Key takeaways
- A tool migration changes your numbers on its own. Different prompt phrasing, sampling, scoring formulas and collection methods can all move a visibility score without a single real change in how AI talks about the client.
- Run the old and new tools side by side on identical prompts for at least two to four weeks before cutting over. That overlap is the only honest way to separate tool differences from real movement.
- Export everything from the old tool before you cancel: prompts, raw responses, citations, competitor lists and scores. Raw responses are worth more than dashboards.
- Keep an annotation log of AI platform changes (the GPT-5.3 rollout on March 4, 2026, ChatGPT's site: operator shift on August 8, 2026, Reddit's citation drop around August 14, 2026) and avoid cutting over near them.
- Tell clients about the break in the series before they find it. A labelled break with a bridge factor is a normal reporting event. An unexplained cliff is a lost retainer conversation.
Why baselines break when you switch tools
Every agency that has moved a client between visibility tools knows the moment. The new dashboard shows 31% where the old one showed 48%, and someone on the client side asks what went wrong. Nothing went wrong. The two tools just measure different things under the same name.
There are at least five reasons the numbers diverge:
- The prompt sets differ. Even when you paste the same list in, tools may rewrite, normalize, or group prompts differently.
- The collection method differs. One tool may scrape the real chat interface, another may call an API. User-facing answers and API answers are not the same thing.
- The sampling differs. Frequency of runs, number of repetitions per prompt, and run time of day all change results, because AI answers vary from run to run.
- The metric definitions differ. Visibility score, share of voice, position and sentiment are each calculated their own way, and a tool that counts a mention only when the brand is named in the first paragraph will score lower than one that counts any mention.
- The location and language settings differ. A monitor pinned to one country and language will not match a global one.
None of these are bugs. They are just why you cannot treat a score from tool A and a score from tool B as one continuous line. The playbook below is about building a bridge so the client's history still means something.
One caveat before we start. The research behind this guide did not cover each vendor's export limits, import formats or current pricing, and I found no expert guidance written specifically about migration. So treat the steps here as a method that follows from how the data behaves, and check each vendor's docs and pricing page before you commit to a timeline.
The AI engines move too, and that confuses everything
Here is the part that makes migrations risky. You are not measuring a stable thing. The AI platforms change their behavior overnight, and if that coincides with your cutover, you will blame the tool or the client's content for something neither caused.
Three examples from Promptwatch Data in 2026:
- ChatGPT rolled out GPT-5.3 on March 4, 2026. According to Promptwatch's citation drop report, ChatGPT cited about 6.4 sources per search-enabled response the week before and about 4.7 to 4.9 by late March, roughly 27% fewer. GPT-5.3, GPT-5.4 and GPT-5-Mini dropped within a day of each other, and there was no recovery a month later.
- On August 8, 2026, ChatGPT Search started using the site: operator at scale. Promptwatch's site: operator report shows the share of fanout queries using it going from 0.37% to 16.8% in a day, with average fanout queries per response rising from about 1.08 to about 1.83. Fanout and keyword mappings built before that date are stale.
- Reddit's share of ChatGPT citations held at 3.83% from July 18 to August 7, then fell under 1% on August 14. The Reddit citations report averaged 0.52% for August 14 to 17, an 86.4% relative drop. Promptwatch itself notes that a data-collection issue can't be ruled out, so the size of that drop is provisional.
That last point is the useful one for agencies. A vendor-side collection change and a platform-side change look identical in a dashboard. Both show up as a sudden step. If you are mid-migration, you will not be able to tell which one you are looking at unless you have a parallel feed to compare against.
For a wider view of how fast citation patterns move, the Promptwatch Data hub has the full set of monthly reports.
Phase 0: audit what you actually have
Before you evaluate any new tool, write down what the old one is doing for each client. Most agencies skip this and regret it at cutover.
For each client, record:
- Number of prompts, and how they are grouped (topics, personas, funnel stage)
- Which AI models are tracked, and whether each is tracked through the real interface or an API
- Countries, languages, and any city-level or coordinate targeting
- The competitor list, including any aliases and the brand name rules (for example, how a brand whose name is a common word is detected)
- Run frequency
- Which metrics appear in client reports, with the exact definition the old tool uses
- Any custom dashboards, scheduled reports or Slack alerts that clients depend on
The metric definition is the one people forget. If the client deck says share of voice, you need to know whether that means share of all brand mentions, share of responses, or something else. You will need that to map it to the new tool.
Phase 1: export everything before you touch the new tool
Exports are your insurance policy. Do this first, while the old subscription is fully active.
