placement in a source block.
5. **Are denominator rules equivalent?** A timeout, partial answer, blocked answer, or no-source answer has to land in the same denominator category.
6. **Are source eligibility rules equivalent?** Decide whether owned pages, third-party earned media, Reddit, forums, directories, review sites, and search-result snippets count.

Here is the handover sequence I would use.

| Phase | Operator action | Evidence to keep | Exit condition |
|---|---|---|---|
| 1. Inventory | Export the complete old prompt set, engine roster, source rules, entity aliases, and failure states | Prompt registry, engine registry, scoring rule file, entity map | Every historical metric can be traced to a prompt, engine, run, and rule version |
| 2. Fixture build | Select a fixed sample of historical runs across successful, partial, timeout, cited, uncited, and ambiguous answers | Raw answer fixtures, citation events, expected scores | The new vendor can ingest or mirror the fixture |
| 3. Dry import | Load the fixture without changing live reporting | Import logs, rejected rows, transformed fields, checksum | Every rejected or transformed field is explained |
| 4. Parallel run | Run old and new systems on the same prompts and engine surfaces | Side-by-side answer bodies, citation events, denominators | Differences are labeled as method, vendor access, or real answer variation |
| 5. Reconciliation | Compare source eligibility, entity matching, and denominator effects before comparing scores | Exception register, rule-diff table, approved mapping | The team can explain why two systems disagree |
| 6. Cutover | Freeze the old system, record the last trusted export, and mark the new system's first production run | Final old export, first new export, version stamps | Reporting notes disclose the measurement break and preserved continuity |

Do not start with “which vendor has the better score.” Start with “which vendor lets us defend the score.”

The fields above also map to source-owned documentation patterns that should not be collapsed into a vendor's private score. Search Console warns operators to use the [Performance report data freshness note](https://developers.google.com/search/blog/2019/09/search-performance-fresh-data) before reading recent query data as final. Google documents that AI experiences can appear as [AI features in Search](https://developers.google.com/search/docs/appearance/ai-features), which is why the surface label belongs in the export. Cloudflare publishes separate [bot traffic controls and categories](https://developers.cloudflare.com/bots/), which is why crawl behavior belongs outside buyer claims. OpenAI describes [web search as a tool surface](https://platform.openai.com/docs/guides/tools-web-search), Anthropic documents a separate [web search tool for Claude](https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/web-search-tool), Perplexity exposes its own [answer API and model surface](https://docs.perplexity.ai/getting-started/overview), and Microsoft documents [Bing Web Search API response behavior](https://learn.microsoft.com/en-us/bing/search-apis/bing-web-search/overview). Those are not interchangeable measurement sources.

For the handover file itself, use boring provenance standards. The W3C's [PROV overview for provenance interchange](https://www.w3.org/TR/prov-overview/) explains why source, activity, and agent relationships should travel with evidence. NIST defines [provenance as source-and-history information](https://csrc.nist.gov/glossary/term/provenance), which is exactly what a migration loses when only chart images move. The JSON Canonicalization Scheme in [RFC 8785](https://www.rfc-editor.org/rfc/rfc8785) is a practical model for stable checksums. Schema.org's [Dataset vocabulary](https://schema.org/Dataset) shows the metadata shape that makes exported evidence legible. The IETF timestamp profile in [RFC 3339](https://datatracker.ietf.org/doc/html/rfc3339) gives the date-time discipline every run log should use.


## How Machine Relations changes the migration question

**Machine Relations measurement is not just rank tracking with AI labels.** [Machine Relations](https://machinerelations.ai) is the discipline of making a brand legible, credible, and citable inside AI-mediated discovery systems, which means measurement has to preserve the evidence trail behind each answer.

That is why the [Machine Relations Stack](https://machinerelations.ai/stack) matters during a vendor migration. If earned authority, entity resolution, citation architecture, distribution, and measurement are separate layers, the measurement layer cannot rewrite the lower layers during a tool switch.