Pull, per client:
- The prompt list with tags, topics and any intent labels
- Raw AI responses with dates, model names and country, if the tool stores them. This is the single most valuable export, because you can re-score raw text later under any definition.
- Citations with URLs and dates
- Daily or weekly time series for every metric you report
- Competitor lists and brand aliases
- Screenshots of the dashboards and the exact report PDFs you have sent the client
Store the files somewhere the client relationship owns, not a personal drive. Name them with the tool, client, and export date.
A practical test: pick one prompt and one date from six months ago and see whether you could reconstruct the answer from your archive alone. If you can't, your export is incomplete. If the old tool does not allow raw response exports, note that as a limit on what you can bridge and plan to lean harder on the parallel run.
A warning on cancellation. Do not cancel the old plan until the parallel run is finished and the client has signed off on the bridge. Some tools delete data fast after a plan ends, and you want a month of cushion.
Phase 2: get to prompt parity
The new tool needs to ask the same questions. Not similar ones. The same ones.
- Import the prompt list as-is first. Resist the urge to improve it during migration. Improvements belong after cutover, and each one gets annotated.
- Check that prompts landed unchanged. Some tools trim, normalize or deduplicate on import.
- Recreate topics and tags so your reporting breakdowns still work.
- Match models, countries and languages. In some tools a monitor's language and location are fixed once created, so set them correctly the first time rather than patching later.
- Recreate competitor sets and aliases, and write the brand detection rules for ambiguous names.
If you use ChatGPT query fan-outs to build or update prompt sets, remember that anything built before August 8, 2026 is probably out of date. That is a reason to treat prompt refresh as its own project, separate from the migration.
Phase 3: run both tools in parallel
This is the step that saves the baseline. For two to four weeks, both tools run the same prompts on the same models in the same locations. Four weeks is better, since AI answers vary and a short window can catch an unrepresentative stretch.
What the overlap gives you:
- A direct measurement of the gap between tools, per metric, per model
- Evidence about whether the gap is stable (a steady offset) or noisy (varying day to day)
- A chance to catch prompt mismatches, location mistakes and brand detection errors while the old tool is still there to compare
Two rules for the window. First, do not change anything else. No new content pushes you want to attribute, no prompt edits, no competitor list changes. Second, check the AI platform calendar. If a known platform change (like the ones above) falls inside the window, note it, because both tools should show it. If only one does, you have found a tool difference, which is exactly what you were looking for.
Also track average citations per response next to the client's brand citations. If the average drops across the board, the platform changed. If only the client's count drops, it is about the client. If only one tool shows a drop, it is about the tool.
Phase 4: build the bridge
At the end of the parallel window you will have two series that overlap. Now decide how to present continuity.
Option 1: an offset or ratio
Compute the ratio of old to new for each metric over the overlap, for example the average of old divided by new. If the ratio is stable, you can restate old history in new-tool terms by multiplying. Label every restated point as estimated. Do this only where the overlap showed a steady ratio, and per metric, never as one blanket factor. Position and sentiment rarely scale the same way as visibility.
Option 2: a labelled break
When the ratio is unstable, do not fake continuity. Draw the old series and the new series on one chart with a clearly marked break and a one-line note. Report change within each series, and compare across the break only directionally.
Option 3: re-score from raw responses
If you exported raw responses, you can sometimes run old text through a new definition to produce a consistent history. It is the cleanest approach and the most work. Reserve it for clients where long-range trend matters, such as annual reviews or board reporting.
| Approach | Best when | Effort | Risk |
|---|---|---|---|
| Ratio restatement | Overlap shows a stable gap | Low | Hides noise if the ratio drifts |
| Labelled break | Gap is unstable or definitions differ a lot | Low | Clients must accept a reset |
| Re-scoring raw responses | You have full raw exports and need long-range trend | High | Time cost, definition debates |
Whichever you pick, record it in writing: which approach, which date range, which ratio, and who signed off.
Phase 5: tell the client before they notice
The baseline problem is half technical and half political. Clients remember last quarter's number.
Send a short note before cutover. It should cover why you are switching, what will change in the reports, the date of the break, and what the bridge method is. Keep it to a few lines. Then add a standing footnote on the chart: Series change on [date]; earlier values restated using a ratio of X, or not restated.
If the new tool is cheaper per response or adds capabilities such as crawler logs or traffic attribution, say so, but keep that separate from the continuity message. Clients hear a pitch and a data warning in the same email as spin.
Keep an annotation log from day one
The cheapest control in this whole playbook is a shared log with a date, an event and an owner. Entries should include:
- Platform changes (March 4, August 8 and August 14, 2026 above)
- Tool changes (cutover date, prompt edits, model additions, location changes)
- Client changes (site migration, big content pushes, PR hits, product launches)
Schedule the cutover away from known platform shifts when you can. If a new model release lands mid-window, extend the parallel run rather than trying to reason around it.