The September 15, 2026 Machine Relations Index release, `mri_score_v2.0+2026-09-15+e512364a281a`, covers May 10 through September 15 across 122 observed days. In that release, Reddit appeared in 30 of 132 AI-visibility-GEO `x_vs_y` monitored question-shape runs, or 22.73%, across seven run dates. That is a domain citation rate within monitored runs. It is not purchase preference, and it is not proof that Reddit caused a buyer to choose anything.

The migration lesson is narrower and more useful: source eligibility must be explicit. If one vendor counts Reddit threads, another excludes them, and a third folds them into a generic “community” bucket, the team has changed the measurement system. It has not learned whether visibility changed.

This is where [share of citation](https://machinerelations.ai/glossary/share-of-citation), [citation architecture](https://machinerelations.ai/glossary/citation-architecture), and [AI visibility](https://machinerelations.ai/glossary/ai-visibility) have to be defined before procurement signs off. Otherwise the same executive report can contain three incompatible meanings of “we were cited.”

## What the migration report should disclose

**Every AI visibility migration report should disclose freshness, scope, and non-claims in plain language.** If the caveat would change an executive decision, it belongs in the report.

For this operating method, I would include these notes:

- The GSC evidence that motivated the topic was generated September 14, 2026 at 09:05 UTC for August 14 through September 11 and carries normal Search Console lag.
- The 3,995 impressions on “competitors of brightedge” are query-page impressions, not unique searches, buyer counts, or conversion evidence.
- Cloudflare and crawl totals show machine or web request behavior, not purchasers.
- Synthetic engine panels show monitored prompt behavior, not organic demand.
- The September 15 MRI release is current by native owner proof and marked READY, but an independent public HTTP fetch returned 403 during verification, so the public bytes were not independently rehashed in that run.
- Novelty for this article is provisional because the cross-publication graph was stale and public sitemap requests returned 403 during source inspection.

That last bullet is not weakness. It is the discipline. When the evidence is bounded, the claim should be bounded too.

## What to keep out of the migration decision

**A vendor migration cannot earn AI citations by itself.** It can preserve visibility history, improve evidence access, or make scoring easier to audit. It cannot replace the work of earning credible third-party sources that engines already trust.

Keep these claims out of the decision memo:

- “The new vendor will improve our AI visibility.” It might improve measurement. It does not create citations.
- “A better score means the market changed.” It might mean the method changed.
- “More AI crawler requests mean more buyers.” They mean more requests.
- “Synthetic prompt success means organic demand.” It means the monitored panel completed.
- “One source class is always good or bad.” Source eligibility depends on the question shape, engine surface, and citation definition.

The better decision memo says: we preserved the old measurement identity, documented the rule differences, ran both systems in parallel, accepted the new system only where the evidence trail survived, and disclosed the break where it did not.

That is the grown-up version of switching vendors.

## FAQ

### What should survive an AI visibility vendor migration?

The exact prompts, engine identities, raw answer bodies, citation events, source eligibility rules, entity aliases, failure states, denominators, scoring versions, and export checksums should survive an AI visibility vendor migration. Without those fields, the new vendor can show activity but cannot preserve measurement history.

### How long should two AI visibility vendors run in parallel?

Run them in parallel long enough to cover successful answers, partial answers, timeouts, uncited answers, cited answers, and ambiguous entity matches across the priority prompt groups. The point is not perfect score matching. The point is explaining every material difference before the old system is shut off.

### Should Reddit count as an AI visibility citation source?

Reddit should count only if the measurement rule says Reddit is eligible for that question shape and engine surface. In the September 15, 2026 MRI release, Reddit appeared in 30 of 132 monitored AI-visibility-GEO `x_vs_y` runs, or 22.73%, but that is a domain citation rate within monitored runs, not buyer preference or causality.

### Is switching AI visibility vendors the same as improving AI visibility?

No. Switching vendors changes instrumentation. Improving AI visibility requires the brand to become more legible, credible, and citable across the sources AI engines use. The vendor migration should protect the measurement history while the [AuthorityTech visibility audit](https://app.authoritytech.io/visibility-audit) shows where the underlying citation system needs work.

## Links

- [Blog Index](https://authoritytech.io/blog.md)
- [Home](https://authoritytech.io/index.md)