Choosing the destination tool
Migration is also the moment to reconsider whether the old tool still fits. Don't just swap one dashboard for another. Ask what the retainer needs now: monitoring only, or also content production, crawler visibility and traffic attribution.
The agency comparison posts from 3RD and Brandlight are worth reading for the multi-client angle, with the obvious caveat that both are vendors on their own lists and say so or have a clear stake. Brandlight claims Peec AI costs $205 with unlimited seats and that OpenLens has a free multi-tenant tier; verify prices on each vendor's own page.
A quick fit map for agencies:
| Need | Tools to look at |
|---|---|
| Monitoring plus fixing: content, crawler logs, traffic | Promptwatch |
| Mid-market monitoring with straightforward economics | Peec AI, Otterly.AI |
| Enterprise monitoring for technical delivery teams | Profound, Scrunch AI |
| Agency-specific multi-client workflow | Search Party |
| Traditional SEO suite with AI tracking added | Semrush, Ahrefs Brand Radar |
Promptwatch is the one 1001 SEO Media's team uses for its own GEO work, so weigh my take accordingly. It is an end-to-end platform rather than a tracker: it monitors prompts in the real interfaces of ChatGPT, AI Overviews, AI Mode, Gemini, Perplexity and others, then traces citations, logs AI crawler visits, measures AI traffic and produces content. For migrations specifically, a few things matter. Monitors fix language and location at creation, which keeps markets from mixing. Excel and CSV exports cover prompts, responses with brand positions, citations and crawler data. A REST API and an MCP server make it possible to script prompt imports and pull history out later. Agency plans run unlimited projects: Kick-off at $199/mo, Growth at $399/mo and Scale at $799/mo, with a 7-day free trial on monthly billing. Confirm current numbers at promptwatch.com/pricing. There is more on the agency setup at promptwatch.com/solutions/agencies.

Peec AI is the name that comes up most for mid-market monitoring.
Profound is the self-serve enterprise option, better suited to agencies with technical delivery teams.
Otterly.AI sits at the affordable monitoring end.

Scrunch AI covers monitoring for brands and agencies.

Search Party positions itself as agency-focused, which is worth a look if multi-client workflow is your main pain.
AthenaHQ is another monitoring platform that shows up in agency shortlists.
If you want to browse beyond this shortlist, the GEO software directory at bestgeosoftware.com and the AI rank tracking directory at ai-rank-tools.com list far more options.
Questions to ask any vendor before you migrate
- Can I export raw responses, not just scores?
- How is each metric calculated, in writing?
- Does tracking use the real interface or an API for each model?
- What happens to my data if I downgrade or cancel, and for how long is it kept?
- Can I import prompts in bulk, and are they kept verbatim?
- Is there an API so I can script client onboarding and pull history later?
- Can I run a trial on real client prompts without committing?
A four-week migration timeline
| Week | What happens |
|---|---|
| Before week 1 | Audit the old setup, export everything, check the platform change calendar, notify the client |
| Week 1 | Import prompts into the new tool, verify parity, set models, locations and competitors |
| Weeks 2 to 3 | Parallel run, no other changes, log any platform events |
| Week 4 | Compare series, pick a bridge method, draft the chart footnote, get client sign-off |
| After week 4 | Cut over reporting, keep the old plan for one more cycle, then cancel |
For an agency with many clients, stagger them. Move the smallest, most forgiving account first, learn what breaks, then do the rest. Do not move every client in the same week, because one platform event will then contaminate all of them at once.
Common mistakes
The first is cutting over mid-month with no overlap, so the client sees a step change in the monthly report. The second is improving the prompt list during migration, which makes it impossible to tell whether movement came from the tool, the prompts or reality. The third is comparing a score from an API-based tool against one from a real-interface tool and assuming they should match. They shouldn't.
The fourth is blaming the client's content for a platform shift. The March 4 citation drop hit three ChatGPT models at once, and anyone who read it as a content problem wasted a month. The fifth is reading a single snapshot as a trend. Promptwatch's own advice on the citation drop is to track citations per response over time, and single-snapshot audits are unreliable.
The last one is silence. If the client learns about the series break from a chart rather than from you, the number becomes the story instead of the work.
What good looks like
After a clean migration, the client has a chart with a labelled break, a one-paragraph explanation, an annotation log, and a restated or clearly separated history. The agency has an archive of raw exports, a prompt set that matches the old one, and a parallel-run record showing how far the two tools differ on each metric. The next time you switch (and in this market you probably will), you already have the template.


